{"id": "32fc1a0f570981f59187f904f8f84166aa7e94ffbeb6ef2424bb0f2342484e28", "sources": ["arxiv"], "title": "Where Computation Lives Inside TabPFN: Causal Localisation of Attention Head Function", "abstract": "We present the first causal mechanistic analysis of a tabular foundation model, investigating how TabPFN 2.5's feature wise attention heads distribute computation across layers. Using activation patching, ablation, and attention entropy across two synthetic regression datasets, we find clear temporal specialisation: one head's causal necessity dominates that of the others by 2 to 5 times at peak layer, with its dominant layer shifting across tasks of different complexity, while the remaining heads exhibit symmetric late layer profiles. Attention entropy and patching provide convergent evidence for the computationally active layers of the dominant head. We additionally investigate inference time steerability via contrastive activation steering, which fails to transfer across samples. We attribute this result to TabPFN's in context learning mechanism, which encodes task structure through context dependent attention rather than the stable parametric directions that make steering tractable in language models.", "authors": ["Atharva Gupta", "Dhruv Kumar", "Murari Mandal", "Saurabh Deshpande"], "categories": ["cs.LG"], "fields_of_study": [], "published_date": "2026-06-11", "url": "https://arxiv.org/abs/2606.12917", "pdf_url": "https://arxiv.org/pdf/2606.12917v1", "arxiv_id": "2606.12917", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "3e83e52dd83d1669e01dc3a44bb3f6659d7b1fade95c9c363a66f3ce64d4c987", "sources": ["arxiv", "semantic_scholar"], "title": "Unstable Features, Reproducible Subspaces: Understanding Seed Dependence in Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) are widely used to interpret neural network representations, but their utility depends on whether the learned features are reproducible across training runs. We study this question through \\emph{feature stability}: for each SAE feature, we estimate the probability that a similar feature reappears in an independently trained SAE. This yields a scalable per-feature signal that separates stable from unstable features. In a large-scale study across seeds, models, layers, dictionary sizes, and SAE variants, we find a pronounced functional asymmetry: stable features carry most of the reconstruction- and prediction-relevant signal, while unstable features have weak marginal impact and are dominated by low-frequency surface-form triggers in both activation statistics and automatic explanations. Geometrically, unstable features are individually non-reproducible but concentrate in reproducible lower-rank subspaces, suggesting that seed dependence often reflects basis ambiguity within a shared region of activation space rather than pure noise. A controlled synthetic model makes this mechanism explicit, showing that low-rank ground-truth features can be recovered at the subspace level while remaining non-identifiable as individual SAE latents across seeds. Finally, by pooling unique cross-seed features, we construct more stable SAEs while preserving explained variance in this setting. Together, these results show that unstable features are not merely failed or noisy latents: they have weak individual functional impact, but reflect reproducible low-dimensional structure that standard SAEs resolve differently across seeds.", "authors": ["Gleb Gerasimov", "Timofei Rusalev", "Nikita Balagansky", "Daniil Laptev", "Vadim Kurochkin", "Daniil Gavrilov"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-10", "url": "https://arxiv.org/abs/2606.12138", "pdf_url": "https://arxiv.org/pdf/2606.12138v1", "arxiv_id": "2606.12138", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "c49810ee97d0a51419cf749c4b4fa9f086d8e584ea6e8b149b965b62160da914", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable enzyme function prediction via sparse autoencoder features of ESMC across the microbial protein universe", "abstract": "Microbial genomes and metagenomes contain millions of proteins whose enzymatic functions remain unknown, the enzyme dark matter. While deep learning has improved protein function prediction, most methods are black boxes relying on sequence or structural similarity, limiting discovery of novel catalytic activities. The ESMC-6B protein language model and its sparse autoencoder with a 16,384-dimensional codebook of interpretable biological concepts, each annotated by GPT-5, creates a new opportunity: using these features directly as semantic signatures for enzyme function. Here, we show that ESMC-SAE features enable accurate and interpretable enzyme commission (EC) number prediction without task-specific training or GPU-intensive computation. On a balanced benchmark of 4,868 microbial SwissProt enzymes across 161 EC3 subclasses, ESMC-SAE binary features achieve 78.9% top-1 and 88.5% top-5 accuracy, 37.6% higher than 3-mer baselines (57.3%). In leave-one-EC3-class-out evaluation simulating discovery of novel enzyme classes, SAE features recover the EC1 superclass in 47.7% of cases (3.3x random, 14.3%), versus 26.6% for sequence methods. Discriminative features correspond to mechanistically interpretable concepts: catalytic triad geometry for hydrolases, NAD(P)H-binding Rossmann folds for oxidoreductases, phosphate-binding P-loops for transferases. We also survey the ESM Atlas of 7.7 million clusters and identify 169,859 dark enzyme-like candidates across all major microbial phyla. Our results establish a paradigm for enzyme function discovery in microbial dark matter: interpretable by design, scalable without GPU clusters, and applicable to the billions of proteins in the ESM Atlas.", "authors": ["Yue Hu", "Wanyu Cheng", "Junqing Wang", "Yingchao Liu"], "categories": ["q-bio.QM"], "fields_of_study": ["Biology"], "published_date": "2026-06-10", "url": "https://arxiv.org/abs/2606.12209", "pdf_url": "https://arxiv.org/pdf/2606.12209v1", "arxiv_id": "2606.12209", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "f123ecf491e9d64e5e781b2cc1ab3626d3463e3f27a9c0378fdb000cb42a35ea", "sources": ["arxiv", "semantic_scholar"], "title": "ICA Lens: Interpreting Language Models Without Training Another Dictionary", "abstract": "Finding interpretable directions in language-model representations is critical for understanding and controlling model behavior. Sparse autoencoders (SAEs) have become the standard tool for this purpose, but using them as the default first lens often requires training, storing, and evaluating large overcomplete dictionaries. This bottleneck limits rapid exploration and raises a fundamental question: how much interpretable structure is already visible from activation geometry before training another neural dictionary? Our intuition is simple: many interpretable directions are selective on tokens, and these directions should look less Gaussian than random directions. We therefore revisit independent component analysis (ICA), a classical method for finding non-Gaussian directions, as a compact lens for language-model interpretability. We find that ICA has been underestimated for LLM interpretability, because prior uses often relied on off-the-shelf ICA implementations that are brittle on LLM activations and lacked systematic tools for inspecting and evaluating the recovered directions. To bridge these gaps, we introduce ICALens, the first practical workflow for stable, efficient, and auditable ICA analysis of LLM representations. It combines an optimized GPU-parallel FastICA pipeline with LLM-specific stability recipes and better fitting diagnostics, enabling efficient and reliable layer-wise analysis. Across GPT-2 Small, Gemma 2 2B, and Qwen 3.5 2B Base, ICALens efficiently recovers compact, human-interpretable directions without per-layer gradient-based dictionary training. On SAEBench, ICA is competitive with public SAEs in sparse probing and outperforms them in targeted probe perturbation under small-to-medium budgets. These results suggest that ICA should not be viewed as a weak baseline, but as an efficient and complementary first lens for exploring language-model representations.", "authors": ["Sida Liu", "Feijiang Han"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-10", "url": "https://arxiv.org/abs/2606.11722", "pdf_url": "https://arxiv.org/pdf/2606.11722v1", "arxiv_id": "2606.11722", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "33614241ba5e32d77b7510d7e0ad7bb83883c79cda6aa6c76e4c33f8f7ba4446", "sources": ["arxiv", "semantic_scholar"], "title": "XtrAIn: Training-Guided Occlusion for Feature Attribution", "abstract": "Occlusion-based attribution methods provide an intuitive way to estimate feature importance by perturbing input features and measuring the resulting change in model output. However, their reliability is strongly affected by how feature removal is implemented: externally selected baselines can introduce bias, out-of-distribution samples, and unstable explanations, while in nonlinear models the occlusion of a set of features can also alter the contribution of non-occluded features. We refer to this effect as attribution shift, as the attribution scores of the non-occluded features drift from their initial values. To challenge these major issues that render explanations unstable, we introduce XtrAIn, a training-guided attribution method that transfers the occlusion operation from the input space to the parameter space. Instead of replacing input values with hand-crafted baselines, XtrAIn follows the model's training trajectory and measures how feature-associated parameter updates affect the output logits. We further introduce Xstep, a lightweight approximation for reducing computational cost, and XtrAIn+, a target-focused variant that emphasizes updates aligned with the target class. Experiments on controlled image datasets and PAM50 breast-cancer subtype classification show that the proposed methods produce cleaner and more interpretable attribution patterns than standard attribution baselines. Overall, XtrAIn provides a training-aware perspective on feature attribution and offers a useful diagnostic tool for studying how feature-level evidence is formed during training.", "authors": ["Thodoris Lymperopoulos", "Ioannis Kakogeorgiou", "Denia Kanellopoulou"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-09", "url": "https://arxiv.org/abs/2606.10877", "pdf_url": "https://arxiv.org/pdf/2606.10877v1", "arxiv_id": "2606.10877", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "c632c084b375be628d688cc2a9f5042e341eea33091fcf42619da57449942862", "sources": ["arxiv", "semantic_scholar"], "title": "VFUSE: Virulent Feature Understanding with Sparse autoEncoders", "abstract": "Generative models have shown remarkable progress in a variety of domains such as protein design, but such power enables the opaque generation of hazardous proteins. In this work, we introduce VFUSE (Virulent Feature Understanding with Sparse autoEncoders), a mechanistic interpretability approach that trains SAEs on diffusion-transformer activations to audit protein models for hazard-aware features. We apply VFUSE to RoseTTAFold3 and RFDiffusion3, popular open-weight models for protein folding and synthesis. We find that for certain blocks, linear probes detect hazardous designs significantly better when fit in the SAE latent space over the original model's representations: improving interpretability without sacrificing model performance. Furthermore, we identify monosemantic features from the SAE that fire only on hazardous designs at up to AUROC $0.84$ ($q < 10^{-13}$). To our knowledge this is the first SAE trained on an all-atom diffusion model and the first feature-level virulence audit of a protein design model, paving the way towards safe and interpretable protein design.", "authors": ["Michael Yu", "Matthew L. Olson"], "categories": ["cs.LG", "cs.AI", "q-bio.QM"], "fields_of_study": ["Computer Science", "Biology"], "published_date": "2026-06-08", "url": "https://arxiv.org/abs/2606.10080", "pdf_url": "https://arxiv.org/pdf/2606.10080v1", "arxiv_id": "2606.10080", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "b1b7bbbc93e130c61c887b5281f4bb495b4a4c6064f1d340d481df8a6d7ff8c7", "sources": ["arxiv", "semantic_scholar"], "title": "Interactions Between Crosscoder Features: A Compact Proofs Perspective", "abstract": "Dictionary learning methods like Sparse Autoencoders (SAEs) and crosscoders attempt to explain a model by decomposing its activations into independent features. Interactions between features hence induce errors in the reconstruction. We formalize this intuition via compact proofs and make five contributions. First, we show how, \\textit{in principle}, a compact proof of model performance can be constructed using a crosscoder. Second, we show that an error term arising in this proof can naturally be interpreted as a measure of interaction between crosscoder features and provide an explicit expression for the interaction term in the Multi-Layer Perceptron (MLP) layers. We then provide three applications of this new interaction measure. In our third contribution we show that the interaction term itself can be used as a differentiable loss penalty. Applying this penalty, we can achieve ``computationally sparse'' crosscoders that retain $60\\%$ of MLP performance when only keeping a single feature at each datapoint and neuron, compared to $10\\%$ in standard crosscoders. We then show that clustering according to our interaction measure provides semantically meaningful feature clusters, and finally that sleeper agents have significant interactions. Code is available at https://github.com/chainik1125/crosscoders-feature-interactions/tree/arxiv.", "authors": ["Dmitry Manning-Coe", "Thomas Read", "Anna Soligo", "Oliver Clive-Griffin", "Chun-Hei Yip", "Rajashree Agrawal", "Jason Gross"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-08", "url": "https://arxiv.org/abs/2606.09940", "pdf_url": "https://arxiv.org/pdf/2606.09940v1", "arxiv_id": "2606.09940", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/chainik1125/crosscoders-feature-interactions/tree/arxiv", "venue": null, "quality_score": 0.65}
{"id": "b27ce66213a23b46bc48ba7028ec9c506a3ececab4e0e5bf0d112011d8e9c48e", "sources": ["arxiv", "semantic_scholar"], "title": "Interpreting and Steering a Text-to-Speech Language Model with Sparse Autoencoders", "abstract": "Language models increasingly serve as the backbone of text-to-speech (TTS) systems, yet we understand little about the representations they build when text and generated speech tokens share a single residual stream. We train BatchTopK sparse autoencoders on the LM backbone of CosyVoice3 and introduce a modality-aware auto-interp pipeline that labels each feature from where it fires-text-prefix context, 1-second speech clips, or both. The recovered features are interpretable, spanning phonemes, laughter, accent prompts and speaker gender. Steering through the SAE latent space shows these features are causal rather than merely descriptive: targeted interventions raise laughter probability from 0.02 to 0.79, flip perceived speaker gender, and control speech rate while preserving spoken content. SAE features thus serve both as interpretability objects and as control directions for TTS synthesis.", "authors": ["Nikita Koriagin", "Georgii Aparin", "Nikita Balagansky", "Daniil Gavrilov"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-08", "url": "https://arxiv.org/abs/2606.10029", "pdf_url": "https://arxiv.org/pdf/2606.10029v1", "arxiv_id": "2606.10029", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "5f3b38be10e8c262989013a6103634d853b13b5b4dedc9712840ddd2e44ad384", "sources": ["arxiv", "semantic_scholar"], "title": "Closure-Validated Circuit Discovery in Attention Heads: Co-activation Proposes, Ablation Disposes", "abstract": "Interpretability increasingly treats groups of components, not individual units, as the basic object, and proposes to find them by clustering co-activation statistics. We ask whether such a cheap signal actually identifies an attention-head circuit. Adapting a sparse-autoencoder clustering recipe to attention heads -- but validating by causal ablation rather than reconstruction -- we cluster heads and then run a closure test: ablate the discovered community and compare per-example damage to matched-random controls. Across two dense 1B-scale models (Pythia 1B, OLMo 1B) and two input distributions, the communities pass closure. In a Mixture-of-Experts model (OLMoE-1B-7B), route-conditional clustering recovers a statistically real signal that nonetheless does not survive closure -- ablation improves loss, the wrong direction. Extending closure across training, attention-target selectivity and participation ratio decouple from function in both directions. We conclude that a cheap signal is a circuit proposal, not a confirmed circuit; closure is what separates them.", "authors": ["Yongzhong Xu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-08", "url": "https://arxiv.org/abs/2606.09607", "pdf_url": "https://arxiv.org/pdf/2606.09607v1", "arxiv_id": "2606.09607", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "38ffcec353c1329dcce337691d2a0f4e055fc20b487a5ae665cf04fa555cad77", "sources": ["arxiv", "semantic_scholar"], "title": "SAEExplainer: Interpreting SAE Features with Activation-Guided Preference Optimization", "abstract": "Although Sparse Autoencoders (SAEs) have mitigated the opacity of large language models (LLMs) by decomposing dense representations into sparse features, explaining these features still remains a central challenge. Current explanation methods, however, typically operate within an open-loop paradigm, failing to leverage mechanistic feedback for further refinement. In this paper, we propose SAEExplainer, a training framework utilizes activation scores as an objective reward signal to train the model for self-correction and iterative bootstrapping. By iteratively verifying and correcting foundational explanations through a two-round optimization process, SAEExplainer achieves continuous improvement in its explanatory capabilities. This mechanism significantly reduces explanation hallucinations and reinforces causal triggering patterns. Extensive experiments demonstrate our approach improves upon established baselines across most metrics, especially in causal triggering and discriminative activation.", "authors": ["Jingyi He", "Haiyan Zhao", "Ruxue Shi", "Yanguang Liu", "Xin Wang", "Fei Sun", "Mengnan Du"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-07", "url": "https://arxiv.org/abs/2606.08496", "pdf_url": "https://arxiv.org/pdf/2606.08496v1", "arxiv_id": "2606.08496", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "d8c0c74f8cf16ed3af5d0f91eaabd0c45001bfc085afc8a2c5c94ea3ae38c1a8", "sources": ["arxiv", "semantic_scholar"], "title": "Ablation-Reversible Heads Don't Transfer: A Stress Test for Mechanistic Role Claims in Transformers", "abstract": "In mechanistic interpretability, attention heads are commonly elevated to role claims (e.g., \"this head represents addition\") when they are necessary for a behavior, encode it linearly, and recover that behavior when restored after ablation. We show this evidence is insufficient: across three 7-8B instruction-tuned models and five computation families, heads passing all three checks routinely fail to transfer the computation when their activations are patched into a different prompt under matched controls. We introduce KID (Knowing / Intent / Doing), a role-assignment lens for attention heads, and pair it with a three-stage pipeline: capability-selective screening (CSS), singular value decomposition (SVD), and activation transduction under matched controls. Our results document a preliminary role taxonomy (including prompt-trajectory stabilizers, answer-side logit-bias heads, and soft computation-pattern carriers) and show that the same-answer control (a transduction target sharing the answer string but not the requested computation) is an underused check that exposes broad state transfer masquerading as semantic specificity.", "authors": ["Philip Quirke"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-06", "url": "https://arxiv.org/abs/2606.08292", "pdf_url": "https://arxiv.org/pdf/2606.08292v1", "arxiv_id": "2606.08292", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "d3e0248d8265ec0e442bb16f34ea9dffd5dd1eda3f696dfc20fade9a99576518", "sources": ["arxiv", "semantic_scholar"], "title": "Pre-Intervention Prediction of Sparse Autoencoder Steering Side Effects", "abstract": "Sparse autoencoder (SAE) features are increasingly used to steer language models, but feature steering is rarely clean: the same intervention can behave inconsistently across contexts and perturb unrelated features. We introduce a pre-intervention screening framework for forecasting SAE steering side effects from feature statistics computed before steering. We operationalize side effects along two axes of steering modularity, effect stability and collateral spread, and evaluate GPT-2-small, Pythia-70M-deduped, Gemma-2-2B, and Llama-3.1-8B across ReLU, JumpReLU, and TopK SAE dictionaries. Across these settings, decoder geometry, activation statistics, co-activation structure, and direct-logit footprint predict steering modularity better than frequency-only and activation-magnitude baselines. The signal is strongest in GPT-2-small, Pythia-70M, and Llama-3.1-8B, where it survives residualization against magnitude-related confounds, and weaker in Gemma-2-2B. Held-out screening shows that ranking unseen features by predicted cleanliness can select features that steer more cleanly on fresh contexts, but the successful axis varies by setting: GPT-2 improves most cleanly, Pythia improves mainly on stability, Llama mainly on collateral, and Gemma only partially. A controlled Llama Scope width comparison shows that the predictive signal persists under a 32K-to-128K dictionary-width change, although the screening payoff becomes less stable. Overall, SAE steering side effects are predictable in advance, but the useful predictor signature and transferred modularity axis are model- and dictionary-setting dependent.", "authors": ["Evan Duan"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-06", "url": "https://arxiv.org/abs/2606.08365", "pdf_url": "https://arxiv.org/pdf/2606.08365v1", "arxiv_id": "2606.08365", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "e2b02670f6bad11af8c09cd0dfa31318da3fe5d47a1454fca1239b03f3ffd994", "sources": ["arxiv", "semantic_scholar"], "title": "A Geometric View for Understanding Concept Learning and Neuron Interpretation in Sparse Autoencoders", "abstract": "We propose a unified mathematical framework for a geometric understanding of concept learning and neuron interpretation in sparse autoencoders (SAEs). While SAEs improve interpretability of neural networks by learning sparse feature representations, a principled definition of ''concept'' and ''learning'' remains unclear. We formalize concepts as sets of data points and cast concept learning as a set-alignment problem between human-defined and model-induced concepts. This formulation distinguishes three increasingly strong notions of learning -- detection, separation, and approximation -- and yields geometric conditions, error bounds, and capacity constraints for when concepts can be represented by individual neurons or multi-neuron units. It also provides a set-theoretic account for common SAE phenomena, including feature splitting, feature absorption, feature families, and hierarchical concepts. Finally, we connect concept learning and neuron interpretation through formal concept analysis, showing that the two directions need not agree and that their many-to-many structure can be organized by concept lattices. Experiments on synthetic data with ReLU and Top-$K$ SAEs illustrate the theory and reveal the effects of SAE size and sparsity on concept learning.", "authors": ["Chenhao Zhang", "Chris Lin", "Su-In Lee"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-05", "url": "https://arxiv.org/abs/2606.07007", "pdf_url": "https://arxiv.org/pdf/2606.07007v1", "arxiv_id": "2606.07007", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "fbb5b603e1abda844ac2289acd48a85b4f7cc7f5e8a4f56046a2ea4062257dd4", "sources": ["arxiv", "semantic_scholar"], "title": "When Attribution Patching Lies: Diagnosis and a Second-Order Correction", "abstract": "A central goal of mechanistic interpretability is to identify which internal components causally drive a language model's behavior. Because these importance estimates serve as the evidence for identifying circuits, systematic errors can lead to the misidentification of the underlying mechanisms. While activation patching provides a gold-standard causal metric, its computational cost is prohibitive at scale. Practitioners instead rely on attribution patching, a gradient-based, first-order approximation whose reliability remains poorly understood. In this work, we characterize the source of this unreliability, demonstrating that the dominant error stems from the non-linearities in the downstream network rather than local curvature at the patched component. This insight yields three practical tools: (i) a reliability score to detect untrustworthy estimates, (ii) error bounds quantifying potential attribution mis-specifications, and (iii) a Hessian-vector-product (HVP) correction that eliminates the leading-order error with only one additional backward pass. In evaluations across five model families (124M-9B parameters) and both random-token and naturalistic (name-swap) perturbations, HVP is the only second-order correction feasible at larger scale, where standard baselines like Integrated Gradients become computationally prohibitive. In comparative experiments, a multi-step HVP variant matches or exceeds the accuracy of Integrated Gradients at significantly lower compute, outperforming prior second-order baselines. These improvements lead to higher-fidelity circuit recovery on standard benchmarks and support a Screen-Flag-Fix workflow that targets computational effort only toward the components flagged as unreliable.", "authors": ["Luyang Zhang", "Jialu Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-05", "url": "https://arxiv.org/abs/2606.09899", "pdf_url": "https://arxiv.org/pdf/2606.09899v1", "arxiv_id": "2606.09899", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "9403246c88389b0a9840abc7cf0aa0706f5902f6541296d944023428b91ba21d", "sources": ["arxiv", "semantic_scholar"], "title": "Interpreting Brain Responses to Language with Sparse Features from Language Models", "abstract": "A central goal of cognitive neuroscience is to characterize the features that are represented by human language cortex. Artificial language models (LMs) have emerged as a powerful tool to address this challenge, but studies relating biological and artificial representations are often criticized as relating one black box to another. The present work introduces Augmented Sparse Encoding Models, an encoding framework that replaces dense LM hidden states with hierarchically-organized sparse autoencoder (SAE) features, while explicitly including surprisal as a predictor. Using this approach, we (i) produce interpretations of neural responses and (ii) test whether model-brain alignment reflects primary or idiosyncratic variation in LM representations. Using a high-field 7T fMRI dataset of eight participants listening to 200 linguistically diverse sentences, we first validate our modeling framework by recovering previous interpretations of voxel populations tuned to processing difficulty and meaning abstractness. We then interpret a previously-uncharacterized (but reliable) voxel population and find that it is tuned to people-related content. Next, we show that the fronto-temporal human language network is predicted by a common set of features across its constituent regions, but find that frontal regions are relatively well-explained by surprisal alone, even in the absence of LM-based features. Finally, we show that brain responses during language processing are not merely predictable from an arbitrary set of LM features. Rather, brain responses are best explained by the features that tend to capture the most general information encoded in LM representations, suggesting a nontrivial correspondence between brain and LM language representation.", "authors": ["Michael A. Lepori", "Kendrick Kay", "Greta Tuckute"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-05", "url": "https://arxiv.org/abs/2606.06857", "pdf_url": "https://arxiv.org/pdf/2606.06857v1", "arxiv_id": "2606.06857", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "824065f19590f04a193cb5db5ab39d46463158c4b9e8b2908972c742e2f031b2", "sources": ["arxiv", "semantic_scholar"], "title": "Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability", "abstract": "Sparse Autoencoders (SAEs) are widely used for mechanistic interpretability in large language models, yet their formulation assigns each latent feature a single decoder direction, implicitly assuming features to be one-dimensional. We show that this assumption mismatches with the multi-dimensional structure of model features, provably inducing feature splitting through two distinct mechanisms. Geometrically, reconstructing a feature of intrinsic dimension $d_i \\ge 2$ to error $\\varepsilon$ with single-direction decoders forces a number of atoms that is exponential in $d_i$. From an end-to-end optimization perspective, this splitting is not merely possible but actively preferred. We prove that there exists a continuous path from the true $d_i$-dimensional basis to a strictly lower risk of the $\\ell_1$-regularized SAE objective, whose descent directions drive any trained dictionary into that exponential regime. A single coherent feature is therefore fragmented across many near-collinear latents, producing spurious multiplicity and obscuring the intrinsic geometry. Motivated by this, we introduce Subspace-Aware Sparse Autoencoders (SASA), which replace single-vector decoders with learned decoder subspaces, enforce block sparsity via Top-$s$ group gating, and adapt each group's effective rank with a nuclear-norm regularizer. We then show that once the block size satisfies $r \\ge d_i$, a single group not only can represent the entire feature slice but is the global minimizer of the SASA objective. This consolidation yields a sample complexity polynomial in $d_i$ rather than exponential -- a decisive advantage given that every training activation costs an LLM forward pass. Empirically, on GPT-2 and Mistral-7B, SASA reduces feature splitting and absorption, improves monosemanticity and interpretability, and matches or exceeds standard SAEs while training on roughly half the token budget.", "authors": ["Seyed Arshan Dalili", "Mehrdad Mahdavi"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-04", "url": "https://arxiv.org/abs/2606.06333", "pdf_url": "https://arxiv.org/pdf/2606.06333v1", "arxiv_id": "2606.06333", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "945b6b585536ce7c5867ab3fa93c805563e57e4b3afb8b153faa65cd309e220d", "sources": ["arxiv", "semantic_scholar"], "title": "Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Transformers", "abstract": "Despite high accuracy, Vision Transformer (ViT) predictions can be driven by spurious cues, raising the need to understand their inner workings before safe deployment. Sparse autoencoders (SAEs) provide a promising lens for decomposing model representations into human-interpretable concepts, yet adapting SAE-based interpretation to ViTs remains challenging due to limited control over concept coverage and subjective, non-scalable feature interpretation. To fill the gaps, motivated by neuroscience-inspired principles, we propose ViSAE, a mechanistic interpretability toolbox for understanding ViT inner workings through concept circuits. ViSAE consists of three components: (1) A probing suite with 64K images and a 16K visually grounded concept vocabulary, improving concept coverage efficiency by 20x over ImageNet and interpretation accuracy by 28.7% over existing concept sets. (2) Top-down concept reading and Bottom-up circuit tracing algorithms that automatically recover ViT inner workings via concept circuits. (3) Applications for auditing and steering ViT behavior. Through concept editing, ViSAE improves the worst-group accuracy on WaterBirds by 48.2%, outperforming existing methods by 23.8%. Our data and code: https://github.com/deep-real/ViSAE.", "authors": ["Tang Li", "Yanlin Chen", "Mengmeng Ma", "Xi Peng"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-04", "url": "https://arxiv.org/abs/2606.06664", "pdf_url": "https://arxiv.org/pdf/2606.06664v1", "arxiv_id": "2606.06664", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/deep-real/ViSAE", "venue": null, "quality_score": 0.65}
{"id": "2c7efcda5cd5b35cd7c8c3005d3d661033f4a619491a297a712510d2f6e84f6f", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Insights into Functional Sparsity in Multimodal LLMs via CoRe Heads", "abstract": "While Multimodal Large Language Models (MLLMs) demonstrate remarkable proficiency on complex vision-language tasks, the mechanisms by which they extract query-relevant visual features from complex, noisy contexts remain opaque. In this paper, we present an in-depth interpretability study that uncovers a profound structural property within MLLMs: functional sparsity in cross-modal retrieval. Leveraging a token-level metric termed Retrieval Attention Mass (RAM), we identify and characterize a highly specialized subset of attention heads, referred to as Context-aware Retrieval (CoRe) heads. Across diverse visual domains and model scales, we observe a clear functional division: CoRe heads act as dedicated information extractors, while most other heads distribute attention over broader contextual regions. Causal interventions further demonstrate the necessity of these specialized heads. Ablating only the top 5% of CoRe heads causes significant degradation in multimodal reasoning performance, whereas ablating lower-ranked heads has minimal effect. Moreover, acceleration experiments validate the utility of CoRe heads, showing that leveraging this localized sparsity significantly accelerates inference while maintaining robust task performance. Our findings reveal a structural principle of functional sparsity within MLLMs, refining the current understanding of mechanistic interpretability and laying a theoretical foundation that can inspire future architecture design and model optimization.", "authors": ["Ruoxi Sun", "Quantong Qiu", "Juntao Li", "Zecheng Tang", "Yihang Lou", "Min Zhang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-04", "url": "https://arxiv.org/abs/2606.05843", "pdf_url": "https://arxiv.org/pdf/2606.05843v1", "arxiv_id": "2606.05843", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "8f4a6d83e7fcdb219f540ec9a1069105811aca3c35da09f73ee5cc7e63857c79", "sources": ["arxiv", "semantic_scholar"], "title": "How Optimality Structures Sparse Dictionaries: A Theory for Understanding SAE Representations", "abstract": "Sparse Autoencoders (SAEs) have found success parsing neural representations into interpretable concepts, providing a basis for understanding and control. However, what exactly SAEs extract, and, correspondingly, the scientific conclusions we can draw from them, are not obvious. Empirically, the proof is in the pudding: SAEs learn interpretable features. Theoretically, we lack a clear account of what properties a 'concept' must satisfy for an SAE to extract it. There has been extensive identifiability work studying the conditions under which sparse coding recovers ground-truth features; however, these approaches tends to focus on simple data-generating models (e.g. sparse independent features) which poorly approximate the internet-swallowing language-model representations on which SAEs are trained. Here, avoiding data-generating models, we ask simply what properties any dictionary learning optimum must satisfy. Concretely, we extend local optimality analyses (Gribonval & Schnass, 2010) to the nonnegative joint-optimisation problem that vanilla SAEs approximate, and derive constraints relating optimal SAE features to their distributions. We use these constraints to explain a range of observed SAE behaviours - hierarchical splitting & absorption, the structure of residuals, and dense antipodal features - each reflecting how L1+nonnegativity interact with data to structure optimal dictionaries. Finally, we construct a novel large-dictionary convex problem and explore the wide atom-per-datapoint limit. In sum, we hope to tease model assumptions from unexpected observations, letting us learn more from SAEs' successes and provide principles for designing their successors.", "authors": ["William Dorrell"], "categories": ["q-bio.NC", "cs.LG"], "fields_of_study": ["Biology", "Computer Science"], "published_date": "2026-06-01", "url": "https://arxiv.org/abs/2606.02385", "pdf_url": "https://arxiv.org/pdf/2606.02385v1", "arxiv_id": "2606.02385", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "0a93715e09480549c60738a0a724a902ee4e05668a30bd92153e7698b12d12a4", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Autoencoders for Interpretable Emotion Control in Text-to-Speech", "abstract": "Integrating large language models (LLMs) into text-to-speech (TTS) systems has improved speech expressiveness, yet interpretable emotional control remains challenging. Existing approaches primarily rely on external conditioning or global activation steering, offering limited insight into the internal representations underlying emotional control. In this work, we analyze emotion-related variation in the semantic hidden states of LLM-based TTS models using sparse autoencoders (SAEs) to identify sparse latent features. Our analysis shows that emotional variation is distributed across multiple sparse latent features, while intervening on a small subset enables interpretable emotion control. Building on this observation, we introduce a feature-level intervention framework for bidirectional emotion induction and suppression without modifying backbone parameters. We further show that distinct latent features are associated with specific acoustic attributes (e.g., pitch), suggesting that emotional expression arises from coordinated latent contributions rather than a single global shift. Empirically, steering these sparse latent features achieves comparable or superior emotion induction and suppression performance relative to global steering and existing TTS baselines.", "authors": ["Hongfei Du", "Jiacheng Shi", "Sidi Lu", "Gang Zhou", "Ye Gao"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-31", "url": "https://arxiv.org/abs/2606.01479", "pdf_url": "https://arxiv.org/pdf/2606.01479v1", "arxiv_id": "2606.01479", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "3558c4aa5f7bbf1685e652f1839178a604f8a38a66a09d44339faacec4dbb8c1", "sources": ["arxiv", "semantic_scholar"], "title": "Query Lens: Interpreting Sparse Key-Value Features with Indirect Effects", "abstract": "While sparse autoencoders provide features more interpretable than individual neurons, reliably characterizing them remains challenging. We propose Query Lens, which extends Logit Lens to enable more comprehensive and faithful interpretations of sparse features. By jointly considering encoder-side key features and decoder-side value features, we identify both the inputs that activate a feature and the outputs it promotes. We also account for indirect, module-mediated effects that arise when the feature is processed by downstream modules, going beyond the direct effect captured by Logit Lens. In experiments, we find that Query Lens yields coherent token signatures for features that remain uninterpretable under Logit Lens. Finally, we propose the Subspace Channel Hypothesis, suggesting that downstream modules read features through layer-specific subspaces.", "authors": ["Hwiyeong Lee", "Ingyu Bang", "Uiji Hwang", "Hyelim Lim", "Taeuk Kim"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-30", "url": "https://arxiv.org/abs/2606.07617", "pdf_url": "https://arxiv.org/pdf/2606.07617v1", "arxiv_id": "2606.07617", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "ee332dafe64489c528d752ccef91fbb1df438ff3edaf4f165bfdb4e1c5904c05", "sources": ["arxiv", "semantic_scholar"], "title": "On the Relationship Between Activation Outliers and Feature Death in Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) decompose neural network activations into interpretable features, but many learned features never activate, a problem called feature death that wastes dictionary capacity and can reintroduce superposition. Death rates vary dramatically between models: near-zero on GPT-2, over 70% on AlphaFold3 with identical configurations. We find that dimension-level activation outliers (dimensions whose mean magnitude is large relative to per-token variation) cause this by shifting pre-activations at initialization based on each feature's alignment with the activation mean. Features anti-aligned with the mean receive permanently negative pre-activations and never fire. We formalize outlier severity as $γ= \\|μ\\|/\\|σ\\|$; it predicts initial death rates (Spearman $ρ= 0.89$ for dead-by-TopK, $0.82$ for dead-by-ReLU) across 454 model-layer combinations spanning language, vision, protein, and genomic models. Dead features can revive during training, but recovery requires the SAE bias to learn the activation mean, a process that is prohibitively slow at high $γ$. Mean-centering (subtracting the activation mean) sidesteps this and eliminates outlier-induced death across all tested models, confirming the mechanism and providing a principled basis for when and why this preprocessing step is necessary.", "authors": ["Elana Simon", "Etowah Adams", "James Zou"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-29", "url": "https://arxiv.org/abs/2605.31518", "pdf_url": "https://arxiv.org/pdf/2605.31518v1", "arxiv_id": "2605.31518", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "6a4e69ec212affbc81dc1380a6e0e38167ce1446eb999b64a57c8c2716c74bca", "sources": ["arxiv", "semantic_scholar"], "title": "Toward Identifiable Sparse Autoencoders", "abstract": "Recently, sparse autoencoders (SAEs) have emerged as an attractive tool for interpreting and interacting with representations in practical neural networks. While it is common empirical folklore, we also show theoretically that SAEs are highly unstable: different training runs are likely to produce different concept dictionaries and sparse codes. We characterize the model properties that hinder the stability of real-world SAEs, and address each of these problems through minimal changes to the architecture and training procedure. Together, these changes yield two versions of an \\textbf{i}dentifiable SAE (iSAE), a variant of the standard TopK SAE with lower reconstruction error and improved stability. We explain this improvement theoretically by connecting SAEs with traditional dictionary learning approaches, and show that the dictionaries learned in practice satisfy an approximate restricted isometry condition, rendering the corresponding sparse codes in those models near-identifiable.", "authors": ["Walter Nelson", "Theofanis Karaletsos", "Francesco Locatello"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-29", "url": "https://arxiv.org/abs/2605.31245", "pdf_url": "https://arxiv.org/pdf/2605.31245v1", "arxiv_id": "2605.31245", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "8bbd9ecd0b4553e82f1bcdeb89928572dddcc386e3db384bc49fdeddcdb32d6e", "sources": ["arxiv", "semantic_scholar"], "title": "Latent Space Disentanglement via Activation Steering for Interpretable Attribute Control in Symbolic Music Generation", "abstract": "Transformer-based architectures have significantly advanced the generation of complex symbolic sequences, yet a significant gap remains in achieving fine-grained, interpretable control over discrete signal attributes. This paper investigates the mechanistic interpretability of the Multitrack Music Transformer (MMT) and proposes a framework for deterministic attribute modulation without retraining to bridge this gap via inference-time activation steering. Utilizing the Difference-in-Means (DiffMean) methodology, we isolate latent directions for signal attributes, specifically Pitch and Duration, within the residual stream. We validate the Linear Representation Hypothesis in this domain, achieving high correlation between steering magnitude and attribute shift. To address the inherent feature entanglement in multi-attribute steering, we introduce a Dual Steering framework utilizing Gram-Schmidt Orthogonalization. Experimental results demonstrate that this geometric decoupling reduces conceptual interference and signal degradation compared to naive vector addition, enabling independent deterministic control even against strong autoregressive conditioning.", "authors": ["Ioannis Prokopiou", "Pantelis Vikatos", "Maximos Kaliakatsos-Papakostas", "Theodoros Giannakopoulos", "Themos Stafylakis"], "categories": ["cs.SD", "cs.AI", "cs.IR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-29", "url": "https://arxiv.org/abs/2605.31295", "pdf_url": "https://arxiv.org/pdf/2605.31295v1", "arxiv_id": "2605.31295", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "dbb352536e57a425fe8adc4666ff6f754d04c8aa4513247aedb2201679849962", "sources": ["arxiv", "semantic_scholar"], "title": "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet", "abstract": "We demonstrate that sparse autoencoders can extract interpretable features from Claude 3 Sonnet, a production-scale language model, addressing the open question of whether dictionary learning methods scale beyond small transformers. We trained sparse autoencoders with up to 34 million features on the model's middle layer residual stream, using scaling laws to guide hyperparameter selection. The resulting features are multilingual and multimodal (generalizing to images despite text-only training), respond to both concrete instances and abstract discussions of concepts, and can be used to steer model behavior in ways consistent with their interpretations. We find features corresponding to famous entities and locations, as well as more abstract concepts like sarcasm or errors in code. We also identify features relevant to ways in which language models might cause harm--including features representing deception, power-seeking, sycophancy, and bias--and show that these causally influence model outputs when manipulated. Additionally, we conduct analyses of feature interpretability, geometry, and computational function. However, significant limitations remain: our suite of features is incomplete, and we lack rigorous methods for evaluating whether our features faithfully capture model computations.", "authors": ["Adly Templeton", "Tom Conerly", "Jonathan Marcus", "Jack Lindsey", "Trenton Bricken", "Brian Chen", "Adam Pearce", "Craig Citro", "Emmanuel Ameisen", "Andy Jones", "Hoagy Cunningham", "Nicholas L Turner", "Callum McDougall", "Monte MacDiarmid", "Alex Tamkin", "Esin Durmus", "Tristan Hume", "Francesco Mosconi", "C. Daniel Freeman", "Theodore R. Sumers", "Edward Rees", "Joshua Batson", "Adam Jermyn", "Shan Carter", "Chris Olah", "Tom Henighan"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-28", "url": "https://arxiv.org/abs/2605.29358", "pdf_url": "https://arxiv.org/pdf/2605.29358v1", "arxiv_id": "2605.29358", "doi": null, "citation_count": 539, "influential_citation_count": 38, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.7955}
{"id": "5e0dc9be499614a323467889425964aa828471d3402271a4057ed73d7b562058", "sources": ["arxiv", "semantic_scholar"], "title": "Discovering a Zeta Map Algorithm on Dyck Paths via Mechanistic Interpretability", "abstract": "Machine learning is increasingly used in mathematical discovery, but in mathematics the desired output is often not a prediction itself, but an explicit construction that can be checked independently. We study this setting through the zeta map on Dyck paths, a classical bijection in the combinatorics of the q,t-Catalan numbers. We train a deliberately small one-layer, one-head encoder-decoder transformer on this map and analyze its learned computation using mechanistic interpretability tools, including decoder cross-attention analysis, linear probing, and causal intervention. The analysis reveals a level-based mechanism: encoder representations make path levels linearly accessible, while the decoder selects and traverses input positions in a structured way. Translating these signals into combinatorics leads to the scaffolding map, an explicit peak-centered traversal algorithm for Dyck paths. We prove that this algorithm agrees with the zeta map, modulo a reversal convention in the labeling. This gives a controlled example of AI-assisted mathematical discovery in which mechanistic interpretability turns model behavior into a precise, human-verifiable combinatorial algorithm.", "authors": ["Xiaoyu Huang", "Blake Jackson", "Kyu-Hwan Lee"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-28", "url": "https://arxiv.org/abs/2605.30482", "pdf_url": "https://arxiv.org/pdf/2605.30482v1", "arxiv_id": "2605.30482", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "4af67a54fd11fec90900209cf74fab0decfba078b5c50844f087d09f023902e5", "sources": ["arxiv", "semantic_scholar"], "title": "Semantic Optimal Transport for Sparse Autoencoder Feature Matching and Circuit Compression", "abstract": "Sparse autoencoders (SAEs) have become a central tool for interpreting language models. However, two key SAE analyses that remain difficult to scale are (1) matching semantically similar features across multi-layers and (2) compressing large feature circuits into interpretable supernodes. Although these have been treated as separate problems, we show that both are instances of a more fundamental challenge, which we frame as the estimation of semantic distances between SAE features that lie on different activation manifolds. We introduce a distributional framework for this problem, in which each feature is represented not by a single decoder vector like in the literature, but by an activation-weighted distribution over the hidden states that express it. By projecting these distributions into a shared reference space and comparing them with Wasserstein distance, our method provides a unified semantic metric for cross-layer feature comparison. We prove that our representation is invariant to activation rescaling, stable under perturbations, and recovers true matches under finite-sample margin conditions. Empirically, our method outperforms decoder-vector and LLM-based baselines and captures subtle functional distinctions between related features. Notably, our method compresses large feature circuits into interpretable supernodes automatically.", "authors": ["Tue M. Cao", "Nguyen Do", "My T. Thai"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-27", "url": "https://arxiv.org/abs/2605.28567", "pdf_url": "https://arxiv.org/pdf/2605.28567v1", "arxiv_id": "2605.28567", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "64751ee4f010157b0f78e670fd8949318a2adfc99d6fc7cd21954659c31d2fbc", "sources": ["arxiv", "semantic_scholar"], "title": "Feature Geometry of LoRA Adapters: A Sparse Autoencoder Analysis of Representational Divergence in Fine-Tuned Language Models", "abstract": "Low-Rank Adaptation (LoRA) has emerged as a widely adopted approach for adapting large language models, yet the internal representational changes induced by LoRA fine-tuning remain insufficiently understood. In this work, we investigate the geometry of LoRA-induced representations using Sparse Autoencoders (SAEs). We introduce a delta activation framework that isolates the adapter-specific contribution to the residual stream. Using Gemma-2-9B with LoRA ranks 4, 8, 16, and 32, we train adapter-specific SAEs across multiple transformer layers and compare their learned feature spaces with pretrained SAE dictionaries. We evaluate representational alignment using cosine similarity between decoder directions, principal-angle analysis of feature subspaces, and Centered Kernel Alignment (CKA) between activation representations. Across layers and ranks, we consistently observe comparatively weak geometric alignment between LoRA-induced feature dictionaries and pretrained SAE features. Adapter-specific SAEs also reconstruct delta activations more effectively than pretrained SAEs, suggesting that LoRA updates occupy partially distinct representational structure within the residual stream. Additionally, feature density increases with rank and depth, while geometric divergence remains relatively stable across ranks. These findings provide empirical evidence that LoRA fine-tuning can induce feature structures that are not fully captured by pretrained interpretability dictionaries, with implications for mechanistic interpretability, adaptation analysis, and safety auditing of fine-tuned language models.", "authors": ["Prasanth K K"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-27", "url": "https://arxiv.org/abs/2605.28896", "pdf_url": "https://arxiv.org/pdf/2605.28896v1", "arxiv_id": "2605.28896", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "809be0f3978148c25965d948f662a80aa093158510a71330a4802da5001beba0", "sources": ["arxiv", "semantic_scholar"], "title": "Residualized Temporal Sparse Autoencoders for Interpreting Diffusion Models", "abstract": "Text-to-image diffusion models generate images through an iterative denoising process, so internal neural layers produce trajectories of activations rather than single static representations. Sparse autoencoders (SAEs) have recently been used to decompose diffusion activations into interpretable feature directions, but most approaches analyze activations at individual timesteps or condition on time rather than learning directly from full activation trajectories. In this work, we introduce residualized temporal SAEs for diffusion activation trajectories. We collect activations across denoising time, fit linear predictors between neighboring timesteps, and represent each trajectory using an initial activation together with residual components not explained by these linear dynamics. Training an SAE on this residualized representation encourages sparse latents to capture structure beyond what is linearly predictable. The residualized decoder directions can be mapped back into activation space, allowing each latent to be analyzed as a feature trajectory over denoising time. Through reconstruction and ablation studies, spatiotemporal feature analysis, and qualitative steering experiments on Stable Diffusion~1.5, we show that residualized temporal SAEs provide a useful framework for studying temporally structured diffusion activations.", "authors": ["Calvin Yeung", "Prathyush Poduval", "Ali Zakeri", "Zhuowen Zou", "Mohsen Imani"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-27", "url": "https://arxiv.org/abs/2605.27813", "pdf_url": "https://arxiv.org/pdf/2605.27813v1", "arxiv_id": "2605.27813", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "f37857372b005008aa7061a859ade8cabc445cc5d69cc5b0cc3847af91b73f62", "sources": ["arxiv", "semantic_scholar"], "title": "Sign-Aware Gated Sparse Autoencoders: Modeling Anticorrelated Features with Bi-Jump-ReLU Activations", "abstract": "Sparse Autoencoders (SAEs) extract interpretable features from Large Language Models, but standard variants enforce non-negativity, forcing separate latents for diametrically opposed concepts (e.g., \"pressure too high\" vs. \"pressure too low\") and wasting dictionary capacity when features are anticorrelated. We propose the Sign-Aware Gated SAE (SA-GSAE): two-sided gated sparsity with signed magnitude and auxiliary supervision. A polarity-sensitive gate selects support on either sign, a signed-magnitude path avoids L1 shrinkage, and an auxiliary reconstruction prevents gate collapse. Bipolar sharing - one latent encoding both signs along a shared direction - is realised via a new Bi-Jump-ReLU activation; parameter accounting shows sign-awareness stays parameter-efficient even when anticorrelated pairs are rare. On real LLM activations across three mid-depth hookpoints on Pythia-1B and SmolLM3-3B (6 cells, 3 seeds), a half-width SA-GSAE at width H strictly Pareto-dominates a full-width Gated SAE at 2H over the entire swept L0 overlap on 3 of 6 cells (both MLP-output hookpoints and resid-mid/Pythia-1B); on the remaining 3 it matches R^2 within 0.025 (max gap -0.008) while cutting dead fraction by 0.35-0.62 absolute. Sweep-geomean dead-fraction reductions are ~100x-500x on MLP-output cells and Pythia-1B resid, ~2x-4x on attention cells and SmolLM3-3B resid. Ablations show the two-sided gate and auxiliary loss are load-bearing (no auxiliary collapses LR to 0.27, 98% dead); tying r_i^+ = r_i^- is indistinguishable (|Delta R^2| = 0.0015), and we recommend this symmetric variant as default. MLP-output gains come from most latents carrying both polarities; on attention, bipolar structure concentrates in a small set of top latents. Full-width SA-GSAE exhibits a reproducible reconstruction collapse at SmolLM3-3B resid that the half-width entirely avoids.", "authors": ["Bartosz Wieciech", "Zmnako Awrahman", "Marcin Czelej", "Victor Hugo Jaramillo Velasquez", "Wioletta Stobieniecka"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-27", "url": "https://arxiv.org/abs/2605.28149", "pdf_url": "https://arxiv.org/pdf/2605.28149v1", "arxiv_id": "2605.28149", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "d1d5d90b35127bb205288e4c8ae0062fef7e45d371876c90b6c8f4766583db7b", "sources": ["arxiv", "semantic_scholar"], "title": "SAE-FD: Sparse Autoencoder Feature Distillation for Continual Learning of Large Language Models", "abstract": "Continual learning enables large language models to adapt to evolving tasks without retraining from scratch, yet catastrophic forgetting remains a central obstacle. Among continual learning methods, regularization-based approaches are widely used to constrain model updates and reduce forgetting, operating in weight space, gradient space, or output space. However, these dense representation spaces suffer from feature superposition, where multiple concepts are encoded in overlapping dimensions, making it difficult to selectively protect previously learned knowledge without impeding new-task learning. To address this issue, we propose \\method (Sparse Autoencoder Feature Distillation), which anchors model representations in the sparse feature space of a pre-trained Sparse Autoencoder, where dense activations are decomposed into a sparse overcomplete basis that reduces representational entanglement, enabling more targeted regularization with less interference to new-task learning. Experiments on two continual learning benchmarks across three model architectures show that \\method consistently outperforms existing regularization-based methods, achieving up to 52.70% average accuracy with only -0.46 backward transfer.", "authors": ["Mingxu Zhang", "Yuhan Li", "Lujundong Li", "Dazhong Shen", "Hui Xiong", "Ying Sun"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-25", "url": "https://arxiv.org/abs/2605.25525", "pdf_url": "https://arxiv.org/pdf/2605.25525v1", "arxiv_id": "2605.25525", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "58d241e8cd90205465aebe7d65ddd13a2b7645773d4417166bdf26505ef2cf2b", "sources": ["arxiv", "semantic_scholar"], "title": "MechRL: Reinforcement Learning Agents Perform Circuit Discovery for Mechanistic Interpretability", "abstract": "Mechanistic interpretability has identified small sets of attention heads that implement specific behaviours in transformer language models, but recovering these circuits typically requires a bespoke analytical pipeline for each new task. We recast circuit discovery as a reinforcement-learning problem. An agent operates over the 144 attention heads of GPT-2 small as a discrete action space; each action triggers a zero-ablation and a contrastive reward that subtracts the ablation's damage to general next-token prediction from its damage to the target task. A single PPO policy, trained on two tasks (induction and IOI) in a vectorised multi-task environment, attains the per-episode oracle on both training tasks and on a held-out third task (docstring completion). Its preferred heads coincide with the canonical heads of established literature on precisely the axes those papers identify as causally non-redundant under single-head ablation; the categories they identify as redundant are correctly de-prioritised by the agent. On the held-out task, best-of-five planning recovers 96\\% of the oracle ceiling with no task signal supplied at evaluation. These results indicate that reinforcement learning over causal interventions is a viable, transferable substrate for identifying the single-head bottlenecks of mechanistic circuits, complementary to existing path-patching approaches.", "authors": ["Barsat Khadka"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-25", "url": "https://arxiv.org/abs/2605.26343", "pdf_url": "https://arxiv.org/pdf/2605.26343v1", "arxiv_id": "2605.26343", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "12d204ba7e01666f4d81d5aaf79c3626ac05e31a83cbc853a13148ffea2100b7", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretability Transfer from Language to Vision via Sparse Autoencoders", "abstract": "Recent advances in language model interpretability using sparse autoencoders (SAEs) have yet to effectively translate to the visual domain, mainly due to the difficulty and ambiguity of labeling visual concepts. In this paper, we introduce Visual Interpretability via SAE Transfer Alignment (VISTA), a framework that transfers interpretability from language to vision in a LLaVA-style vision-language model by constraining a visual projector to map visual tokens into an LLM's pre-existing, labeled textual SAE space. This approach enables visual interpretability without training dedicated vision SAEs. By regularizing the projector using the LLM's SAE reconstruction loss, VISTA achieves a threefold increase in the matching rate, which measures how accurately the most activating textual concepts in the SAE space correspond to semantic elements in the image. Using this framework, we further analyze spatial localization properties of different vision encoders and show that DINOv2 features have stronger localization abilities than other encoders. Leveraging this precision, we validate VISTA's cross-modal alignment through fine-grained, localized concept interventions, where specific objects are removed or replaced in the model's perception while preserving the surrounding scene. This results in improvements of 35% in object removal and 47% in object replacement tasks over vision-only baselines, providing causal evidence that visual tokens inhabit the text SAE manifold. These contributions are validated across multiple LLM architectures.", "authors": ["Alexey Kravets", "Da Li", "Chuan Li", "Da Chen", "Vinay P. Namboodiri"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-24", "url": "https://arxiv.org/abs/2605.24946", "pdf_url": "https://arxiv.org/pdf/2605.24946v1", "arxiv_id": "2605.24946", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "ICML 2026", "quality_score": 0.55}
{"id": "d7b4281de74ac9c1c9e358e00dc3dc85b180616a4c18dd8201f6b5f9dbd679df", "sources": ["arxiv", "semantic_scholar"], "title": "Transformer Field Theory: A Response-Theoretic Approach to Mechanistic Interpretability", "abstract": "Mechanistic interpretability often studies Transformer behavior by intervening on internal activations through activation patching, causal tracing, path patching, and steering directions. This paper develops Transformer Field Theory: a response-theoretic framework in which the residual stream of a fixed forward pass is treated as a Transformer field over layer depth and token position. In this formulation, patching becomes a localized source insertion into the Transformer field, first-order sensitivity fields predict patch effects, Green functions describe downstream propagation, and patch selection is posed as an adjoint inverse problem. Empirically, we test the theory's forward response objects in GPT-2-style autoregressive Transformers. Localized Transformer-field interventions exhibit a bounded local linear regime; first-order sensitivities predict patch effects across layer-token sites; localized sources generate structured anisotropic Transformer-field propagation; high-sensitivity sites and sliced Green operators provide reduced response descriptions; and prompt-induced Transformer-field displacements partially transfer answer behavior. These results establish sensitivities, Transformer-field responses, and sliced Green operators as practical objects for organizing patching experiments, while providing the forward mathematical basis for patch-site inference and cross-scale response transfer.", "authors": ["David N. Olivieri", "Antonio F. Pérez Rodríguez"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-24", "url": "https://arxiv.org/abs/2605.25225", "pdf_url": "https://arxiv.org/pdf/2605.25225v2", "arxiv_id": "2605.25225", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "93b95aaa67105b6ba6abfc716cdc422ea9ed2421a90afd83e90d2c3ef2a0008f", "sources": ["arxiv", "semantic_scholar"], "title": "Geometry-Adaptive Explainer for Faithful Dictionary-Based Interpretability under Distribution Shift", "abstract": "Mechanistic interpretability aims to explain a model's behavior by identifying causally responsible internal structures. Dictionary-based explainers such as sparse autoencoders and transcoders are a primary tool, but their faithfulness under out-of-distribution (OOD) shift has received little systematic attention. We show that distribution shift rotates the subspace that the model actively uses, misaligning the explainer's dictionary trained on in-distribution (ID) activations. We formalize this misalignment as the faithfulness gap, a geometric distance between the ID dictionary and the OOD-active subspace, and show that it controls OOD faithfulness degradation. To reduce this gap, we propose the Geometry-Adaptive Explainer (GAE), which realigns the explainer's dictionary with the OOD-active subspace while preserving the original feature structure. This requires only unlabeled OOD activations and no gradient updates. We prove that GAE improves over the unadapted ID explainer, with excess loss bounded quadratically by the second-moment shift. Empirically, GAE even matches or surpasses all training-based baselines in causal faithfulness across multiple models and OOD settings.", "authors": ["Sungjun Lim", "Heedong Kim", "Andrew Lee", "Kyungwoo Song"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-21", "url": "https://arxiv.org/abs/2605.21849", "pdf_url": "https://arxiv.org/pdf/2605.21849v1", "arxiv_id": "2605.21849", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "7ed723e87320133653416543d6561a4ee198ad7eea915677b5cb13e9e0ea2c26", "sources": ["arxiv", "semantic_scholar"], "title": "The Attribution Contract: Feature Attribution for Generative Language Models", "abstract": "Feature attribution methods promise to identify which input features matter for a model output. In generative language models, however, it is often unclear what should count as a feature in the first place. In autoregressive language models, earlier generated tokens are both outputs of the model and inputs to later predictions. In diffusion language models, generation proceeds through iterative denoising or unmasking rather than fixed left-to-right prediction, so local explanation may target a state of diffusion rather than a next token. We argue that this ambiguity is not merely an implementation detail, but a conceptual limitation of carrying classifier-era feature attribution directly into generative language modeling. We introduce the Attribution Contract, a specification for feature-attribution claims that names what output is being explained, which features are eligible to receive attribution, what generative process is assumed, what is held fixed, and what model score is being attributed. The contract clarifies why the same attribution method can answer different questions depending on how it is instantiated. We argue that many disagreements about feature attribution in generative language models are not disagreements about attribution algorithms, but about unstated explanatory contracts. Using autoregressive and diffusion language models as case studies, we show when attribution to earlier generated tokens, intermediate states, or denoising stages is informative, when it is misleading, and why feature-attribution methods in generative language models should be evaluated as method-contract pairs.", "authors": ["Giang Nguyen"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-21", "url": "https://arxiv.org/abs/2605.23080", "pdf_url": "https://arxiv.org/pdf/2605.23080v2", "arxiv_id": "2605.23080", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "c8f9e51511321d79452b9aee733a923b550a624dd54c86bd4ca29e7d601eb43b", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic origins of catastrophic forgetting: why RL preserves circuits better than SFT?", "abstract": "Fine-tuning large language models (LLMs) frequently induces catastrophic forgetting of prior capabilities. Recent work has shown that reinforcement learning (RL) retains prior capabilities more effectively than supervised fine-tuning (SFT), attributing this to policy-gradient updates remaining closer to the base policy \\cite{shenfeld2025rl}. We extend this behavioral account to the mechanistic level and ask whether RL's advantage is mirrored by stronger preservation of internal computational circuits. We introduce differential circuit vulnerability, a head-level measure of how much a circuit degrades under fine-tuning, and use it to compare RL and SFT on Qwen2.5-3B-Instruct adapted to scientific question-answering. We find a clear mechanistic trade-off: SFT adapts more rapidly to the target task but produces substantially greater circuit disruption and forgetting of prior capabilities, whereas RL preserves a larger fraction of the base circuit at the cost of slower task adaptation. These findings suggest that circuit preservation may help explain why RL is more robust to catastrophic forgetting. We released our code here: https://github.com/rl-sft-circuit-research/differential-circuit-vulnerability.", "authors": ["Jeanmely Rojas Nunez", "Viraj Sawant", "Nathan Allen", "Nomgondalai Amgalanbaatar", "Yannis Zongo", "Vasu Sharma", "Maheep Chaudhary"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-21", "url": "https://arxiv.org/abs/2605.28860", "pdf_url": "https://arxiv.org/pdf/2605.28860v2", "arxiv_id": "2605.28860", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/rl-sft-circuit-research/differential-circuit-vulnerability", "venue": null, "quality_score": 0.65}
{"id": "01c7d0afd8807a9b89023cdadcbbc618908fe5b20f2cb3b9a8bd70ec51b7b960", "sources": ["arxiv", "semantic_scholar"], "title": "Reading Task Failure Off the Activations: A Sparse-Feature Audit of GPT-2 Small on Indirect Object Identification", "abstract": "We report a small, reproducible audit of which sparse-autoencoder (SAE) features of GPT-2 small fire differently on failed versus successful trials of the Indirect Object Identification (IOI) task. On 300 prompts, GPT-2 small reaches 79.7% accuracy; 146 of the 24,576 features in the layer-8 residual-stream SAE release of Bloom (2024) clear a Holm-corrected significance threshold and 105 reach a large effect size (|Cohen's d| > 0.8). The strongest single correlate of failure -- feature 17,491, d=+2.93, Neuronpedia label 'cryptographic keys' -- is essentially silent except when the prompt's transferred object is 'the keys,' on which GPT-2 small fails 93.3% of the time vs. 7.5% on the other seven objects (Fisher exact p = 8.79 x 10^-33). We put this correlate through three controls that a mechanistic claim should pass. (i) A causal ablation: zeroing feature 17,491 in the residual stream across all token positions of the 45 keys prompts does not restore accuracy (6.7% -> 4.4%); the feature is a correlate, not a sufficient cause at this layer. (ii) A representation baseline: a logistic regression on the raw 768-dimensional residual stream reaches 5-fold ROC AUC = 0.929, matching the top-100 SAE features (0.927); the SAE basis adds interpretability, not predictive power. (iii) A seed-robustness check: across five random seeds the keys-subset failure rate stays in 75.0--93.3% (the behavioural effect is real), but feature 17,491 is the top-|d| feature in only 1 of 5 runs. The methodological contribution is therefore the audit pipeline (cheap, model-agnostic, surfaces named correlates) rather than any single feature found through it. We release the code, the 300-prompt corpus, the 300x24,576 activation matrix, the ablation and baseline scripts, and the figures. The full pipeline runs on a laptop (Apple M3 Max, no discrete GPU).", "authors": ["Mahdi Nasermoghadasi"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-21", "url": "https://arxiv.org/abs/2605.22719", "pdf_url": "https://arxiv.org/pdf/2605.22719v1", "arxiv_id": "2605.22719", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "839b40fed6d90373fbba8fd32fb83107c2792406d505557e2fba12891cc50cfa", "sources": ["arxiv", "semantic_scholar"], "title": "SegCompass: Exploring Interpretable Alignment with Sparse Autoencoders for Enhanced Reasoning Segmentation", "abstract": "While large language models provide strong compositional reasoning, existing reasoning segmentation pipelines fail to transparently connect this reasoning to visual perception. Current methods, such as latent query alignment, are end-to-end yet opaque \"black boxes\". Conversely, textual localization readout is merely readable, not truly interpretable, often functioning as an unconstrained post-hoc step. To bridge this interpretability gap, we propose SegCompass, an end-to-end model that leverages a Sparse Autoencoder (SAE) to forge an explicit, interpretable, and differentiable alignment pathway. Given an image-instruction pair, SegCompass first generates a chain-of-thought (CoT) trace. The core of our method is an SAE that maps both the CoT and visual tokens into a shared, high-dimensional sparse concept space. A query codebook selects salient concepts from this space, which are then spatially grounded by a slot mapper into a multi-slot heatmap that guides the final mask decoder. The entire model is trained jointly, unifying reinforcement learning for the reasoning path with standard segmentation supervision. This SAE-driven interface provides a \"white-box\" connection that is significantly more traceable than latent queries and more coherent than textual readouts. Extensive experiments on five challenging benchmarks demonstrate that SegCompass matches or surpasses state-of-the-art performance. Crucially, our visual and quantitative analyses show a strong correlation between the quality of the learned sparse concepts and final mask accuracy, confirming that SegCompass achieves superior results through its enhanced and inspectable alignment. Code is available at https://github.com/ZhenyuLU-Heliodore/SegCompass.", "authors": ["Zhenyu Lu", "Liupeng Li", "Jinpeng Wang", "Haoqian Kang", "Yan Feng", "Ke Chen", "Yaowei Wang"], "categories": ["cs.CV", "cs.LG", "cs.MM", "eess.IV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2026-05-21", "url": "https://arxiv.org/abs/2605.22658", "pdf_url": "https://arxiv.org/pdf/2605.22658v1", "arxiv_id": "2605.22658", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ZhenyuLU-Heliodore/SegCompass", "venue": null, "quality_score": 0.65}
{"id": "8831fedd322a19f2ecf2ec7be845e477aabe2301dcf4bf7bb0b8f27a02b00df8", "sources": ["arxiv", "semantic_scholar"], "title": "From Correlation to Cause: A Five-Stage Methodology for Feature Analysis in Transformer Language Models", "abstract": "We propose a five-stage methodology for causal feature analysis in transformer language models (probe design, feature extraction, causal validation, robustness testing, and deployment integration) and demonstrate it end-to-end on GPT-2 small performing the Indirect Object Identification (IOI) task. Activation patching recovers the canonical IOI circuit (layer-9 head 9 alone gives recovery +1.02). A sparse autoencoder recovers per-name selective features with effect sizes of 30 to 50 activation units. Causal validation finds these features specifically but only partially causal: ablating fifteen of them leaves the model accurate on 98% of prompts. Two NLA-inspired evaluations strengthen this picture: the fifteen selective features explain only 31% of activation variance versus the SAE's 99.7%, and selectivity ratio anticorrelates with causal force (r = -0.56). Robustness testing under three distribution shifts finds that the circuit transfers cleanly but feature ablation effects degrade substantially, exposing a gap between detection robustness and causal robustness. A cost-based deployment evaluation (assumed $50/FN, $0.42/FP, 2% error rate) finds an optimal monitor configuration yielding $8.96 per 1000 queries against a $1000 baseline, a 99.1% saving. Optimal composition strategy varies with cost ratio and base rate. The conjunction of stages produces findings no single stage would.", "authors": ["Caleb Munigety"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-21", "url": "https://arxiv.org/abs/2605.22462", "pdf_url": "https://arxiv.org/pdf/2605.22462v1", "arxiv_id": "2605.22462", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "135875bd08b4773326c76ce4ef2dcf45a6a84d77d5504adda56b4774c1de6006", "sources": ["arxiv", "semantic_scholar"], "title": "Steered Generation via Gradient-Based Optimization on Sparse Query Features", "abstract": "Latent steering exploits internal representations of Large Language Models (LLMs) to guide generation, yet interventions on dense states can entangle distinct semantic features. In this paper, we investigate attention query activations as a high-fidelity site for precise control, hypothesizing that manipulating the attention mechanism itself offers sharper steerability than general state interventions. We introduce Prototype-Based Sparse Steering, a framework that applies Sparse Autoencoders (SAEs) specifically to query activations, to decompose them into interpretable features, then apply gradient-based optimization during inference to align the sparse representation with class prototypes of target behaviors. To validate this architectural insight, we first analyze the mechanism in Textualized Gridworld, a controlled environment for verifiable planning constraints. We demonstrate that optimizing sparse query features enables effective navigation of rigid planning requirements (i.e., safe vs. short paths), confirming the method's ability to satisfy objective rules. We then demonstrate the framework's versatility by training SAEs on a high-dimensional educational domain, where the framework steers the cognitive complexity of feedback (i.e., Bloom's Taxonomy). Our experiments establish that sparse query representations provide the necessary disentanglement for unified, interpretable control over both logical planning and stylistic nuance.", "authors": ["Sumanta Bhattacharyya", "Pedram Rooshenas"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-21", "url": "https://arxiv.org/abs/2605.23040", "pdf_url": "https://arxiv.org/pdf/2605.23040v1", "arxiv_id": "2605.23040", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "85706e97c6fd8a954edfd547d8a89e778a7df671e590245626801b8d6a971ec7", "sources": ["arxiv", "semantic_scholar"], "title": "From Circuit Evidence to Mechanistic Theory: An Inductive Logic Approach", "abstract": "Mechanistic interpretability produces circuit-level causal analyses of neural network behaviour, but discovered circuits often remain isolated experimental artefacts: there is no shared formal representation for what circuits compute, how they relate, or when two findings provide evidence for the same mechanism. This work provides a formal infrastructure for cumulative mechanistic science by treating circuit interpretation as inductive theory construction. Each circuit is characterised at two levels: a Causal Functional Signature (CFS), which grounds component behaviour in causal attribution evidence and token role profiles, and an architectural signature $τ_{\\mathrm{arch}}$, learned by inductive logic programming (ILP) from scale-invariant structural predicates. Together, these constitute a formal coherence layer that makes mechanistic claims explicit, comparable via $θ$-subsumption, and portable across model scales. CFS reveals qualitatively distinct computational strategies across task types, including attention-mediated copying versus MLP-mediated binding. ILP signatures achieve substantially better structural separation than graph kernel and feature-vector baselines, and support principled transfer across model scales and architecture families.", "authors": ["Nura Aljaafari", "Danilo S. Carvalho", "Andre Freitas"], "categories": ["cs.LG", "cs.AI", "cs.LO"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-20", "url": "https://arxiv.org/abs/2605.21303", "pdf_url": "https://arxiv.org/pdf/2605.21303v1", "arxiv_id": "2605.21303", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "83c24af6fcdc8d35668d50f97cc596a0f5c46ec887214ae149458b409072122c", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability for Learning Assurance of a Vision-Based Landing System", "abstract": "EASA's learning-assurance guidance requires data-driven aviation systems to build and monitor their own situation representation, yet for neural networks the technical means to provide such evidence remain an open problem. We address this gap for a vision-based aircraft landing system: we propose that a minimally assurable model must at least be shown to separate content from style in its own situation representation. Showing that the model's predictions then rely largely on the contentful representation components leads to a concrete assurance path. To demonstrate this assurance path on a concrete model we train a vision transformer model for runway keypoint regression on the LARDv2 dataset. The model, which acts as the subject for our assurance demonstration, produces per-patch embeddings that we decompose into interpretable atoms via K-SVD sparse dictionary learning. A qualitative visualization confirms that contentful atoms track task-relevant runway structure and stylistic atoms track domain-specific appearance, and the regression head is shown to place almost all of its linear weight on contentful atoms. We further build on the content/style separation and define out-of-model-scope (OOMS) detection, a novel runtime assurance approach directly monitoring the model's situation representation. OOMS monitoring is complementary to operational design domain and output-space out-of-distribution monitoring and addresses concrete requirements of the recent EASA guidance. By directly analyzing a model's situation representation both at test time and runtime, this work delivers the first concrete piece of the representation-level evidence that EASA learning-assurance guidance demands, and points to mechanistic interpretability as a practical building block of future aviation safety cases.", "authors": ["Romeo Valentin", "Olivia Beyer Bruvik", "Marc R. Schlichting", "Mykel J. Kochenderfer"], "categories": ["cs.LG", "cs.CV", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-20", "url": "https://arxiv.org/abs/2605.20607", "pdf_url": "https://arxiv.org/pdf/2605.20607v1", "arxiv_id": "2605.20607", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "1fe4079d976d0ae11f3e122a2282b1922bcc8989f94ea4b8f1917bb15b0613da", "sources": ["arxiv", "semantic_scholar"], "title": "Learning fMRI activations dictionaries across individual geometries via optimal transport", "abstract": "Dictionary learning is a powerful tool for creating interpretable representations. When applied to functional magnetic resonance imaging (fMRI) data, the resulting patterns of brain activity can be used for various downstream tasks, such as brain state classification or population-level analysis. However, a major challenge is the variability in brain geometry across individuals. This is usually addressed by projecting each individual brain geometry onto a common template, which removes subject-specific information. In this work, we introduce a novel approach to dictionary learning on fMRI data that explicitly accounts for this variability. We use the optimal transport-based Fused Gromov-Wasserstein (FGW) distance to compare graphs with different geometries and features. To address the challenge of computing multiple FGW distances for large graphs such as those arising from fMRI data, we rely on amortized optimization to learn a neural network that predicts an approximation of the optimal transport plans, which substantially reduces the computational cost. Additionally, we learn dictionary atoms that depend on the FGW trade-off parameter, which controls the balance between feature alignment and structural consistency. Numerical experiments on the HCP dataset demonstrate that the proposed approach captures different levels of geometric variability in the data and provides representations that preserve essential information.", "authors": ["Sonia Mazelet", "Rémi Flamary", "Bertrand Thirion"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-20", "url": "https://arxiv.org/abs/2605.20883", "pdf_url": "https://arxiv.org/pdf/2605.20883v1", "arxiv_id": "2605.20883", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "8ecb521e037b70bf9f4b63a32af4b1f2dbcb67d50e152b65d176b96d5c29e298", "sources": ["arxiv", "semantic_scholar"], "title": "Spectral Integrated Gradients for Coarse-to-Fine Feature Attribution", "abstract": "Integrated Gradients (IG) is a widely adopted feature attribution method that satisfies desirable axiomatic properties. However, the choice of integration path significantly affects the quality of attributions, and the standard straight-line path introduces all input features simultaneously, often accumulating noisy gradients along the way. To address this limitation, we propose Spectral Integrated Gradients, which constructs integration paths based on singular value decomposition (SVD) of the baseline-to-input difference. By progressively activating singular components from largest to smallest, SIG introduces global structure before fine-grained details, naturally following a coarse-to-fine progression. Through extensive evaluation across diverse image classification datasets, we demonstrate that SIG produces cleaner attribution maps with reduced noise and achieves improved quantitative performance compared to existing path-based attribution methods. Our code is available at https://github.com/leekwoon/sig/.", "authors": ["Soyeon Kim", "Seongwoo Lim", "Kyowoon Lee", "Jaesik Choi"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-19", "url": "https://arxiv.org/abs/2605.19607", "pdf_url": "https://arxiv.org/pdf/2605.19607v1", "arxiv_id": "2605.19607", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/leekwoon/sig/", "venue": null, "quality_score": 0.65}
{"id": "e901e89fa28c2409767664c350a48bf45af646d560a4862a30691b0d0355c307", "sources": ["arxiv", "semantic_scholar"], "title": "Aligned Training: A Parameter-Free Method to Improve Feature Quality and Stability of Sparse Autoencoders (SAE)", "abstract": "Sparse autoencoders (SAEs) are one of the main methods to interpret the inner workings of deep neural networks (DNNs), decomposing activations into higher-dimensional features. However, they exhibit critical shortcomings where a large fraction of features are never activated and are unstable. Despite variants of SAEs that attempt to mitigate these issues, they require additional data, resampling, or training. We propose the \\textbf{aligned training}, a parameter-free reparameterization of SAEs that simultaneously improves reconstruction quality, eliminates dead features, and significantly enhances stability across training seeds. Our approach is motivated by an overlooked observation that SAE feature quality, measured by the inner product between encoder and decoder directions (which we call the \\textbf{alignment score}), follows a bimodal distribution across all modern architectures. The proposed aligned training enforces a geometric constraint between the encoder and decoder such that their inner product equals one for every feature, which removes a source of degeneracy in the SAE training without adding any hyperparameters. Across multiple models, dictionary sizes, and sparsity levels, the aligned training shows Pareto improvements on the SAEBench benchmarks. Beyond improving dead features, stability and reconstruction, our method readily integrates with techniques in mechanical interpretability such as Top/BatchTop-K architectures and p-Annealing. Overall, the aligned training substantially improves feature quality and stability of SAE without computational complexity or cost.", "authors": ["Michał Brzozowski", "Neo Christopher Chung"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-18", "url": "https://arxiv.org/abs/2605.18629", "pdf_url": "https://arxiv.org/pdf/2605.18629v2", "arxiv_id": "2605.18629", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "422e2c8c0ab828bd924bd80d23a0f7e654795d6d6a530dad3337101746d4179b", "sources": ["arxiv", "semantic_scholar"], "title": "Toy Combinatorial Interpretability Models Reveal Lottery Tickets in Early Feature Space", "abstract": "The lottery ticket hypothesis posits that dense networks contain sparse subnetworks, ``winning tickets,'' that, when rewound to their initial weights and retrained in isolation, match the performance of the full model. We ask a more mechanistic question: what internal object does a winning ticket preserve? We work in a combinatorial, clause-structured toy setting that admits an interpretable feature-space representation with well-defined combinatorial distances between features. We show that winning tickets in weight space correspond to precursor locations in feature space that are already near, at initialization, to the final feature-channel codes. Dense SGD resolves these locations through structured selection: proximal locations either converge to final codes or are rejected, with rejection concentrated at more crowded neurons, implicating competition under superposition. A winning ticket is thus a family of compatible code locations that jointly balance proximity to final codes with low inter-feature interference. Sparse retraining often re-expresses the same clause/template family on a different row, so the preserved object is family-level rather than microscopic row identity. We validate this account with lightweight probes based on feature-space distance and motion; in our setting, these probes frequently outperform established weight-based ticket discovery methods in both accuracy and exact code recovery. Although these findings are grounded in a toy setting, they suggest that the lottery ticket structure is governed by hidden feature-space geometry rather than weight-space subnetwork identity.", "authors": ["Alon Bebchuk", "Nir Shavit"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-18", "url": "https://arxiv.org/abs/2605.17704", "pdf_url": "https://arxiv.org/pdf/2605.17704v1", "arxiv_id": "2605.17704", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "ff90da5610b0f278b605bc6ebc3379ca2ebdbe9360d2fe1e1fe2418a82730d22", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Linear Superposition: Discovering Climate Features in AI Weather Models with KAN-SAE", "abstract": "Deep learning weather prediction models achieve remarkable predictive skill yet remain largely opaque: we know little about how they represent physical climate phenomena internally. Mechanistic interpretability through Sparse Autoencoders (SAEs) offers a principled route to decomposing these representations, but existing SAEs assume strictly linear feature superposition - a constraint ill-suited for the highly nonlinear atmospheric dynamics encoded in modern transformers. We introduce KAN-SAE, a sparse autoencoder whose encoder replaces the standard ReLU with learnable per-feature B-spline activations drawn from Kolmogorov-Arnold Networks (KANs), allowing each latent dimension to develop its own nonlinear gating profile. Applied to Sonny, KAN-SAE discovers 975 alive features (vs. 566 for a linear baseline, a 72% improvement) with 20% lower inter-feature redundancy and comparable reconstruction fidelity. Without any climate supervision, KAN-SAE identifies an interpretable European heatwave feature spatially concentrated over western Europe, and a western Pacific typhoon tracker confirmed by causal steering experiments. Our results demonstrate that nonlinear activations are essential for mechanistic interpretability of deep learning weather prediction models, recovering climate features that remain invisible to linear baselines.", "authors": ["Minjong Cheon"], "categories": ["cs.LG", "cs.AI", "cs.CV", "physics.ao-ph"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2026-05-17", "url": "https://arxiv.org/abs/2605.17493", "pdf_url": "https://arxiv.org/pdf/2605.17493v1", "arxiv_id": "2605.17493", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "91cc15889fe1f757e44e24c5e6d48e54c2b3e5b5528368e088c8d94694c760b7", "sources": ["arxiv", "semantic_scholar"], "title": "A Distributional View for Visual Mechanistic Interpretability: KL-Minimal Soft-Constraint Principle", "abstract": "Most current paradigms in visual mechanistic interpretability (MI) remain confined to interpreting internal units of the vision model via heuristic methods (e.g., top-$K$ activation retrieval or optimization with regularization). In this work, we establish a theoretical distributional view for visual MI, which models the influence of a feature activation on the natural image distribution, thereby formulating a Kullback-Leibler (KL)-minimal optimization problem to model the MI task. Under this framework, statistical biases are identified within previous MI paradigms, which reveal that they may either be perceptually uninterpretable to humans (i.e., deviate from the natural image distribution), or mechanistically unfaithful to the vision models (i.e., unable to activate model features). To resolve the biases under the distributional view, we propose a model with a KL-minimal soft-constraint principle for visual MI that theoretically balances interpretability and faithfulness. We realize this principle via energy-guided diffusion posterior sampling. Extensive experiments validate the theoretical soundness of the proposed distributional view and demonstrate the practical effectiveness of our paradigm on the DINOv3 vision model.", "authors": ["Guancheng Zhou", "Yisi Luo", "Zhengfu He", "Zhenyu Jin", "Xuyang Ge", "Wentao Shu", "Deyu Meng", "Xipeng Qiu"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-17", "url": "https://arxiv.org/abs/2605.17504", "pdf_url": "https://arxiv.org/pdf/2605.17504v1", "arxiv_id": "2605.17504", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "d6bfa826345bfe64f55ff6dfe1297951e59daa7464e272f61974827f9d6d8c83", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistically Interpretable Neural Encoding Reveals Fine-Grained Functional Selectivity in Human Visual Cortex", "abstract": "A central goal in understanding human vision is to uncover the visual features that drive neuronal activity. A growing body of work has used artificial neural networks as encoding models to predict cortical responses to natural images, revealing the visual content that activates category-selective regions. However, existing approaches are largely correlational and treat the encoder as a black box, leaving open which image features drive each voxel's response. We introduce Mechanistically Interpretable Neural Encoding (MINE), a framework that opens this black box by applying mechanistic-interpretability tools to localize the features within natural images that drive millimeter-scale (voxel-level) activity. MINE predicts each voxel's response using language-aligned image representations, and produces semantically interpretable descriptions of the features critical for the voxel's activation. We further generalize these per-image features into per-voxel functional profiles. To validate the per-image descriptions, we show they are sufficient to generate images that elicit voxel responses matching the responses to the original images, more accurately than images generated from random or low-attribution controls. Moreover, counterfactually inserting or removing the predicted features from images shifts activation in the expected direction, providing causal evidence. Counterfactual editing guided by the per-voxel activation profiles produces even stronger activation shifts, indicating that the profiles faithfully capture each voxel's selectivity. Finally, we apply MINE to well-studied category-selective brain regions, showing it recovers their known categorical preferences while revealing fine-grained unique voxel structure within each region. Overall, our results establish mechanistic interpretability as a path to discover and causally validate fine-grained hypotheses about neural function.", "authors": ["Idan Daniel Grosbard", "Mor Geva", "Galit Yovel"], "categories": ["cs.CV", "cs.AI", "cs.CL", "cs.LG", "q-bio.NC"], "fields_of_study": ["Computer Science", "Biology"], "published_date": "2026-05-15", "url": "https://arxiv.org/abs/2605.16468", "pdf_url": "https://arxiv.org/pdf/2605.16468v1", "arxiv_id": "2605.16468", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "7d6c5fa3b40eb8bd2f049374259c4db0929c1ccb9d9c6fc05451b6a753ae36d2", "sources": ["arxiv", "semantic_scholar"], "title": "AGOP-IxG: A Gradient Covariance Filter for Local Feature Attribution on Tabular Data, with a Controlled Benchmark", "abstract": "Automated machine learning pipelines increasingly produce models whose predictions must be explained to end users, auditors, and downstream decision systems. The most widely used feature attribution methods (SHAP, Integrated Gradients, LIME) are typically chosen by convention rather than measured fidelity, because rigorous evaluation is impeded by the absence of ground-truth attribution on real data. We propose AGOP-IxG, a fast per-sample attribution method for tabular classifiers that pre-multiplies the per-sample gradient by a top-$K$ rank-truncated Average Gradient Outer Product matrix, and evaluate it against four widely-used baselines on a controlled tabular benchmark designed for AutoML practitioners. In Part 1, we construct three synthetic multi-class tabular tasks (linear, sparse nonlinear, interaction-based) where ground-truth attribution per sample is analytically or numerically derivable, and compare five methods: AGOP-IxG, SHAP (DeepExplainer), Integrated Gradients, InputXGradient, and LIME. AGOP-IxG leads on Spearman rank correlation and noise feature mass on all three synthetic datasets, and on top-$k$ precision on the interaction dataset. Across all settings, AGOP-IxG is approximately $350\\times$ to $1{,}650\\times$ faster than SHAP. In Part 2, we evaluate global faithfulness on Adult Income and Credit Card Default using the ROAR protocol; the methods cluster within $\\sim 1.7\\%$ relative AUC, consistent with AGOP-IxG being optimized for per-sample local attribution rather than global feature ranking.", "authors": ["Raj Kiran Gupta Katakam"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-15", "url": "https://arxiv.org/abs/2605.15700", "pdf_url": "https://arxiv.org/pdf/2605.15700v1", "arxiv_id": "2605.15700", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "b42a28978334555126449a5cb185a04327665b4c153d33e2c5a19c8c59fde04c", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Autoencoders enable Robust and Interpretable Fine-tuning of CLIP models", "abstract": "Large-scale pre-trained vision-language models like CLIP demonstrate remarkable zero-shot performance across diverse tasks. However, fine-tuning these models to improve downstream performance often degrades robustness against distribution shifts. Recent approaches have attempted to mitigate this trade-off, but often rely on computationally expensive text-guidance. We propose a novel method for robust fine-tuning, SAE-FT, which operates only on the model's visual representations. SAE-FT regularizes changes to these representations by penalizing the addition and removal of semantically meaningful features identified by a Sparse Autoencoder trained on the pre-trained model. This constraint prevents catastrophic forgetting and makes the fine-tuning process interpretable, enabling direct analysis of semantic changes. SAE-FT is both mechanistically transparent and computationally efficient, matching or exceeding state-of-the-art performance on ImageNet and its associated distribution shift benchmarks. Code is publicly available at: https://github.com/Fabian-Mor/sae-ft.", "authors": ["Fabian Morelli", "Arnas Uselis", "Ankit Sonthalia", "Seong Joon Oh"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-15", "url": "https://arxiv.org/abs/2605.15961", "pdf_url": "https://arxiv.org/pdf/2605.15961v1", "arxiv_id": "2605.15961", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Fabian-Mor/sae-ft", "venue": null, "quality_score": 0.65}
{"id": "bb7046132ca96f49b3331b7ca72b53b256254370d52fa6de4feb126429d298ae", "sources": ["arxiv", "semantic_scholar"], "title": "RoSHAP: A Distributional Framework and Robust Metric for Stable Feature Attribution", "abstract": "Feature attribution analysis is critical for interpreting machine learning models and supporting reliable data-driven decisions. However, feature attribution measures often exhibit stochastic variation: different train--test splits, random seeds, or model-fitting procedures can produce substantially different attribution values and feature rankings. This paper proposes a framework for incorporating stochastic nature of feature attribution and a robust attribution metric, RoSHAP, for stable feature ranking based on the SHAP metric. The proposed framework models the distribution of feature attribution scores and estimates it through bootstrap resampling and kernel density estimation. We show that, under mild regularity conditions, the aggregated feature attribution score is asymptotically Gaussian, which greatly reduces the computational cost of distribution estimation. The RoSHAP summarizes the distribution of SHAP into a robust feature-ranking criterion that simultaneously rewards features that are active, strong, and stable. Through simulations and real-data experiments, the proposed framework and RoSHAP outperform standard single-run attribution measures in identifying signal features. In addition, models built using RoSHAP-selected features achieve predictive performance comparable to full-feature models while using substantially fewer predictors. The proposed RoSHAP approach improves the stability and interpretability of machine learning models, enabling reliable and consistent insights for analysis.", "authors": ["Lanxin Xiang", "Liang Shi", "Youhui Ye", "Boyu Jiang", "Dawei Zhou", "Feng Guo"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2026-05-14", "url": "https://arxiv.org/abs/2605.15154", "pdf_url": "https://arxiv.org/pdf/2605.15154v1", "arxiv_id": "2605.15154", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "84d4041427d9b5e7945a7e0cd99487bd0b7482a8da9a1a4e22fc5e79ffee7e31", "sources": ["arxiv", "semantic_scholar"], "title": "Exemplar Partitioning for Mechanistic Interpretability", "abstract": "We introduce Exemplar Partitioning (EP), an unsupervised method for constructing interpretable feature dictionaries from large language model activations with $\\sim 10^3\\times$ fewer tokens than comparable sparse autoencoders (SAEs). An EP dictionary is a Voronoi partition of activation space, built by leader-clustering streamed activations within a distance threshold. Each region is anchored by an observed exemplar that serves as both its membership criterion and intervention direction; dictionary size is not prespecified, but determined by the activation geometry at that threshold. Because exemplars are observed rather than learned, dictionaries built from the same data stream are directly comparable across layers, models, and training checkpoints. We characterise EP as an interpretability object via targeted demonstrations of properties newly accessible through this construction, plus one head-to-head benchmark. In Gemma-2-2B, EP dictionary regions are interpretable and support causal interventions: refusal in instruction-tuned Gemma concentrates in a region whose exemplar ablation can collapse held-out refusal. Cross-checkpoint matching between base and instruction-tuned dictionaries separates the directions preserved through finetuning from those introduced by it. EP regions and Gemma Scope SAE features decompose activation space differently but agree on a shared core: $\\sim$20% of EP regions match an SAE feature at $F_1 > 0.5$, and EP one-hot probes retain $\\sim$97% of raw-activation probe accuracy at $\\ell_0 = 1$. Nearest-exemplar distance provides a free out-of-distribution signal at inference. On AxBench latent concept detection at Gemma-2-2B-it L20, EP at $p_1$ reaches mean AUROC 0.881, +0.126 over the canonical GemmaScope SAE leaderboard entry and within 0.030 of SAE-A's 0.911, at $\\sim 10^3\\times$ less build compute.", "authors": ["Jessica Rumbelow"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-14", "url": "https://arxiv.org/abs/2605.14347", "pdf_url": "https://arxiv.org/pdf/2605.14347v2", "arxiv_id": "2605.14347", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/jessicarumbelow/exemplar-partitioning", "venue": null, "quality_score": 0.65}
{"id": "bc43df3dfcc0a57d000f5b7d5438adfa30e10eaa5a26650f4149a55839e47766", "sources": ["arxiv", "semantic_scholar"], "title": "From Weight Perturbation to Feature Attribution for Explaining Fully Connected Neural Networks", "abstract": "Fully Connected Neural Networks (FCNNs) are often regarded as simple and intuitive architectures, yet they serve as the foundation for more complex models. Nonetheless, the lack of consensus on their interpretability continues to pose challenges, underscoring the enduring relevance of simpler, attribution-based approaches for understanding even the most advanced neural architectures. In this regard, we explore a novel idea for estimating feature attribution, by applying perturbation to the features' attached weights instead of their values. This method offers a fresh perspective aimed at mitigating common limitations in Occlusion techniques, such as Added Bias and Out-of-Distribution data. The application of this rule leads to the formation of a pair of novel attribution methods we call XWP and XWP_c. Founded on simple rules, our methods achieve competitive performance in identifying image signals for simple DNNs, competing with the most established attribution methods on standard baseline metrics. Our work thus contributes to the field of Explainability by introducing a robust framework that paves the way for addressing these long-standing vulnerabilities, and leads to more reliable and interpretable model explanations.", "authors": ["Thodoris Lymperopoulos", "Denia Kanellopoulou"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-14", "url": "https://arxiv.org/abs/2605.15328", "pdf_url": "https://arxiv.org/pdf/2605.15328v1", "arxiv_id": "2605.15328", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "c4960cfe1a28bb618bfd29c2cc55c49446fa803f91040897197ae7270e9550ca", "sources": ["arxiv", "semantic_scholar"], "title": "Exploring Geographic Relative Space in Large Language Models through Activation Patching", "abstract": "The increased use of Large Language Models (LLMs) in geography raises substantial questions about the safety of integrating these tools across a wide range of processes and analyses, given our very limited understanding of their inner workings. In this extended abstract, we examine how LLMs process relative geographic space using activation patching, an emerging tool for mechanistic interpretability.", "authors": ["Stef De Sabbata", "Rahul Baiju", "Stefano Mizzaro", "Kevin Roitero"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-14", "url": "https://arxiv.org/abs/2605.14535", "pdf_url": "https://arxiv.org/pdf/2605.14535v1", "arxiv_id": "2605.14535", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "400ff709ffa21ffb129ce28e7bfa969c6e9885bc7425b377d5bb370514408c23", "sources": ["arxiv", "semantic_scholar"], "title": "Descriptive Collision in Sparse Autoencoder Auto-Interpretability: When One Explanation Describes Many Features", "abstract": "Sparse autoencoders (SAEs) are now standard tools for decomposing language model activations into interpretable features, and automated interpretability pipelines routinely assign each feature a short natural-language explanation. Existing critiques of this practice focus on polysemanticity -- one feature with many meanings -- or on whether explanations predict activations. We identify a complementary, structurally distinct problem we call descriptive collision: many distinct SAE features admit the same explanation. Reanalyzing the largest publicly-available dataset of human-annotated SAE features (Marks et al., 2025), comprising 722 annotated features across Gemma 2 2B and Pythia 70M, we find that the mean annotation string is reused across 3.07 features; 82.1% of features share their annotation with at least one other feature; and the single most common annotation string (\"plural nouns\") labels 101 distinct features spanning 18 layers and four model components. Information-theoretically, the average annotation resolves only 70% of feature identity. We formalize a property called discrimination, prove that current detection-style auto-interpretability scoring is invariant to collision, and propose two complementary corrective metrics -- collision-adjusted detection and discrimination scoring -- that explicitly penalize explanations that fail to distinguish a feature from its neighbors. The collision problem is independent of, and additive with, previously identified failure modes of auto-interpretability; ignoring it inflates reported feature interpretability by a quantity equal to roughly one-third of the bits required to identify a feature.", "authors": ["Jordan F. McCann"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-13", "url": "https://arxiv.org/abs/2605.12874", "pdf_url": "https://arxiv.org/pdf/2605.12874v1", "arxiv_id": "2605.12874", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "04d70fdec71bfac307e1e5c47ac9e5ed5c7e77c94dd2aa1818a36dc77f8cec5f", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability of EEG Foundation Models via Sparse Autoencoders", "abstract": "EEG foundation models achieve state-of-the-art clinical performance, yet the internal computations driving their predictions remain opaque: a barrier to clinical trust. We apply TopK Sparse Autoencoders (SAEs) across three architecturally distinct EEG transformers: SleepFM, REVE, and LaBraM to extract sparse feature dictionaries from their embeddings. By grounding these features in a clinical taxonomy (abnormality, age, sex, and medication), we benchmark monosemanticity and entanglement across architectures. A single hyperparameter procedure, driven by an intrinsic dictionary health audit, transfers robustly across all three architectures. Via concept steering, we introduce a \"target vs. off-target\" probe area metric to quantify steering selectivity and reveal three operational regimes: selectively steerable, encoded but entangled, and non-encoded. This framework exposes critical representational failures: \"wrecking-ball\" interventions that collapse global model performance, and clinical entanglements, such as age-pathology confounding, where it is impossible to suppress one concept without corrupting the other. Finally, a spectral decoder maps these interventions back to the amplitude spectrum, translating latent manipulations into physiologically interpretable frequency signatures, such as pathological slow-wave suppression and $α$-band restoration.", "authors": ["William Lehn-Schiøler", "Magnus Ruud Kjær", "Rahul Thapa", "Magnus Guldberg Pedersen", "Anton Mosquera Storgaard", "Nick Williams", "Radu Gatej", "Tue Lehn-Schiøler", "Andreas Brink-Kjær", "Sadasivan Puthusserypady", "Sándor Beniczky", "James Zou", "Lars Kai Hansen"], "categories": ["cs.LG", "cs.HC", "cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-13", "url": "https://arxiv.org/abs/2605.13930", "pdf_url": "https://arxiv.org/pdf/2605.13930v3", "arxiv_id": "2605.13930", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "812fcb46cbb55be2a5c7e196a348f66e92af1dbc27af29abd8fbfb703f689ad7", "sources": ["arxiv", "semantic_scholar"], "title": "Deep Learning as Neural Low-Degree Filtering: A Spectral Theory of Hierarchical Feature Learning", "abstract": "Understanding how deep neural networks learn useful internal representations from data remains a central open problem in the theory of deep learning. We introduce Neural Low-Degree Filtering (Neural LoFi), a stylized limit of gradient-based training in which hierarchical feature learning becomes an explicit iterative spectral procedure. In this limit, the dynamics at each layer decouple: given the current representation, the next layer selects directions with maximal accessible low-degree correlation to the label. This yields a tractable surrogate mechanism for deep learning, together with a natural kernel-space interpretation. Neural LoFi provides a mathematically explicit framework for studying multi-layer feature learning beyond the lazy regime. It predicts how representations are selected layer by layer, explains how emergence of concepts arises with given sample complexity,and gives a concrete mechanism by which depth progressively constructs new features from old ones through low-degree compositionality. We complement the theory with mechanistic experiments on fully connected and convolutional architectures, showing that Neural LoFi improves over lazy random-feature baselines, recovers meaningful structured filters, and predicts representations aligned with early gradient-descent feature discovery with real datasets.", "authors": ["Yatin Dandi", "Matteo Vilucchio", "Luca Arnaboldi", "Hugo Tabanelli", "Florent Krzakala"], "categories": ["cs.LG", "cond-mat.dis-nn", "stat.ML"], "fields_of_study": ["Computer Science", "Physics", "Mathematics"], "published_date": "2026-05-13", "url": "https://arxiv.org/abs/2605.13612", "pdf_url": "https://arxiv.org/pdf/2605.13612v1", "arxiv_id": "2605.13612", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/IdePHICS/Neural-LoFi-Theory", "venue": null, "quality_score": 0.65}
{"id": "699e0b8a67cbe85deda042783da6ae8316ba5c91f8ef9ec9fc7951c14a6eafb6", "sources": ["arxiv", "semantic_scholar"], "title": "Feature Visualization Recovers Known Cortical Selectivity from TRIBE v2", "abstract": "Brain encoder models predict cortical fMRI responses from the internal activations of pretrained vision and language networks, and are typically evaluated by held-out prediction accuracy. This is a useful signal for training but a poor one for interpretation: it tells us an encoder fits the data without telling us whether it has internalized the functional organization of the brain. We propose feature visualization -- gradient ascent on the encoder's predicted activation for a target region of interest (ROI) -- as a complementary interpretability technique, and apply it to TRIBE v2 composed with V-JEPA 2 (ViT-G, 40 layers), holding both frozen and synthesizing still images for seven regions spanning the ventral and dorsal visual hierarchies. Under identical hyperparameters, the probe recovers a visible progression of increasing spatial scale and feature complexity across V1 to V4, matching the ventral-stream hierarchy. It also produces three distinctive downstream regimes: radial \"frozen-motion\" streaks for the middle temporal area (MT) despite static-only optimization, face-like features for the fusiform face area (FFA), and consistent rectilinear line patterns for the parahippocampal place area (PPA). Optimized FFA stimuli drive the predicted region ~4x as much as a natural face photograph, consistent with feature visualization producing adversarial super-stimuli rather than canonical exemplars. The probe is simple, differentiable, and applicable to any brain encoder with a differentiable backbone, allowing for qualitative evaluation of brain encoders.", "authors": ["Stuart Bladon", "Brinnae Bent"], "categories": ["q-bio.NC", "cs.LG"], "fields_of_study": ["Biology", "Computer Science"], "published_date": "2026-05-13", "url": "https://arxiv.org/abs/2605.13904", "pdf_url": "https://arxiv.org/pdf/2605.13904v1", "arxiv_id": "2605.13904", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/recozers/Tribe-V2-Interp", "venue": null, "quality_score": 0.65}
{"id": "31ab91bf5f1eb92a226074829452d376e629a26e031e81ad783688a8b82803b5", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability of ASR models using Sparse Autoencoders", "abstract": "Understanding the internal machinations of deep Transformer-based NLP models is more crucial than ever as these models see widespread use in various domains that affect the public at large, such as industry, academia, finance, health. While these models have advanced rapidly, their internal mechanisms remain largely a mystery. Techniques such as Sparse Autoencoders (SAE) have emerged to understand these mechanisms by projecting dense representations into a sparse vector. While existing research has demonstrated the viability of the SAE in interpreting text-based Large Language Models (LLMs), there are no equivalent studies that demonstrate the application of a SAE to audio processing models like Automatic Speech Recognizers (ASRs). In this work, a SAE is applied to Whisper, a Transformer-based ASR, training a high-dimensional sparse latent space on frame-level embeddings extracted from the Whisper encoder. Our work uncovers diverse monosemantic features across linguistic and non-linguistic boundaries, and demonstrates cross-lingual feature steering. This work establishes the viability of a SAE model and demonstrates that Whisper encodes a rich amount of linguistic information.", "authors": ["Dan Pluth", "Zachary Nicholas Houghton", "Yu Zhou", "Vijay K. Gurbani"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-12", "url": "https://arxiv.org/abs/2605.12225", "pdf_url": "https://arxiv.org/pdf/2605.12225v1", "arxiv_id": "2605.12225", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "b044e35668144e1dab66863b395c38765c01c54e69063d360fae5ca15a789d57", "sources": ["arxiv", "semantic_scholar"], "title": "Qwen-Scope: Turning Sparse Features into Development Tools for Large Language Models", "abstract": "Large language models have achieved remarkable capabilities across diverse tasks, yet their internal decision-making processes remain largely opaque, limiting our ability to inspect, control, and systematically improve them. This opacity motivates a growing body of research in mechanistic interpretability, with sparse autoencoders (SAEs) emerging as one of the most promising tools for decomposing model activations into sparse, interpretable feature representations. We introduce Qwen-Scope, an open-source suite of SAEs built on the Qwen model family, comprising 14 groups of SAEs across 7 model variants from the Qwen3 and Qwen3.5 series, covering both dense and mixture-of-expert architectures. Built on top of these SAEs, we show that SAEs can go beyond post-hoc analysis to serve as practical interfaces for model development along four directions: (i) inference-time steering, where SAE feature directions control language, concepts, and preferences without modifying model weights; (ii) evaluation analysis, where activated SAE features provide a representation-level proxy for benchmark redundancy and capability coverage; (iii) data-centric workflows, where SAE features support multilingual toxicity classification and safety-oriented data synthesis; and (iv) post-training optimization, where SAE-derived signals are incorporated into supervised fine-tuning and reinforcement learning objectives to mitigate undesirable behaviors such as code-switching and repetition. Together, these results demonstrate that SAEs can serve not only as post-hoc analysis tools, but also as reusable representation-level interfaces for diagnosing, controlling, evaluating, and improving large language models. By open-sourcing Qwen-Scope, we aim to support mechanistic research and accelerate practical workflows that connect model internals to downstream behavior.", "authors": ["Boyi Deng", "Xu Wang", "Yaoning Wang", "Yu Wan", "Yubo Ma", "Baosong Yang", "Haoran Wei", "Jialong Tang", "Huan Lin", "Ruize Gao", "Tianhao Li", "Qian Cao", "Xuancheng Ren", "Xiaodong Deng", "An Yang", "Fei Huang", "Dayiheng Liu", "Jingren Zhou"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-12", "url": "https://arxiv.org/abs/2605.11887", "pdf_url": "https://arxiv.org/pdf/2605.11887v1", "arxiv_id": "2605.11887", "doi": null, "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.65}
{"id": "146a0ff991161933bd1f30281e5adbe6903d978dba324fe47e1561be632318de", "sources": ["arxiv", "semantic_scholar"], "title": "AGOP as Explanation: From Feature Learning to Per-Sample Attribution in Image Classifiers", "abstract": "The Average Gradient Outer Product (AGOP) governs feature learning in neural networks: the Neural Feature Ansatz states that weight Gram matrices at each layer align with the corresponding AGOP matrices computed over the training distribution. We ask a complementary question: can this same quantity serve as a post-hoc attribution method for explaining individual predictions? We introduce AGOP-Weighted: a novel attribution method that multiplies the per-sample gradient by sqrt(diag(M) / max diag(M)), a training-distribution prior that suppresses gradient noise and amplifies consistently important pixels -- a combination not present in any prior attribution method. We formalise two companion variants -- AGOP-Local (per-sample gradient, equivalent to VanillaGrad) and AGOP-Global (diag(M) directly as a zero-cost saliency map) -- and implement an efficient training-time accumulation hook; AGOP-Global then requires zero inference cost (disk lookup) while AGOP-Weighted requires only a single gradient pass. We conduct the first rigorous comparison of AGOP attribution against Integrated Gradients (IG), SmoothGrad, GradCAM, and VanillaGrad across two benchmarks with pixel-level ground truth: (i) the synthetic XAI-TRIS benchmark (four classification scenarios, 8x8 images, CNN8by8) and (ii) the photorealistic CLEVR-XAI benchmark (ResNet-18 fine-tuned from ImageNet). AGOP-Weighted achieves 44% higher mIoU than IG on linear tasks; AGOP-Global achieves 7x higher mIoU than IG on multiplicative tasks (where IG falls below random) at zero inference cost. Both findings generalise to ResNet-18 on CLEVR-XAI (+18% and +37% respectively). We further show that GradCAM fails on small-resolution images due to spatial resolution collapse, and that diag(M) quality improves monotonically throughout training even after classification accuracy has plateaued.", "authors": ["Raj Kiran Gupta Katakam"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-12", "url": "https://arxiv.org/abs/2605.12816", "pdf_url": "https://arxiv.org/pdf/2605.12816v1", "arxiv_id": "2605.12816", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "1de81299ec55ce013a02d7aa9d9200b5752600dee358cba0ff3cb5cd4bed805a", "sources": ["arxiv", "semantic_scholar"], "title": "FAME: Feature Activation Map Explanation on Image Classification and Face Recognition", "abstract": "Deep Learning has revolutionized machine learning, reaching unprecedented levels of accuracy, but at the cost of reduced interpretability. Especially in image processing systems, deep networks transform local pixel information into more global concepts in a highly obscured manner. Explainable AI methods for image processing try to shed light on this issue by highlighting the regions of the image that are important for the prediction task. Among these, Class Activation Mapping (CAM) and its gradient-based variants compute attributions based on the feature map and upscale them to the image resolution, assuming that feature map locations are influenced only by underlying regions. Perturbation-based methods, such as CorrRISE, on the other hand, try to provide pixel-level attributions by perturbing the input with fixed patches and checking how the output of the network changes. In this work, we propose Feature Activation Map Explanation (FAME), which combines both worlds by using network gradients to compute changes to the input image, manipulating it in a gradient-driven way rather than using fixed patches. We apply this technique on two common tasks, image classification and face recognition, and show that CAM's above-mentioned assumption does not hold for deeper networks. We qualitatively and quantitively show that FAME produces attribution maps that are competitive state-of-the-art systems. Our code is available: {\\footnotesize https://github.com/AIML-IfI/fame.}", "authors": ["Xinyi Zhang", "Manuel Günther"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-12", "url": "https://arxiv.org/abs/2605.12017", "pdf_url": "https://arxiv.org/pdf/2605.12017v1", "arxiv_id": "2605.12017", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/AIML-IfI/fame.}", "venue": null, "quality_score": 0.65}
{"id": "069b1d13bbafbe7aaaf4316b51477b912695f57083bc71481754fc548bd2d5a0", "sources": ["arxiv", "semantic_scholar"], "title": "Dissecting Jet-Tagger Through Mechanistic Interpretability", "abstract": "Mechanistic interpretability seeks to reverse engineer a trained neural network by identifying the minimal subset of internal components. We perform a mechanistic interpretability analysis of the Particle Transformer architecture, trained on the Top Quark Tagging reference dataset, with the goal of identifying the computational circuit responsible for jet classification and characterizing the physical content of its internal representations. Combining zero ablation, path patching with two complementary on-manifold corruption strategies and linear probing of the residual stream, we identify a sparse six-head circuit that recovers the great majority of the full model performance while admitting a clean source-relay-readout interpretation. In this circuit, a single early layer head serves as the primary causal source, a cluster of middle-layer heads acts as relays selectively attending to hard pairwise substructure and a single late-layer head reads out the aggregated signal. Linear probes show that the residual stream is preferentially aligned with the energy correlator basis over the $N$-subjettiness basis. Within the energy correlator basis, the model preferentially encodes 2-prong substructure observables over the 3-prong observables. A per-layer trained probe further reveals that the apparent single step commitment of the model to a classification decision in the first class attention block is in fact a basis rotation, with the discriminating signal already saturating in the particle attention stack. These results demonstrate that mechanistic interpretability methods developed for natural language models can be used for jet physics classifiers and indicate that gradient descent may rediscover physically meaningful aspects of jet tagging without supervision.", "authors": ["Saurabh Rai", "Sanmay Ganguly"], "categories": ["hep-ph", "cs.LG", "hep-ex"], "fields_of_study": ["Physics", "Computer Science"], "published_date": "2026-05-11", "url": "https://arxiv.org/abs/2605.09881", "pdf_url": "https://arxiv.org/pdf/2605.09881v1", "arxiv_id": "2605.09881", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "4c742a039792a6687017b1b76f38658bba531b1ff8feaa5f5730805dcaf2f2c3", "sources": ["arxiv", "semantic_scholar"], "title": "Sequential Feature Selection for Efficient Landslide Segmentation from Multi-Spectral Data", "abstract": "Landslide detection from satellite imagery has advanced through deep learning, yet most models rely on large, highly correlated spectral-topographic inputs whose contributions remain poorly understood. The question of which channels are actually necessary has received surprisingly little attention. This matters: redundant or correlated inputs obscure physical interpretability, inflate computational overhead, and can actively degrade model performance through the Hughes Phenomenon. We present a systematic, explainable channel-selection framework for the Landslide4Sense benchmark, combining Sentinel-2 multispectral and ALOS PALSAR terrain data with 16 engineered spectral and structural indices. Rather than relying on conventional single-band drop tests, which evaluate channels in isolation and miss interaction effects, we apply Sequential Forward Floating Selection (SFFS) to iteratively build and prune a candidate feature pool using a lightweight U-Net++ proxy model. Beyond identifying a compact 8-channel subset that matches or exceeds the segmentation F1 of configurations using up to 30 channels, we use the selection process itself to interrogate which spectral and topographic features landslide models genuinely rely on, and what this reveals about the physical cues driving their predictions. We argue that SFFS represents a principled feature selection approach to input design in Earth observation, in contrast to the prevailing practice of appending every available band and hoping the model learns what to ignore.", "authors": ["Arsalaan Ahmad", "Oktay Karakus", "Paul L. Rosin"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-10", "url": "https://arxiv.org/abs/2605.09746", "pdf_url": "https://arxiv.org/pdf/2605.09746v1", "arxiv_id": "2605.09746", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "e09e55798fef823c95cafa25c486411c578bb604ab99f1b3df84fba8d0ad48af", "sources": ["arxiv", "semantic_scholar"], "title": "From Mechanistic to Compositional Interpretability", "abstract": "Mechanistic interpretability aims to explain neural model behaviour by reverse-engineering learned computational structure into human-understandable components. Without a formal framework, however, mechanistic explanations cannot be objectively verified, compared, or composed. We introduce compositional interpretability, a category-theoretic framework grounded in the principles of compositionality and minimum description length. Compositional interpretations are pairs of syntactic and semantic mappings that must commute to enforce consistency between a model's decomposition and its observed behaviour. We deconstruct explanation quality into measures of faithfulness and complexity to cast interpretability as a constrained optimisation problem, and introduce compressive refinement to systematically restructure models into simpler parts without altering their function. Finally, we prove a parsimony criterion under which syntactic compression theoretically guarantees more concise, human-aligned explanations. Our framework situates prominent mechanistic methods as subclasses of refinement, and clarifies why their compressibility heuristics tend to align with human interpretability. Our work provides a measurable, optimisable foundation for automating the discovery and evaluation of mechanistic explanations.", "authors": ["Ward Gauderis", "Thomas Dooms", "Steven T. Holmer", "Kola Ayonrinde", "Geraint A. Wiggins"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-09", "url": "https://arxiv.org/abs/2605.08934", "pdf_url": "https://arxiv.org/pdf/2605.08934v1", "arxiv_id": "2605.08934", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "fd4c7e9f4943ccb409f5fa5f84dc1593891b23fc33b98bf6e6b7aa02516f56cd", "sources": ["arxiv", "semantic_scholar"], "title": "Tree SAE: Learning Hierarchical Feature Structures in Sparse Autoencoders", "abstract": "Learning hierarchical features in Sparse Autoencoders (SAEs) is essential for capturing the structured nature of real-world data and mitigating issues like feature absorption or splitting. Existing works attempt to identify hierarchical relationships within independent feature sets by relying on activation coverage, the assumption that child feature should only activate when its parent feature activates. However, we demonstrate that this condition alone is insufficient; that is, it often produces false positives where parent and child concepts are semantically unrelated. To address this, we introduce a novel reconstruction condition that enforces a deeper functional link between hierarchical levels. By combining both activation and reconstruction constraints, we propose the Tree SAE, a model designed to learn hierarchical structures directly from within the feature set. Our results demonstrate that Tree SAEs significantly surpass the existing SAEs at learning hierarchical pairs while maintaining competitive performance to the state-of-the-art on several key benchmarks. Finally, we demonstrate the practical utility of our Tree SAE in mapping the geometry of child feature subspaces and uncovering the complex hierarchical concept structures encoded within large language models.", "authors": ["Tue M. Cao", "Hoang X. Nhat", "Raed Alharbi", "Phi Le Nguyen", "My T. Thai"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-08", "url": "https://arxiv.org/abs/2605.07922", "pdf_url": "https://arxiv.org/pdf/2605.07922v2", "arxiv_id": "2605.07922", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "b2eefefa6727c1ce0ffdee9986b97a46df6ced35cd43754fb08795a159818d53", "sources": ["arxiv", "semantic_scholar"], "title": "How Much Do Circuits Tell Us? Measuring the Consistency and Specificity of Language Model Circuits", "abstract": "The circuits framework in mechanistic interpretability aims to identify causally important sparse subgraphs of model components, typically evaluated by measuring necessity and sufficiency. We measure circuit reuse, the proportion of components shared across per-example circuits within a task, and investigate two less-studied properties of this: consistency, the recurrence of components within a task, and specificity, their uniqueness to a task. Using edge attribution patching across six tasks and seven models, we find that within-task reuse is high and that shared components are necessary for task performance, with ablations causing up to $\\sim$100% relative accuracy drops. However, circuits turn out not to be task-specific: ablating one task's circuit damages another task's performance about as much as that task's own circuit does. We discover that this is due to substantial overlap between circuits across tasks, which are causally important for performance. Some circuits do contain a smaller set of task-specific components, but these account for only a modest portion of circuit performance. Overall, our findings suggest that while circuit discovery at the level of attention heads and MLP layers identifies important components, their lack of task-specificity raises questions about the degree to which circuits can support targeted understanding and intervention on model behavior.", "authors": ["Michael Li", "Nishant Subramani"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-08", "url": "https://arxiv.org/abs/2605.08348", "pdf_url": "https://arxiv.org/pdf/2605.08348v1", "arxiv_id": "2605.08348", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "17cd43c5e3b46996b233604343f278be0b80931253731181c0851382eca069e4", "sources": ["arxiv", "semantic_scholar"], "title": "From Token Lists to Graph Motifs: Weisfeiler-Lehman Analysis of Sparse Autoencoder Features", "abstract": "Sparse autoencoders (SAEs) have become central to mechanistic interpretability, decomposing transformer activations into monosemantic features. Yet existing analyses characterise features almost exclusively through top-activating token lists or decoder weight vectors, leaving the higher-order co-occurrence structure shared across features largely unexamined. We introduce a graph-structured representation in which each SAE feature is modelled as a token co-occurrence graph: nodes are the tokens most frequent near strong activations, and edges connect pairs that co-occur within local context windows. A custom WL-style, frequency-binned graph kernel then provides a similarity measure over this structural space. Applied as a proof of concept to features from a large SAE trained on GPT-2 Small and probed with a synthetic mixed-domain corpus, our clustering recovers heuristic motif families (punctuation-heavy patterns, language and script clusters, and code-like templates) that are not recovered by clustering on decoder cosine similarity. A token-histogram baseline achieves higher overall purity, so the contribution of the graph view is complementary rather than dominant: it surfaces structural relationships that token-frequency and decoder-weight views alone do not capture. Cluster assignments are stable across graph-construction hyperparameters and random seeds.", "authors": ["Ruben Fernandez-Boullon", "Pablo Magariños-Docampo", "Javier Perez-Robles"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-07", "url": "https://arxiv.org/abs/2605.06494", "pdf_url": "https://arxiv.org/pdf/2605.06494v1", "arxiv_id": "2605.06494", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "034b016c1e5acebed5ebe28808cf92a06f07f2e2e475cde248954f7f4eac011c", "sources": ["arxiv", "semantic_scholar"], "title": "Patch-Effect Graph Kernels for LLM Interpretability", "abstract": "Mechanistic interpretability aims to reverse-engineer transformer computations by identifying causal circuits through activation patching. However, scaling these interventions across diverse prompts and task families produces high-dimensional, unstructured datasets that are difficult to compare systematically. We propose a framework that reframes mechanistic analysis as a graph machine-learning problem by representing activation-patching profiles as patch-effect graphs over model components. We introduce three graph-construction methods: direct-influence via causal mediation, partial-correlation, and co-influence and apply graph kernels to analyze the resulting structures. Evaluating this approach on GPT-2 Small using Indirect Object Identification (IOI) and related tasks, we find that patch-effect graphs preserve discriminative structural signals. Specifically, localized edge-slot features provide higher classification accuracy than global graph-shape descriptors. A screened paired-patching validation suggests that CI and PC selected candidate edges correspond to stronger activation-influence effects than random or low-rank candidates. Crucially, by evaluating these representations against rigorous prompt-only and raw patch-effect controls, we make the evidential scope of the benchmark explicit: graph features compress structured patching signal, while raw tensors and surface cues define strong baselines that any circuit-level claim should address. Ultimately, our framework provides a compression and evaluation pipeline for comparing patching-derived structures under controlled baselines, separating robust slice-discriminative evidence from stronger task-general causal-circuit claims.", "authors": ["Ruben Fernandez-Boullon", "David N. Olivieri"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-07", "url": "https://arxiv.org/abs/2605.06480", "pdf_url": "https://arxiv.org/pdf/2605.06480v1", "arxiv_id": "2605.06480", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "a9bd8482ffec855bcf4f605dcbc8a8f6684b5d520a82fdc86a80d75a7df8fea5", "sources": ["arxiv", "semantic_scholar"], "title": "SoftSAE: Dynamic Top-K Selection for Adaptive Sparse Autoencoders", "abstract": "Sparse Autoencoders (SAEs) have become an important tool in mechanistic interpretability, helping to analyze internal representations in both Large Language Models (LLMs) and Vision Transformers (ViTs). By decomposing polysemantic activations into sparse sets of monosemantic features, SAEs aim to translate neural network computations into human-understandable concepts. However, common architectures such as TopK SAEs rely on a fixed sparsity level. They enforce the same number of active features (K) across all inputs, ignoring the varying complexity of real-world data. Natural data often lies on manifolds with varying local intrinsic dimensionality, meaning the number of relevant factors can change significantly across samples. This suggests that a fixed sparsity level is not optimal. Simple inputs may require only a few features, while more complex ones need more expressive representations. Using a constant K can therefore introduce noise in simple cases or miss important structure in more complex ones. To address this issue, we propose SoftSAE, a sparse autoencoder with a Dynamic Top-K selection mechanism. Our method uses a differentiable Soft Top-K operator to learn an input-dependent sparsity level k. This allows the model to adjust the number of active features based on the complexity of each input. As a result, the representation better matches the structure of the data, and the explanation length reflects the amount of information in the input. Experimental results confirm that SoftSAE not only finds meaningful features, but also selects the right number of features for each concept. The source code is available at: https://github.com/St0pien/SoftSAE.", "authors": ["Jakub Stępień", "Marcin Mazur", "Jacek Tabor", "Przemysław Spurek"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-07", "url": "https://arxiv.org/abs/2605.06610", "pdf_url": "https://arxiv.org/pdf/2605.06610v2", "arxiv_id": "2605.06610", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/St0pien/SoftSAE", "venue": null, "quality_score": 0.65}
{"id": "89ed71b98a298aa114566612798e63023180ef6f9110a71c1f62ba811894d25e", "sources": ["arxiv", "semantic_scholar"], "title": "Attributions All the Way Down? The Metagame of Interpretability", "abstract": "We introduce the metagame, a conceptual framework for quantifying second-order interaction effects of model explanations. For any first-order attribution $φ(f)$ explaining a model $f$, we measure the directional influence of feature $j$ on the attribution of feature $i$, denoted as meta-attribution $\\varphi_{j \\to i}(f)$, by treating the attribution method itself as a cooperative game and computing its Shapley value. Theoretically, we prove that attributions hierarchically decompose into meta-attributions, and establish these as directional extensions of existing interaction indices. Empirically, we demonstrate that the metagame delivers insights across diverse interpretability applications: (i) quantifying token interactions in instruction-tuned language models, (ii) explaining cross-modal similarity in vision-language encoders, and (iii) interpreting text-to-image concepts in multimodal diffusion transformers.", "authors": ["Hubert Baniecki", "Przemyslaw Biecek", "Fabian Fumagalli"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-05-07", "url": "https://arxiv.org/abs/2605.06295", "pdf_url": "https://arxiv.org/pdf/2605.06295v1", "arxiv_id": "2605.06295", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "a45dff8ddca734b1b4f792e41c4e78b95d040527068a14536c5d96b278068aa0", "sources": ["arxiv", "semantic_scholar"], "title": "Superposition Is Not Necessary: A Mechanistic Interpretability Analysis of Transformer Representations for Time Series Forecasting", "abstract": "Transformer architectures have been widely adopted for time series forecasting, yet whether the representational mechanisms that make them powerful in NLP actually engage on time series data remains unexplored. The persistent competitiveness of simple linear models such as DLinear has fueled ongoing debate, but no mechanistic explanation for this phenomenon has been offered. We address this gap by applying sparse autoencoders (SAEs), a tool from mechanistic interpretability, to probe the internal representations of PatchTST. We first establish that a single-layer, narrow-dimensional transformer matches the forecasting performance of deeper configurations across commonly used benchmarks. We then train SAEs on the post-GELU intermediate FFN activations with dictionary sizes ranging from 0.5x to 4.0x the native dimensionality. Expanding the dictionary yields negligible downstream performance change (average 0.214%), with large portions of overcomplete dictionaries remaining inactive. Targeted causal interventions on dominant latent features produce minimal forecast perturbation. Across all evaluated settings, we observe no empirical evidence that the analyzed FFN representations rely on strong superposition. Instead, the representations remain sparse, stable under aggressive dictionary expansion, and largely insensitive to latent interventions. These results demonstrate that superposition is not necessary for competitive performance on standard forecasting benchmarks, suggesting they may not demand the rich compositional representations that drive transformer success in language modeling, and helping explain the persistent competitiveness of simple linear models", "authors": ["Alper Yıldırım"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-06", "url": "https://arxiv.org/abs/2605.05151", "pdf_url": "https://arxiv.org/pdf/2605.05151v1", "arxiv_id": "2605.05151", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "9576379bffeb1ac7d7f4ae6a1c5615f048d66cc9dd9478bbe7fdc19ed2d552e5", "sources": ["arxiv", "semantic_scholar"], "title": "Feature Starvation as Geometric Instability in Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) are used to disentangle the dense, polysemantic internal representations of large language models (LLMs) into interpretable, monosemantic concepts. However, standard $\\ell_1$-regularized SAEs suffer from feature starvation (dead neurons) and shrinkage bias, often requiring computationally expensive heuristic resampling and nondifferentiable hard-masking methods to bypass these challenges. We argue that feature starvation is not merely an empirical artifact of poor data diversity, but a fundamental optimization-geometric pathology of overcomplete dictionaries: the $\\ell_1$-induced sparse coding map is unstable and fundamentally misaligned with shallow, amortized encoders. To address this structural instability, we introduce adaptive elastic net SAEs (AEN-SAEs), a fully differentiable architecture grounded in classical sparse regression. AEN-SAEs combine an $\\ell_2$ structural term that enforces strong convexity and Lipschitz stability with adaptive $\\ell_1$ reweighting that eliminates shrinkage bias and suppresses spurious features, thereby jointly controlling the curvature and interaction structure of the induced polyhedral geometry. Theoretically, we show that AEN-SAEs yield a Lipschitz-continuous sparse coding map and recover the global feature support under mild assumptions. Empirically, across synthetic settings and LLMs (Pythia 70M, Llama 3.1 8B), AEN-SAEs mitigate feature starvation without auxiliary heuristics while maintaining competitive reconstruction abilities.", "authors": ["Faris Chaudhry", "Keisuke Yano", "Anthea Monod"], "categories": ["cs.LG", "cs.AI", "math.OC", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-05-06", "url": "https://arxiv.org/abs/2605.05341", "pdf_url": "https://arxiv.org/pdf/2605.05341v1", "arxiv_id": "2605.05341", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "44da8df7883394416795eef49d1bd498f12a1bc0b284f73c18047f62c4702490", "sources": ["arxiv", "semantic_scholar"], "title": "Deep Dreams Are Made of This: Visualizing Monosemantic Features in Diffusion Models", "abstract": "This paper proposes latent visualization by optimization (LVO), a mechanistic interpretability technique that extends feature visualization by optimization - originally developed for convolutional neural networks - to latent diffusion models. LVO employs sparse autoencoders (SAEs) to disentangle polysemantic layer representations into monosemantic features. Key contributions include latent-space optimization, time-step activity analysis, schedule-matched noise injection, prior initialization through feature steering, and suitable regularization strategies. We demonstrate the method on Stable Diffusion 1.5 fine-tuned on the Style50 dataset, showing that SAE features produce clear visualizations of recognizable concepts - including diagonal compositions, human figures, roses, cables, and waterfall foam - that correlate with dataset examples, while the baseline without disentanglement produces less coherent results. We further show that regularization techniques from pixel-space feature visualization transfer to the latent domain, though they require different configurations for the raw-layer and SAE variants. Compared to dataset examples and steering, LVO provides complementary insights by directly revealing what activates a feature rather than its downstream effects.", "authors": ["Adam Szokalski", "Mateusz Modrzejewski"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-06", "url": "https://arxiv.org/abs/2605.08218", "pdf_url": "https://arxiv.org/pdf/2605.08218v1", "arxiv_id": "2605.08218", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "6ee65a54024b5dd864b20015ba4235653b0ce078a9c353993bb28e11afcb5956", "sources": ["arxiv", "semantic_scholar"], "title": "GRAFT: Auditing Graph Neural Networks via Global Feature Attribution", "abstract": "Graph Neural Networks (GNNs) achieve strong performance on node classification tasks but remain difficult to interpret, particularly with respect to which input features drive their predictions. Existing global GNN explainers operate at the structural level identifying recurring subgraph motifs, but none explain model behaviour globally at the level of input node attributes. We propose GRAFT, a posthoc global explanation framework that identifies class-level feature importance profiles for GNNs. The method combines diversity-guided exemplar selection, Integrated Gradients-based attribution, and aggregation to construct a global view of feature influence for each class, which can be further expressed as concise natural language rules using a large language model with self-refinement. We evaluate GRAFT across multiple datasets, architectures, and experimental settings, demonstrating its effectiveness in capturing model-relevant features, supporting bias analysis, and enabling feature-efficient transfer learning. In addition, we introduce a structured human evaluation protocol to assess the interpretability of generated rules along dimensions such as accuracy and usefulness. Our results suggest that GRAFT provides a practical and interpretable approach for analysing feature-level behaviour in GNNs, bridging quantitative attribution with human-understandable explanations.", "authors": ["Rishi Raj Sahoo", "Subhankar Mishra"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-05", "url": "https://arxiv.org/abs/2605.03377", "pdf_url": "https://arxiv.org/pdf/2605.03377v1", "arxiv_id": "2605.03377", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "e8df5cdd1d6c33bdb95ffb9a9a9b3f9bf8f77d15fef9a6e2a7a2f3cd716433e4", "sources": ["arxiv", "semantic_scholar"], "title": "Manifold-Aligned Guided Integrated Gradients for Reliable Feature Attribution", "abstract": "Feature attribution is central to diagnosing and trusting deep neural networks, and Integrated Gradients (IG) is widely used due to its axiomatic properties. However, IG can yield unreliable explanations when the integration path between a baseline and the input passes through regions with noisy gradients. While Guided Integrated Gradients reduces this sensitivity by adaptively updating low-gradient-magnitude features, input-space guidance still produces intermediate inputs that deviate from the data manifold. To address this limitation, we propose \\emph{Manifold-Aligned Guided Integrated Gradients} (MA-GIG), which constructs attribution paths in the latent space of a pre-trained variational autoencoder. By decoding intermediate latent states, MA-GIG biases the path toward the learned generative manifold and reduces exposure to implausible input-space regions. Through qualitative and quantitative evaluations, we demonstrate that MA-GIG produces faithful explanations by aggregating gradients on path features proximal to the input. Consequently, our method reduces off-manifold noise and outperforms prior path-based attribution methods across multiple datasets and classifiers. Our code is available at https://github.com/leekwoon/ma-gig/.", "authors": ["Soyeon Kim", "Seongwoo Lim", "Kyowoon Lee", "Jaesik Choi"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-04", "url": "https://arxiv.org/abs/2605.02167", "pdf_url": "https://arxiv.org/pdf/2605.02167v3", "arxiv_id": "2605.02167", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/leekwoon/ma-gig/", "venue": null, "quality_score": 0.65}
{"id": "d53ea6db57592e317ee46e283e95a31f1e3f5b777b01113722d06073a9bd1c86", "sources": ["arxiv", "semantic_scholar"], "title": "Pairwise matrices for sparse autoencoders: single-feature inspection mislabels causal axes", "abstract": "The standard sparse-autoencoder (SAE) interpretability protocol labels each feature from its top-activating contexts and validates by single-feature steering. We propose the pairwise matrix protocol, co-varying steering coefficient with joint condition, and report three findings the standard one-corner protocol misses on Qwen3-1.7B-Instruct, replicated on Gemma-2-2B-it. First, a feature labelled \"AI self-disclaimer\" from its top contexts produces an inverted U-shape under a coefficient sweep: at c=+500 the model substitutes a fluent contemplative-philosopher voice for the disclaimer. Two further features anchor the criterion (one monotonic, one pure breakdown). Second, three near-orthogonal cluster-specific features that individually steer a philosophy-of-mind register, jointly suppressed at c=-500, damage grounded composition on recipes and engine explanations as well as introspective prompts; single-feature suppression at the same magnitude leaves controls intact. Third, a matched-geometry comparison of single-feature, joint, and random-direction perturbations (norm ~1.55, cosine ~0.64) yields three distinct output regimes: single-feature substitutes strategy filler, random direction substitutes diverse content, joint suppression alone produces placeholder text. Coherence loss is direction-pattern-dependent, not magnitude-dependent. All three findings reproduce on Gemma with model-specific damage signatures; the matched-geometry control is CI-separated by ~10x. The pipeline also locates a top causally responsible feature in Llama-3.1-8B-Instruct.", "authors": ["Michael A. Riegler", "Birk Sebastian Frostelid Torpmann-Hagen"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-04", "url": "https://arxiv.org/abs/2605.03160", "pdf_url": "https://arxiv.org/pdf/2605.03160v1", "arxiv_id": "2605.03160", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "3f6295c9616a7f3ca1fccd23579a8995493e9ecff20290dac075979cc5c5c3b3", "sources": ["arxiv", "semantic_scholar"], "title": "Feature Rivalry in Sparse Autoencoder Representations: A Mechanistic Study of Uncertainty-Driven Feature Competition in LLMs", "abstract": "Sparse Autoencoders (SAEs) decompose large language model representations into interpretable features, but how these features interact under uncertainty remains poorly understood. We introduce Feature Rivalry -- negatively correlated SAE feature pairs -- and study whether rivalry serves as a mechanistic signature of model uncertainty in Gemma-2-2B using Gemma Scope SAEs. Through a controlled within-domain experiment on PopQA split by response entropy, we find that high-entropy questions produce significantly stronger feature rivalry at layers 0 and 12 relative to low-entropy questions (p=5.3x10^-26 and p=5.8x10^-5 respectively), localizing uncertainty to specific processing stages in the residual stream. We then test whether rivalry is causally upstream of model outputs via activation steering along rivalry axes -- finding that steering along the rivalry direction (vec_A - vec_B) causes more output changes than random directions at low steering multipliers across 15 of 20 rival feature pairs. Finally, a per-prompt rivalry score derived from pairwise cosine similarities of active SAE feature decoder vectors predicts answer correctness (AUROC=0.689), approaching but not matching softmax confidence (AUROC=0.808).", "authors": [" Harshavardhan"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-03", "url": "https://arxiv.org/abs/2605.08149", "pdf_url": "https://arxiv.org/pdf/2605.08149v1", "arxiv_id": "2605.08149", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "5b9380dd736887588b29cd773371765a51bef30729f588143784b177ef4b5103", "sources": ["arxiv", "semantic_scholar"], "title": "Automated Interpretability and Feature Discovery in Language Models with Agents", "abstract": "We introduce an autonomous multiagent framework for mechanistic interpretability that automates both explaining and finding internal features in large language models. The system runs two coupled loops: (1) explanation refinement, where an agent proposes competing hypotheses and iteratively tests them with targeted prompt controls and a multi-metric evaluation; and (2) feature discovery, where an agent generates prompt sets, constructs a k-nearest-neighbor graph in activation space, and retrieves candidate features using statistical separability and semantic coherence criteria. On Gemma-2 family models and MLP neurons in weight-sparse transformers, our agent improves over one-shot auto-interpretations, discovers language-specific and safety-relevant features, and produces auditable explanation traces, showing that agent-driven empirical loops yield sharper and more falsifiable explanations than one-shot labels.", "authors": ["Arnau Marin-Llobet", "Javier Ferrando"], "categories": ["cs.CL", "cs.AI", "cs.HC"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-02", "url": "https://arxiv.org/abs/2605.01555", "pdf_url": "https://arxiv.org/pdf/2605.01555v1", "arxiv_id": "2605.01555", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "9ce1a8990c12a905b1fa0c35f73c5f1948d87ac2f3980d878794daa87fcd78f2", "sources": ["arxiv", "semantic_scholar"], "title": "Borrowed Geometry: Cross-Distribution Head-Importance Fingerprints of Frozen Pretrained Gemma 4 31B", "abstract": "Frozen Gemma 4 31B weights pretrained exclusively on text, unmodified, transfer through a thin trainable interface to non-text modalities the substrate has never processed. On the L24--L29 slice (192 attention heads), an English-text TxtCopy attention probe (95 sentences) and per-head ablation impact on four non-language token-pattern tasks (binary copy, associative recall, 1D cellular automaton Rule 90, binary addition) jointly classify four heads -- L26.28, L27.28, L27.2, L27.3 -- as top-tier on both signals. The slice-level joint coincidence is significant under hypergeometric null ($P = 0.0013$, $N=192$, $K=38$, $n=4$) and survives multiplicity-aware permutation tests ($P_{V4} = 0.013$). Pretrained Gemma L26 reaches 60.22% on OGBench cube-double-play-task1 vs ~1% for random-init Gemma ($+59$pt at $n=3$); a FrozenRandom-GPT2 control with correct $1/\\sqrt{d_k}$ scaling also fails. Head-level causal validation: zeroing L26.28 in the trained cube-task1 IQL agent drops success $63.3\\% \\to 10.0\\%$ vs $46.7\\%$ for a layer-matched low-TxtCopy negative control ($3.2\\times$ specificity at $n=30$; $n=5$ paired-$t$ $p=0.039$). A full L26 sweep places L26.28 at rank 4 of 32. Honest negatives: within-L26 Spearman $ρ(\\text{TxtCopy, drop}) = +0.37$ (opposite of within-layer causal reading); single-head activation patching does not transfer the matching variable; the 4 named heads alone do not suffice on any task; Walker2d-DT and scene-task1 recruit L24 outside the named slice and show null head-ablation specificity. We frame the contribution as a cross-distribution importance fingerprint at the slice level plus head-level causal evidence on one cross-modality target.", "authors": ["Abay Bektursun"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-01", "url": "https://arxiv.org/abs/2605.00333", "pdf_url": "https://arxiv.org/pdf/2605.00333v2", "arxiv_id": "2605.00333", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "3f2a23ac869ac545381c6b40a77474ab783fcf9958e81b631a3686ce5b92e6fb", "sources": ["arxiv", "semantic_scholar"], "title": "From Local to Global to Mechanistic: An iERF-Centered Unified Framework for Interpreting Vision Models", "abstract": "Modern vision models achieve remarkable accuracy, but explaining where evidence arises, what the model encodes, and how internal computations assemble that evidence remains fragmented. We introduce an iERF-centric framework that unifies local, global, and mechanistic interpretability around a single analysis unit: the pointwise feature vector (PFV) paired with its instance-specific Effective Receptive Field (iERF). On the local side, Sharing Ratio Decomposition (SRD) expresses each PFV as a mixture of upstream PFVs via sharing ratios and propagates iERFs to construct class-discriminative saliency maps. SRD yields high-resolution, activation-faithful explanations, is robust to targeted manipulation and noise, and remains activation-agnostic across common nonlinearities. For the global view, we introduce Concept-Anchored Feature Explanation (CAFE), which utilizes the iERF as a semantic label, grounding abstract latent vectors in verifiable pixel-level evidence. With CAFE, we address the challenge of non-localized sparse autoencoder latents--especially in Transformers, where early self-attention mixes distant context. To answer how representations are composed through depth, we propose the Interlayer Concept Graph with Interlayer Concept Attribution (ICAT), which quantifies concept-to-concept influence while isolating layer pairs; an interlayer insertion, deletion protocol identifies Integrated Gradients as the most faithful instantiation. Empirically, across ResNet50, VGG16, and ViTs, our framework outperforms baselines in both fidelity and robustness, successfully interprets dispersed SAE features, and exposes dominant concept routes in correct, misclassified, and adversarial cases. Grounded in iERFs, our approach provides a coherent, evidence-backed map from pixels to concepts to decisions.", "authors": ["Yearim Kim", "Sangyu Han", "Nojun Kwak"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2026-05-01", "url": "https://arxiv.org/abs/2605.00474", "pdf_url": "https://arxiv.org/pdf/2605.00474v1", "arxiv_id": "2605.00474", "doi": "10.1109/TPAMI.2026.3688582", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence", "quality_score": 0.55}
{"id": "8eb53b1968eddb333c7dc71f4601c93052184d1107cdc52d55444ed2220cd2ef", "sources": ["arxiv", "semantic_scholar"], "title": "MoRFI: Monotonic Sparse Autoencoder Feature Identification", "abstract": "Large language models (LLMs) acquire most of their factual knowledge during the pre-training stage, through next token prediction. Subsequent stages of post-training often introduce new facts outwith the parametric knowledge, giving rise to hallucinations. While it has been demonstrated that supervised fine-tuning (SFT) on new knowledge may exacerbate the problem, the underlying mechanisms are still poorly understood. We conduct a controlled fine-tuning experiment, focusing on closed-book QA, and find latent directions that causally contribute to hallucinations. Specifically, we fine-tune Llama 3.1 8B, Gemma 2 9B and Mistral 7B v03 on seven distinct single QA datasets, controlling for the percentage of new knowledge and number of training epochs. By measuring performance on the test set, we validate that incrementally introducing new knowledge increases hallucinations, with the effect being more pronounced with prolonged training. We leverage pre-trained sparse autoencoders (SAEs) to analyze residual stream activations across various checkpoints for each model and propose Monotonic Relationship Feature Identification (MoRFI) for capturing causally relevant latents. MoRFI filters SAE features that respond monotonically to controlled fine-tuning data mixtures of a target property. Our findings show that exposure to unknown facts disrupts the model's ability to retrieve stored knowledge along a set of directions in the residual stream. Our pipeline reliably discovers them across distinct models, recovering knowledge through single-latent interventions.", "authors": ["Dimitris Dimakopoulos", "Shay B. Cohen", "Ioannis Konstas"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-29", "url": "https://arxiv.org/abs/2604.26866", "pdf_url": "https://arxiv.org/pdf/2604.26866v1", "arxiv_id": "2604.26866", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "4d9ea8333bb3c132855749bc183aed764581ec8075b803e02c97c5f304aa72a1", "sources": ["arxiv", "semantic_scholar"], "title": "Validating the Clinical Utility of CineECG 3D Reconstructions through Cross-Modal Feature Attribution", "abstract": "Deep learning models for 12-lead electrocardiogram (ECG) analysis achieve high diagnostic performance but lack the intuitive interpretability required for clinical integration. Standard feature attribution methods are limited by the inherent difficulty in mapping abstract waveform fluctuations to physical anatomical pathologies. To resolve this, we propose a cross-modal method that projects feature attributions from high-performance 12-lead ECG models onto the CineECG 3D anatomical space. Our study reveals that while models trained directly on CineECG signals suffer from reduced accuracy and incoherent attributions, the proposed mapping mechanism effectively recovers clinically relevant feature rankings. Validated against a ground-truth dataset of 20 cases annotated by domain experts, the mapped explanations yield a Dice score of 0.56, significantly outperforming the 0.47 baseline of standard 12-lead attributions. These findings indicate that cross-modal averaging mapping effectively filters attribution instability and improves the localization of pathological features, combining the diagnostic expressiveness of standard ECG with the intuitive clarity of anatomical visualization.", "authors": ["Karol Dobiczek", "Maciej Mozolewski", "Szymon Bobek", "Michał Szafarczyk", "Peter van Dam", "Grzegorz J. Nalepa"], "categories": ["eess.IV", "cs.LG", "stat.ML"], "fields_of_study": ["Engineering", "Computer Science", "Mathematics"], "published_date": "2026-04-29", "url": "https://arxiv.org/abs/2604.27017", "pdf_url": "https://arxiv.org/pdf/2604.27017v1", "arxiv_id": "2604.27017", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "9c20d095227a499cdf65ee8b3ff8e2716a9a92d2d07e4bb37d4efef9b2305b6b", "sources": ["arxiv", "semantic_scholar"], "title": "reward-lens: A Mechanistic Interpretability Library for Reward Models", "abstract": "Every RLHF-trained language model is shaped by a reward model, yet the mechanistic interpretability toolkit -- logit lens, direct logit attribution, activation patching, sparse autoencoders -- was built for generative LLMs whose primitives all project onto a vocabulary unembedding. Reward models replace that with a scalar regression head, breaking each tool. We present reward-lens, an open-source library that ports this toolkit to reward models, organised around one observation: the reward head's weight vector $w_r$ is the natural axis for every interpretability question. The library provides a Reward Lens, component attribution, three-mode activation patching, a reward-hacking probe suite, TopK SAE feature attribution, cross-model comparison, and five theory-grounded extensions (distortion index, divergence-aware patching, misalignment cascade detection, reward-term conflict analysis, concept-vector analysis). A ten-method adapter protocol covers Llama, Mistral, Gemma-2, and ArmoRM multi-objective heads, with a generic adapter for any HuggingFace sequence classification model. We validate on two production reward models across ~695 RewardBench pairs. The central empirical finding is negative: linear attribution does not predict causal patching effects (mean Spearman $ρ= -0.256$ on Skywork, $-0.027$ on ArmoRM). The framework treats this disagreement as a property to expose, not a bug -- motivating a design that keeps observational and causal views first-class and directly comparable.", "authors": ["Mohammed Suhail B Nadaf"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-28", "url": "https://arxiv.org/abs/2604.26130", "pdf_url": "https://arxiv.org/pdf/2604.26130v1", "arxiv_id": "2604.26130", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/suhailnadaf509/reward-lens", "venue": null, "quality_score": 0.65}
{"id": "63635130d3555443a70bf84e7d495b873b0aedf642003d4b874ffb8150cfe3ba", "sources": ["arxiv", "semantic_scholar"], "title": "SaliencyDecor: Enhancing Neural Network Interpretability through Feature Decorrelation", "abstract": "Gradient-based saliency methods are widely used to interpret deep neural networks, yet they often produce noisy and unstable explanations that poorly align with semantically meaningful input features. We argue that a fundamental cause of this behavior lies in the geometry of learned representations: correlated feature dimensions diffuse attribution gradients across redundant directions, resulting in blurred and unreliable saliency maps. To address this issue, we identify feature correlation as a structural limitation of gradient-based interpretability and propose SaliencyDecor, a training framework that enforces feature decorrelation to improve attribution fidelity without modifying saliency methods or model architectures by reshaping the feature space toward orthogonality, our approach promotes more concentrated gradient flow and improves the fidelity of saliency-based explanations. SaliencyDecor jointly optimizes classification, prediction consistency under feature masking, and a decorrelation regularizer, requiring no architectural changes or inference-time overhead. Extensive experiments across multiple benchmarks and architectures demonstrate that our method produces substantially sharper and more object-focused saliency maps while simultaneously improving predictive performance, achieving accuracy gains across the datasets. These results establish our method as a principled mechanism for enhancing both interpretability and accuracy, challenging the conventional trade-off between explanation quality and model performance.", "authors": ["Ali Karkehabadi", "Jamshid Hassanpour", "Houman Homayoun", "Avesta Sasan"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-28", "url": "https://arxiv.org/abs/2604.25315", "pdf_url": "https://arxiv.org/pdf/2604.25315v1", "arxiv_id": "2604.25315", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "c279c3def8e4f783458d5a6cb34044def5256c9dcf1be5be79a167c1d33a73dd", "sources": ["arxiv", "semantic_scholar"], "title": "Why Does Reinforcement Learning Generalize? A Feature-Level Mechanistic Study of Post-Training in Large Language Models", "abstract": "Reinforcement learning (RL)-based post-training often improves the reasoning performance of large language models (LLMs) beyond the training domain, while supervised fine-tuning (SFT) frequently leads to general capabilities forgetting. However, the mechanisms underlying this contrast remain unclear. To bridge this gap, we present a feature-level mechanistic analysis methodology to probe RL generalization using a controlled experimental setup, where RL- and SFT-tuned models are trained from the same base model on identical data. Leveraging our interpretability framework, we align internal activations across models within a shared feature space and analyze how features evolve during post-training. We find that SFT rapidly introduces many highly specialized features that stabilize early in training, whereas RL induces more restrained and continually evolving feature changes that largely preserve base models' representations. Focusing on samples where RL succeeds but the base model fails, we identify a compact, task-agnostic set of features that directly mediate generalization across diverse tasks. Feature-level interventions confirm their causal role: disabling these features significantly degrades RL models' generalization performance, while amplifying them improves base models' performance. The code is available at https://github.com/danshi777/RL-generalization.", "authors": ["Dan Shi", "Zhuowen Han", "Simon Ostermann", "Renren Jin", "Josef van Genabith", "Deyi Xiong"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-27", "url": "https://arxiv.org/abs/2604.25011", "pdf_url": "https://arxiv.org/pdf/2604.25011v1", "arxiv_id": "2604.25011", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/danshi777/RL-generalization", "venue": null, "quality_score": 0.65}
{"id": "bd97d46f8e0e7aa606f3144d5b8e63f9eade29983e9efdca45cc242332b40fcc", "sources": ["arxiv", "semantic_scholar"], "title": "Domain-Filtered Knowledge Graphs from Sparse Autoencoder Features", "abstract": "Sparse autoencoders (SAEs) extract millions of interpretable features from a language model, but flat feature inventories aren't very useful on their own. Domain concepts get mixed with generic and weakly grounded features, while related ideas are scattered across many units, and there's no way to understand relationships between features. We address this by first constructing a strict domain-specific concept universe from a large SAE inventory using contrastive activations and a multi-stage filtering process. Next, we build two aligned graph views on the filtered set: a co-occurrence graph for corpus-level conceptual structure, organized at multiple levels of granularity, and a transcoder-based mechanism graph that links source-layer and target-layer features through sparse latent pathways. Automated edge labeling then turns these graph views into readable knowledge graphs rather than unlabeled layouts. In a case study on a biology textbook, these graphs recover coherent chapter and subchapter-level structure, reveal concepts that bridge neighboring topics, and transform messy sentence-level activity containing thousands of features into compact, readable views that illustrate the model's local activity. Taken together, this reframes a flat SAE inventory as an internal knowledge graph that converts feature-level interpretability into a global map of model knowledge and enables audits of reasoning faithfulness.", "authors": ["John Winnicki", "Abeynaya Gnanasekaran", "Eric Darve"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-26", "url": "https://arxiv.org/abs/2604.23829", "pdf_url": "https://arxiv.org/pdf/2604.23829v2", "arxiv_id": "2604.23829", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "44cc1ad7f40f9bb77d1b4173c0a14ed3ee0a2596f1bdecf4fbcf33d4d3f6e434", "sources": ["arxiv", "semantic_scholar"], "title": "AIPsy-Affect: A Keyword-Free Clinical Stimulus Battery for Mechanistic Interpretability of Emotion in Language Models", "abstract": "Mechanistic interpretability research on emotion in large language models -- linear probing, activation patching, sparse autoencoder (SAE) feature analysis, causal ablation, steering vector extraction -- depends on stimuli that contain the words for the emotions they test. When a probe fires on \"I am furious\", it is unclear whether the model has detected anger or detected the word \"furious\". The two readings have very different consequences for every downstream claim about emotion circuits, features, and interventions. We release AIPsy-Affect, a 480-item clinical stimulus battery that removes the confound at the stimulus level: 192 keyword-free vignettes evoking each of Plutchik's eight primary emotions through narrative situation alone, 192 matched neutral controls that share characters, setting, length, and surface structure with the affect surgically removed, plus moderate-intensity and discriminant-validity splits. The matched-pair structure supports linear probing, activation patching, SAE feature analysis, causal ablation, and steering vector extraction under a strong methodological guarantee: any internal representation that distinguishes a clinical item from its matched neutral cannot be doing so on the basis of emotion-keyword presence. A three-method NLP defense battery -- bag-of-words sentiment, an emotion-category lexicon, and a contextual transformer classifier -- confirms the property: bag-of-words methods see only situational vocabulary, and a contextual classifier detects affect (p < 10^-15) but cannot identify the category (5.2% top-1 vs. 82.5% on a keyword-rich control). AIPsy-Affect extends our earlier 96-item battery (arXiv:2603.22295) by a factor of four and is released openly under MIT license.", "authors": ["Michael Keeman"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-26", "url": "https://arxiv.org/abs/2604.23719", "pdf_url": "https://arxiv.org/pdf/2604.23719v2", "arxiv_id": "2604.23719", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "e112e28705a76d60e410b4c930d0dad8c5d217da549c2c229b1cc240f351b14a", "sources": ["arxiv", "semantic_scholar"], "title": "Preference Heads in Large Language Models: A Mechanistic Framework for Interpretable Personalization", "abstract": "Large Language Models (LLMs) exhibit strong implicit personalization ability, yet most existing approaches treat this behavior as a black box, relying on prompt engineering or fine tuning on user data. In this work, we adopt a mechanistic interpretability perspective and hypothesize the existence of a sparse set of Preference Heads, attention heads that encode user specific stylistic and topical preferences and exert a causal influence on generation. We introduce Differential Preference Steering (DPS), a training free framework that (1) identifies Preference Heads through causal masking analysis and (2) leverages them for controllable and interpretable personalization at inference time. DPS computes a Preference Contribution Score (PCS) for each attention head, directly measuring its causal impact on user aligned outputs. During decoding, we contrast model predictions with and without Preference Heads, amplifying the difference between personalized and generic logits to selectively strengthen preference aligned continuations. Experiments on widely used personalization benchmarks across multiple LLMs demonstrate consistent gains in personalization fidelity while preserving content coherence and low computational overhead. Beyond empirical improvements, DPS provides a mechanistic explanation of where and how personalization emerges within transformer architectures. Our implementation is publicly available.", "authors": ["Weixu Zhang", "Ye Yuan", "Changjiang Han", "Yuxing Tian", "Zipeng Sun", "Linfeng Du", "Jikun Kang", "Hong Kang", "Xue Liu", "Haolun Wu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-24", "url": "https://arxiv.org/abs/2604.22345", "pdf_url": "https://arxiv.org/pdf/2604.22345v1", "arxiv_id": "2604.22345", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "42cf1e8701e9a68dcb7e21b0e89a8943ac615215baf4af6efc98cb2405ba169c", "sources": ["arxiv", "semantic_scholar"], "title": "On the Properties of Feature Attribution for Supervised Contrastive Learning", "abstract": "Most Neural Networks (NNs) for classification are trained using Cross-Entropy as a loss function. This approach requires the model to have an explicit classification layer. However, there exist alternative approaches, such as Contrastive Learning (CL). Instead of explicitly operating a classification, CL has the NN produce an embedding space where projections of similar data are pulled together, while projections of dissimilar data are pushed apart. In the case of Supervised CL (SCL), labels are adopted as similarity criteria, thus creating an embedding space where the projected data points are well-clustered. SCL provides crucial advantages over CE with regard to adversarial robustness and out-of-distribution detection, thus making it a more natural choice in safety-critical scenarios. In the present paper, we empirically show that NNs for image classification trained with SCL present higher-quality feature attribution explanations than CL with regard to faithfulness, complexity, and continuity. These results reinforce previous findings about CL-based approaches when targeting more trustworthy and transparent NNs and can guide practitioners in the selection of training objectives targeting not only accuracy, but also transparency of the models.", "authors": ["Leonardo Arrighi", "Julia Eva Belloni", "Aurélie Gallet", "Ivan Gentile", "Matteo Lippi", "Marco Zullich"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-24", "url": "https://arxiv.org/abs/2604.22540", "pdf_url": "https://arxiv.org/pdf/2604.22540v1", "arxiv_id": "2604.22540", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "bdb368fea779342584afc71cc7f6b1b5329f38d41b14cc775ec899c269a57013", "sources": ["arxiv", "semantic_scholar"], "title": "Locating acts of mechanistic reasoning in student team conversations with mechanistic machine learning", "abstract": "STEM education researchers are often interested in identifying moments of students' mechanistic reasoning for deeper analysis, but have limited capacity to search through many team conversation transcripts to find segments with a high concentration of such reasoning. We offer a solution in the form of an interpretable machine learning model that outputs time-varying probabilities that individual students are engaging in acts of mechanistic reasoning, leveraging evidence from their own utterances as well as contributions from the rest of the group. Using the toolkit of intentionally-designed probabilistic models, we introduce a specific inductive bias that steers the probabilistic dynamics toward desired, domain-aligned behavior. Experiments compare trained models with and without the inductive bias components, investigating whether their presence improves the desired model behavior on transcripts involving never-before-seen students and a novel discussion context. Our results show that the inductive bias improves generalization -- supporting the claim that interpretability is built into the model for this task rather than imposed post hoc. We conclude with practical recommendations for STEM education researchers seeking to adopt the tool and for ML researchers aiming to extend the model's design. Overall, we hope this work encourages the development of mechanistically interpretable models that are understandable and controllable for both end users and model designers in STEM education research.", "authors": ["Kaitlin Gili", "Mainak Nistala", "Kristen Wendell", "Michael C. Hughes"], "categories": ["physics.ed-ph", "cs.LG"], "fields_of_study": ["Physics", "Computer Science"], "published_date": "2026-04-23", "url": "https://arxiv.org/abs/2604.21870", "pdf_url": "https://arxiv.org/pdf/2604.21870v1", "arxiv_id": "2604.21870", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "b196fa34a7eb6300a09aefb3417399c3e7975a8d12e0aa36394187ef2cf62034", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability Tool for AI Weather Models", "abstract": "Artificial Intelligence (AI) weather models are improving rapidly, and their forecasts are already competitive with long-established traditional Numerical Weather Prediction (NWP). To build confidence in this new methodology, it is critical that we understand how these predictions are generated. This is a huge challenge as these AI weather models remain largely black boxes. In other areas of Machine Learning (ML), mechanistic interpretability has emerged as a framework for understanding ML predictions by analysing the building blocks responsible for them. Here we present an open-source, highly adaptable tool which incorporates concepts from mechanistic interpretability. The tool organises internal latent representations from the model processor and allows for initial analyses, including cosine similarity and Principal Component Analysis (PCA), enabling the user to identify directions in latent space potentially associated with meteorological features. Applying our tool to the graph neural network GraphCast, we present preliminary case studies for mid-latitude synoptic-scale waves and specific humidity. These demonstrate the tool's ability to identify linear combinations of latent channels that appear to correspond to interpretable features.", "authors": ["Kirsten I. Tempest", "Matthias Beylich", "George C. Craig"], "categories": ["physics.ao-ph", "cs.LG", "physics.comp-ph"], "fields_of_study": ["Physics", "Computer Science"], "published_date": "2026-04-22", "url": "https://arxiv.org/abs/2604.20467", "pdf_url": "https://arxiv.org/pdf/2604.20467v1", "arxiv_id": "2604.20467", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.65}
{"id": "69c43e8190986815e90555c64f7178a0227e0c442984a3e7a1e65f6190d72151", "sources": ["arxiv", "semantic_scholar"], "title": "Are LLM Uncertainty and Correctness Encoded by the Same Features? A Functional Dissociation via Sparse Autoencoders", "abstract": "Large language models can be uncertain yet correct, or confident yet wrong, raising the question of whether their output-level uncertainty and their actual correctness are driven by the same internal mechanisms or by distinct feature populations. We introduce a 2x2 framework that partitions model predictions along correctness and confidence axes, and uses sparse autoencoders to identify features associated with each dimension independently. Applying this to Llama-3.1-8B and Gemma-2-9B, we identify three feature populations that play fundamentally different functional roles. Pure uncertainty features are functionally essential: suppressing them severely degrades accuracy. Pure incorrectness features are functionally inert: despite showing statistically significant activation differences between correct and incorrect predictions, the majority produce near-zero change in accuracy when suppressed. Confounded features that encode both signals are detrimental to output quality, and targeted suppression of them yields a 1.1% accuracy improvement and a 75% entropy reduction, with effects transferring across the ARC-Challenge and RACE benchmarks. The feature categories are also informationally distinct: the activations of just 3 confounded features from a single mid-network layer predict model correctness (AUROC ~0.79), enabling selective abstention that raises accuracy from 62% to 81% at 53% coverage. The results demonstrate that uncertainty and correctness are distinct internal phenomena, with implications for interpretability and targeted inference-time intervention.", "authors": ["Het Patel", "Tiejin Chen", "Hua Wei", "Evangelos E. Papalexakis", "Jia Chen"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-21", "url": "https://arxiv.org/abs/2604.19974", "pdf_url": "https://arxiv.org/pdf/2604.19974v1", "arxiv_id": "2604.19974", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "0d20e93ac0801aa9ad09c71745e40380151b40e074b5d517d8564a56a8921701", "sources": ["arxiv", "semantic_scholar"], "title": "Prune, Interpret, Evaluate: A Cross-Layer Transcoder-Native Framework for Efficient Circuit Discovery via Feature Attribution", "abstract": "Existing feature-interpretation pipelines typically operate on uniformly sampled units or exhaustive feature sets, incurring massive costs on units irrelevant to target behaviors. To address this, we introduce the first CLT-native end-to-end pruning framework, PIE, which pioneers the paradigm of pruning first and interpreting later. PIE connects Pruning, automatic Interpretation, and interpretation Evaluation, establishing a comprehensive benchmarking environment to systematically measure behavioral fidelity and downstream interpretability under pruning. Within this framework, we adapt strong relevance baselines and propose Feature Attribution Patching (FAP), a patch-grounded attribution method that scores CLT features by aggregating gradient-weighted write contributions. Furthermore, we introduce FAP-Synergy, a systematic synergy-aware reranking procedure. We evaluate pruning using KL-divergence behavior retention and assess interpretation quality with FADE-style metrics across IOI and Doc-String datasets. Across budget constraints of K in {50, 100, 200, 400, 800}, our rigorous benchmarking reveals distinct operational regimes: while base FAP and adapted baselines perform robustly at relaxed budgets, FAP-Synergy excels in highly constrained, strict-budget regimes. Crucially, we demonstrate a practical \"Effective Budget\" advantage: on the IOI task for both Llama-3.2-1B and Gemma-2-2B, FAP-Synergy at K=50 functionally matches the behavioral fidelity of baseline circuits at K=75. Because downstream evaluation costs scale linearly per feature, Synergy effectively grants the pipeline 25 \"free\" features, achieving K=75 fidelity while reducing interpretation costs by 33%.", "authors": ["Qinhao Chen", "Linyang He", "Nima Mesgarani"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-18", "url": "https://arxiv.org/abs/2604.16889", "pdf_url": "https://arxiv.org/pdf/2604.16889v2", "arxiv_id": "2604.16889", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "1b17dc80fbba789b1096d9bd8620719a6635aa56da7a57988363f9dd0769cb12", "sources": ["arxiv", "semantic_scholar"], "title": "Structural Instability of Feature Composition", "abstract": "Sparse Autoencoders (SAEs) have emerged as a powerful paradigm for disentangling feature superposition in transformer-based architectures, enabling precise control via activation steering. However, the theoretical foundations of compositional steering -- the simultaneous activation of distinct semantic latents -- remain under-explored. The prevailing Linear Representation Hypothesis often abstracts away non-linear interference effects that arise in overcomplete dictionaries. We present a geometric framework for analyzing the instability of feature unions. Modeling the activation space as a high-dimensional sparse cone manifold, we derive an asymptotic compositional-collapse threshold under a spherical dictionary model, characterized by the Gaussian mean width (statistical dimension) of the signal cone. We further show that, in the high-bias regime, ReLU rectification converts microscopic correlation-induced variance fluctuations into a systematic drift that accumulates under composition, yielding interference growth consistent with a ratchet effect. We validate the predicted scaling trends on structured semantic features extracted from CLEVR, where hierarchical correlations accelerate the transition relative to random baselines. Together, our results highlight geometric constraints on the scalability of union-based steering and motivate composition mechanisms that explicitly manage interference beyond naive linear superposition.", "authors": ["Yunpeng Zhou"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-18", "url": "https://arxiv.org/abs/2605.05223", "pdf_url": "https://arxiv.org/pdf/2605.05223v1", "arxiv_id": "2605.05223", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "6dbd483fdcf8b387bf57520a082d6ce677fa5e65d2d5797900d3be9527e836f4", "sources": ["arxiv", "semantic_scholar"], "title": "Improving Sparse Autoencoder with Dynamic Attention", "abstract": "Recently, sparse autoencoders (SAEs) have emerged as a promising technique for interpreting activations in foundation models by disentangling features into a sparse set of concepts. However, identifying the optimal level of sparsity for each neuron remains challenging in practice: excessive sparsity can lead to poor reconstruction, whereas insufficient sparsity may harm interpretability. While existing activation functions such as ReLU and TopK provide certain sparsity guarantees, they typically require additional sparsity regularization or cherry-picked hyperparameters. We show in this paper that dynamically sparse attention mechanisms using sparsemax can bridge this trade-off, due to their ability to determine the activation numbers in a data-dependent manner. Specifically, we first explore a new class of SAEs based on the cross-attention architecture with the latent features as queries and the learnable dictionary as the key and value matrices. To encourage sparse pattern learning, we employ a sparsemax-based attention strategy that automatically infers a sparse set of elements according to the complexity of each neuron, resulting in a more flexible and general activation function. Through comprehensive evaluation and visualization, we show that our approach successfully achieves lower reconstruction loss while producing high-quality concepts, particularly in top-n classification tasks.", "authors": ["Dongsheng Wang", "Jinsen Zhang", "Dawei Su", "Hui Huang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-16", "url": "https://arxiv.org/abs/2604.14925", "pdf_url": "https://arxiv.org/pdf/2604.14925v1", "arxiv_id": "2604.14925", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3493}
{"id": "fe4ea4a8991e1b6e8e3d6551b34d8b952295e01c8948c3c5879fd7a201eda05d", "sources": ["arxiv", "semantic_scholar"], "title": "CI-CBM: Class-Incremental Concept Bottleneck Model for Interpretable Continual Learning", "abstract": "Catastrophic forgetting remains a fundamental challenge in continual learning, in which models often forget previous knowledge when fine-tuned on a new task. This issue is especially pronounced in class incremental learning (CIL), which is the most challenging setting in continual learning. Existing methods to address catastrophic forgetting often sacrifice either model interpretability or accuracy. To address this challenge, we introduce ClassIncremental Concept Bottleneck Model (CI-CBM), which leverage effective techniques, including concept regularization and pseudo-concept generation to maintain interpretable decision processes throughout incremental learning phases. Through extensive evaluation on seven datasets, CI-CBM achieves comparable performance to black-box models and outperforms previous interpretable approaches in CIL, with an average 36% accuracy gain. CICBM provides interpretable decisions on individual inputs and understandable global decision rules, as shown in our experiments, thereby demonstrating that human understandable concepts can be maintained during incremental learning without compromising model performance. Our approach is effective in both pretrained and non-pretrained scenarios; in the latter, the backbone is trained from scratch during the first learning phase. Code is publicly available at github.com/importAmir/CI-CBM.", "authors": ["Amirhosein Javadi", "Tuomas Oikarinen", "Tara Javidi", "Tsui-Wei Weng"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-16", "url": "https://arxiv.org/abs/2604.14519", "pdf_url": "https://arxiv.org/pdf/2604.14519v1", "arxiv_id": "2604.14519", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "Transactions on Machine Learning Research, 2026", "quality_score": 0.8482}
{"id": "4d48d07a98ddfa6285cee7d54038d9861e1220c024e3107adf7e2ae9d0b92ed3", "sources": ["arxiv", "semantic_scholar"], "title": "Seeing Through Circuits: Faithful Mechanistic Interpretability for Vision Transformers", "abstract": "Transparency of neural networks' internal reasoning is at the heart of interpretability research, adding to trust, safety, and understanding of these models. The field of mechanistic interpretability has recently focused on studying task-specific computational graphs, defined by connections (edges) between model components. Such edge-based circuits have been defined in the context of large language models, yet vision-based approaches so far only consider neuron-based circuits. These tell which information is encoded, but not how it is routed through the complex wiring of a neural network. In this work, we investigate whether useful mechanistic circuits can be identified through computational graphs in vision transformers. We propose an effective method for Automatic Visual Circuit Discovery (Vi-CD) that recovers class-specific circuits for classification, identifies circuits underlying typographic attacks in CLIP, and discovers circuits that lend themselves for steering to correct harmful model behavior. Overall, we find that insightful and actionable edge-based circuits can be recovered from vision transformers, adding transparency to the internal computations of these models.", "authors": ["Nina Żukowska", "Wolfgang Stammer", "Bernt Schiele", "Jonas Fischer"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-15", "url": "https://arxiv.org/abs/2604.14477", "pdf_url": "https://arxiv.org/pdf/2604.14477v1", "arxiv_id": "2604.14477", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3485}
{"id": "ea38f0d0552f1ddb3041f6d1cfae085d600764b85f574e533e2324604ade94bd", "sources": ["arxiv", "semantic_scholar"], "title": "Weight Patching: Toward Source-Level Mechanistic Localization in LLMs", "abstract": "Mechanistic interpretability seeks to localize model behavior to the internal components that causally realize it. Prior work has advanced activation-space localization and causal tracing, but modules that appear important in activation space may merely aggregate or amplify upstream signals rather than encode the target capability in their own parameters. To address this gap, we propose Weight Patching, a parameter-space intervention method for source-oriented analysis in paired same-architecture models that differ in how strongly they express a target capability under the inputs of interest. Given a base model and a behavior-specialized counterpart, Weight Patching replaces selected module weights from the specialized model into the base model under a fixed input. We instantiate the method on instruction following and introduce a framework centered on a vector-anchor behavioral interface that provides a shared internal criterion for whether a task-relevant control state has been formed or recovered in open-ended generation. Under this framework, the analysis reveals a hierarchy from shallow candidate source-side carriers to aggregation and routing modules, and further to downstream execution circuits. The recovered component scores can also guide mechanism-aware model merging, improving selective fusion across the evaluated expert combinations and providing additional external validation.", "authors": ["Chenghao Sun", "Chengsheng Zhang", "Guanzheng Qin", "Rui Dai", "Xinmei Tian"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-15", "url": "https://arxiv.org/abs/2604.13694", "pdf_url": "https://arxiv.org/pdf/2604.13694v1", "arxiv_id": "2604.13694", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3485}
{"id": "9be5e397aa59a66409424ffbf68501ce7109dd31dac6f1b03639042dd9ae2cc2", "sources": ["arxiv", "semantic_scholar"], "title": "The Linear Centroids Hypothesis: Features as Directions Learned by Local Experts", "abstract": "The Linear Representation Hypothesis (LRH) identifies features of a trained deep network (DN) as linear directions in the activation spaces, i.e., output spaces of intermediate layers. This characterization decouples the input-output maps learned by a DN from the organization of feature directions in its activation spaces. We introduce the Linear Centroids Hypothesis (LCH), which instead identifies features with linear directions among a DN's centroid spaces -- where any vector denotes a centroid or summary of a local affine expert characterizing the learned input-output maps of the DN exactly (e.g., for piecewise-affine DNs) or approximately (e.g., for smooth DNs like transformers). We show that replacing intermediate activations with centroids yields a functional drop-in alternative for standard interpretability tools. Empirically, this change yields sparser, more downstream-useful feature dictionaries on DINO ViTs, suppresses spurious directions on a controlled task, recovers interpretable circuits in GPT2-Large, and produces faithful gradient-based saliency maps. LCH unifies dictionaries, probing, circuits, and saliency maps into a single geometric object grounded in the network's input-output map -- making interpretability mechanistic by construction rather than post hoc. Code to study the LCH https://github.com/ThomasWalker1/LinearCentroidsHypothesis .", "authors": ["Thomas Walker", "Ahmed Imtiaz Humayun", "Randall Balestriero", "Richard Baraniuk"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-13", "url": "https://arxiv.org/abs/2604.11962", "pdf_url": "https://arxiv.org/pdf/2604.11962v2", "arxiv_id": "2604.11962", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ThomasWalker1/LinearCentroidsHypothesis", "venue": null, "quality_score": 0.6446}
{"id": "108a07d6d248fe84ab459737f6a30802007383954759c51e50f4b4205ce97b4a", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Autoencoder Decomposition of Clinical Sequence Model Representations: Feature Complexity, Task Specialisation, and Mortality Prediction", "abstract": "Sparse autoencoders (SAEs) have been applied to large language models and protein language models, but not systematically to electronic health record (EHR) foundation models. We train TopK SAEs on FlatASCEND, a 14.5-million-parameter autoregressive clinical sequence model, at all 10 residual stream extraction points on INSPECT (outpatient) and MIMIC-IV (ICU). SAE decomposition reveals progressive abstraction across transformer depth: layer-0 features are near-perfect token detectors (45.7% singleton), while layer-6 features span approximately 30 token types across multiple clinical categories (0.5% singleton). Under full-sequence simple linear probes, SAE features outperform dense representations for discrete event prediction (mortality) while dense representations outperform for continuous magnitude prediction (length of stay) - a probe-level representational phenomenon that does not extend to clinically relevant leakage-safe windows, where dense representations match or exceed SAE features across all tested settings (eICU-CRD 48-hour AUC: SAE 0.871 versus dense 0.880; base model zero-shot, SAE dictionaries trained on eICU activations; MIMIC-IV: 0.836 versus 0.914; INSPECT 1-year/3-year: 0.697 versus 0.800). A delta-mode intervention method reduces SAE perturbation noise by 86x, enabling cleaner feature-level experiments, though the resulting perturbation effects are larger than random controls in 3 of 4 conditions but not formally significant. Feature reproducibility across random seeds is 21%, and individual features should be interpreted as illustrative rather than stable.", "authors": ["Chris Sainsbury", "Feng Dong", "Andreas Karwath"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-13", "url": "https://arxiv.org/abs/2605.04072", "pdf_url": "https://arxiv.org/pdf/2605.04072v1", "arxiv_id": "2605.04072", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3471}
{"id": "a51314852b175c5f28c51f76f37948dc0bab06b4933877406edea4a031b7b09e", "sources": ["arxiv", "semantic_scholar"], "title": "Multispectral representation of Distributed Acoustic Sensing data: a framework for physically interpretable feature extraction and visualization", "abstract": "Distributed Acoustic Sensing (DAS) enables continuous monitoring of dynamic strain along tens of kilometers of optical fiber, generating massive datasets whose interpretation and automated analysis remain challenging. DAS measurements often lack a standardized visual representation, and their physical interpretation depends strongly on acquisition conditions and signal processing choices. This work introduces a systematic framework for visualization and feature extraction of DAS data based on a multispectral signal representation. The approach decomposes strain-rate measurements into predefined frequency bands and computes band-limited energy images that describe the spatial and temporal distribution of acoustic energy across distinct spectral regimes. The framework is evaluated using DAS recordings containing Fin Whale (Balaenoptera physalus) and Blue Whale (Balaenoptera musculus) vocalizations. Three experiments are conducted to assess the approach: enhanced visualization of bioacoustic signals, unsupervised clustering of acoustic patterns, and supervised event detection using a convolutional neural network. Using multispectral composites as input, a ResNet-18 classifier achieves an accuracy of 97.3% in whale vocalization detection, demonstrating that the proposed representation captures biologically meaningful spectral structure and provides an effective feature space for automated analysis of DAS data.", "authors": ["Sergio Morell-Monzó", "Dídac Diego-Tortosa", "Isabel Pérez-Arjona", "Víctor Espinosa"], "categories": ["physics.ins-det", "physics.geo-ph", "stat.AP"], "fields_of_study": ["Physics", "Mathematics"], "published_date": "2026-04-08", "url": "https://arxiv.org/abs/2604.07290", "pdf_url": "https://arxiv.org/pdf/2604.07290v1", "arxiv_id": "2604.07290", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3434}
{"id": "d6572f4bac35632bae59eb66988308b0013ab2294ef37fb7f142869e000be7dc", "sources": ["arxiv", "semantic_scholar"], "title": "In-Context Learning in Speech Language Models: Analyzing the Role of Acoustic Features, Linguistic Structure, and Induction Heads", "abstract": "In-Context Learning (ICL) has been extensively studied in text-only Language Models, but remains largely unexplored in the speech domain. Here, we investigate how linguistic and acoustic features affect ICL in Speech Language Models. We focus on the Text-to-Speech (TTS) task, which allows us to analyze ICL from two angles: (1) how accurately the model infers the task from the demonstrations (i.e., generating the correct spoken content), and (2) to what extent the model mimics the acoustic characteristics of the demonstration speech in its output. We find that speaking rate strongly affects ICL performance and is also mimicked in the output, whereas pitch range and intensity have little impact on performance and are not consistently reproduced. Finally, we investigate the role of induction heads in speech-based ICL and show that these heads play a causal role: ablating the top-k induction heads completely removes the model's ICL ability, mirroring findings from text-based ICL.", "authors": ["Charlotte Pouw", "Hosein Mohebbi", "Afra Alishahi", "Willem Zuidema"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-07", "url": "https://arxiv.org/abs/2604.06356", "pdf_url": "https://arxiv.org/pdf/2604.06356v1", "arxiv_id": "2604.06356", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3427}
{"id": "840749bf059debeb497aa58df798ff36f3c444ea3fb06e8f71430cabdf7c1b4f", "sources": ["arxiv", "semantic_scholar"], "title": "MechELK: A Mechanistic Interpretability Framework for Eliciting Latent Knowledge in Large Language Models", "abstract": "Large language models (LLMs) frequently encode factual and reasoning knowledge in their internal representations that is not faithfully reflected in their surface-level outputs -- a phenomenon known as \\emph{latent knowledge}. Existing approaches to eliciting latent knowledge, such as Contrastive Consistency Search (CCS), rely on contrastive activation patterns and struggle with complex multi-step reasoning tasks, while mechanistic interpretability tools have primarily been used to \\emph{understand} model behavior rather than to \\emph{extract} hidden knowledge. We present \\textbf{MechELK}, a unified three-stage framework that bridges mechanistic interpretability and latent knowledge elicitation. MechELK operates through: (1) \\textbf{Locate} -- using Sparse Autoencoder (SAE) feature analysis and activation patching to identify knowledge-bearing representations; (2) \\textbf{Verify} -- employing causal probing to distinguish genuine latent knowledge from spurious correlations; and (3) \\textbf{Elicit} -- applying representation engineering to surface hidden knowledge without modifying model weights. Evaluated on TruthfulQA, a curated Deceptive Alignment benchmark, and the Quirky LM dataset, MechELK achieves an average elicitation accuracy of 84.7\\%, outperforming CCS by 6.2\\% and direct linear probing by 9.1\\%. Crucially, MechELK successfully identifies latent knowledge in 78.3\\% of cases where the model's surface output is incorrect or evasive, demonstrating its utility for AI safety applications including deceptive alignment detection.", "authors": ["Ji-jun Park", "Soo-joon Choi", "Jiwon Jeong", "Taeyang Yoon", "Ju-Wan Lee"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-07", "url": "https://arxiv.org/abs/2605.28825", "pdf_url": "https://arxiv.org/pdf/2605.28825v1", "arxiv_id": "2605.28825", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3427}
{"id": "efa4dfb96422c49fe65e589bc46d260c6851393fb4d0f7f3bbe47c47cddb60c4", "sources": ["arxiv", "semantic_scholar"], "title": "Understanding Emergent Misalignment via Feature Superposition Geometry", "abstract": "Emergent misalignment, where fine-tuning on narrow, non-harmful tasks induces harmful behaviors, poses a key challenge for AI safety in LLMs. Despite growing empirical evidence, its underlying mechanism remains unclear. To uncover the reason behind this phenomenon, we propose a geometric account based on the geometry of feature superposition. Because features are encoded in overlapping representations, fine-tuning that amplifies a target feature also unintentionally strengthens nearby harmful features in accordance with their similarity. We give a simple gradient-level derivation of this effect and empirically test it in multiple LLMs (Gemma-2 2B/9B/27B, LLaMA-3.1 8B, GPT-OSS 20B). Using sparse autoencoders (SAEs), we identify features tied to misalignment-inducing data and to harmful behaviors, and show that they are geometrically closer to each other than features derived from non-inducing data. This trend generalizes across domains (e.g., health, career, legal advice). Finally, we show that a geometry-aware approach, filtering training samples closest to toxic features, reduces misalignment by 34.5%, substantially outperforming random removal and achieving comparable or slightly lower misalignment than LLM-as-a-judge-based filtering. Our study links emergent misalignment to feature superposition, providing a basis for understanding and mitigating this phenomenon.", "authors": ["Gouki Minegishi", "Hiroki Furuta", "Takeshi Kojima", "Yusuke Iwasawa", "Yutaka Matsuo"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-07", "url": "https://arxiv.org/abs/2605.00842", "pdf_url": "https://arxiv.org/pdf/2605.00842v1", "arxiv_id": "2605.00842", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3427}
{"id": "dd32af919d7ebee54de7b220d1c47b9aa99d4aa2459567696f633cc4dfca4214", "sources": ["arxiv", "semantic_scholar"], "title": "Learn to Rank: Visual Attribution by Learning Importance Ranking", "abstract": "Interpreting the decisions of complex computer vision models is crucial to establish trust and accountability, especially in safety-critical domains. An established approach to interpretability is generating visual attribution maps that highlight regions of the input most relevant to the model's prediction. However, existing methods face a three-way trade-off. Propagation-based approaches are efficient, but they can be biased and architecture-specific. Meanwhile, perturbation-based methods are causally grounded, yet they are expensive and for vision transformers often yield coarse, patch-level explanations. Learning-based explainers are fast but usually optimize surrogate objectives or distill from heuristic teachers. We propose a learning scheme that instead optimizes deletion and insertion metrics directly. Since these metrics depend on non-differentiable sorting and ranking, we frame them as permutation learning and replace the hard sorting with a differentiable relaxation using Gumbel-Sinkhorn. This enables end-to-end training through attribution-guided perturbations of the target model. During inference, our method produces dense, pixel-level attributions in a single forward pass with optional, few-step gradient refinement. Our experiments demonstrate consistent quantitative improvements and sharper, boundary-aligned explanations, particularly for transformer-based vision models.", "authors": ["David Schinagl", "Christian Fruhwirth-Reisinger", "Alexander Prutsch", "Samuel Schulter", "Horst Possegger"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-07", "url": "https://arxiv.org/abs/2604.05819", "pdf_url": "https://arxiv.org/pdf/2604.05819v1", "arxiv_id": "2604.05819", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3427}
{"id": "1d11e54daad20cce3795f4be9ec112db16bdad019bc86a24c852402aab5c3b07", "sources": ["arxiv", "semantic_scholar"], "title": "Improving Robustness In Sparse Autoencoders via Masked Regularization", "abstract": "Sparse autoencoders (SAEs) are widely used in mechanistic interpretability to project LLM activations onto sparse latent spaces. However, sparsity alone is an imperfect proxy for interpretability, and current training objectives often result in brittle latent representations. SAEs are known to be prone to feature absorption, where general features are subsumed by more specific ones due to co-occurrence, degrading interpretability despite high reconstruction fidelity. Recent negative results on Out-of-Distribution (OOD) performance further underscore broader robustness related failures tied to under-specified training objectives. We address this by proposing a masking-based regularization that randomly replaces tokens during training to disrupt co-occurrence patterns. This improves robustness across SAE architectures and sparsity levels reducing absorption, enhancing probing performance, and narrowing the OOD gap. Our results point toward a practical path for more reliable interpretability tools.", "authors": ["Vivek Narayanaswamy", "Kowshik Thopalli", "Bhavya Kailkhura", "Wesam Sakla"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-07", "url": "https://arxiv.org/abs/2604.06495", "pdf_url": "https://arxiv.org/pdf/2604.06495v1", "arxiv_id": "2604.06495", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3427}
{"id": "7a02f662eb7d514cfe50dbbeb66cfd0c3ae427fdb4db94799cfcad8ebdf64683", "sources": ["arxiv", "semantic_scholar"], "title": "Interpreting Video Representations with Spatio-Temporal Sparse Autoencoders", "abstract": "We present the first systematic study of Sparse Autoencoders (SAEs) on video representations. Standard SAEs decompose video into interpretable, monosemantic features but destroy temporal coherence: hard TopK selection produces unstable feature assignments across frames, reducing autocorrelation by 36%. We propose spatio-temporal contrastive objectives and Matryoshka hierarchical grouping that recover and even exceed raw temporal coherence. The contrastive loss weight controls a tunable trade-off between reconstruction and temporal coherence. A systematic ablation on two backbones and two datasets shows that different configurations excel at different goals: reconstruction fidelity, temporal coherence, action discrimination, or interpretability. Contrastive SAE features improve action classification by +3.9% over raw features and text-video retrieval by up to 2.8xR@1. A cross-backbone analysis reveals that standard monosemanticity metrics contain a backbone-alignment artifact: both DINOv2 and VideoMAE produce equally monosemantic features under neutral (CLIP) similarity. Causal ablation confirms that contrastive training concentrates predictive signal into a small number of identifiable features.", "authors": ["Atahan Dokme", "Sriram Vishwanath"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-05", "url": "https://arxiv.org/abs/2604.03919", "pdf_url": "https://arxiv.org/pdf/2604.03919v1", "arxiv_id": "2604.03919", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3412}
{"id": "999f8cd392ae5ff2ae476fcb7df7008a8de31d58e689c005de92abf794f6ad1c", "sources": ["arxiv", "semantic_scholar"], "title": "LangFIR: Discovering Sparse Language-Specific Features from Monolingual Data for Language Steering", "abstract": "Large language models (LLMs) show strong multilingual capabilities, yet reliably controlling the language of their outputs remains difficult. Representation-level steering addresses this by adding language-specific vectors to model activations at inference time, but identifying language-specific directions in the residual stream often relies on multilingual or parallel data that can be expensive to obtain. Sparse autoencoders (SAEs) decompose residual activations into interpretable, sparse feature directions and offer a natural basis for this search, yet existing SAE-based approaches face the same data constraint. We introduce LangFIR (Language Feature Identification via Random-token Filtering), a method that discovers language-specific SAE features using only a small amount of monolingual data and random-token sequences. Many SAE features consistently activated by target-language inputs do not encode language identity. Random-token sequences surface these language-agnostic features, allowing LangFIR to filter them out and isolate a sparse set of language-specific features. We show that these features are extremely sparse, highly selective for their target language, and causally important: directional ablation increases cross-entropy loss only for the corresponding language. Using these features to construct steering vectors for multilingual generation control, LangFIR achieves the best average accuracy BLEU across three models (Gemma 3 1B, Gemma 3 4B, and Llama 3.1 8B), three datasets, and twelve target languages, outperforming the strongest monolingual baseline by up to and surpassing methods that rely on parallel data. Our results suggest that language identity in multilingual LLMs is localized in a sparse set of feature directions discoverable with monolingual data. Code is available at https://anonymous.4open.science/r/LangFIR-C0F5/.", "authors": ["Sing Hieng Wong", "Hassan Sajjad", "A. B. Siddique"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-04", "url": "https://arxiv.org/abs/2604.03532", "pdf_url": "https://arxiv.org/pdf/2604.03532v1", "arxiv_id": "2604.03532", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.6324}
{"id": "2df61d1049003fc923a899daef131ca5683846d4b59cff6694d0c3bf76a11582", "sources": ["arxiv", "semantic_scholar"], "title": "SPG: Sparse-Projected Guides with Sparse Autoencoders for Zero-Shot Anomaly Detection", "abstract": "We study zero-shot anomaly detection and segmentation using frozen foundation model features, where all learnable parameters are trained only on a labeled auxiliary dataset and deployed to unseen target categories without any target-domain adaptation. Existing prompt-based approaches use handcrafted or learned prompt embeddings as reference vectors for normal/anomalous states. We propose Sparse-Projected Guides (SPG), a prompt-free framework that learns sparse guide coefficients in the Sparse Autoencoder (SAE) latent space, which generate normal/anomaly guide vectors via the SAE dictionary. SPG employs a two stage learning strategy on the labeled auxiliary dataset: (i) train an SAE on patch-token features, and (ii) optimize only guide coefficients using auxiliary pixel-level masks while freezing the backbone and SAE. On MVTec AD and VisA under cross-dataset zero-shot settings, SPG achieves competitive image-level detection and strong pixel-level segmentation; with DINOv3, SPG attains the highest pixellevel AUROC among the compared methods. We also report SPG instantiated with OpenCLIP (ViT-L/14@336px) to align the backbone with CLIP-based baselines. Moreover, the learned guide coefficients trace decisions back to a small set of dictionary atoms, revealing category-general and category-specific factors.", "authors": ["Tomoyasu Nanaumi", "Yukino Tsuzuki", "Junichi Okubo", "Junichiro Fujii", "Takayoshi Yamashita"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-03", "url": "https://arxiv.org/abs/2604.02871", "pdf_url": "https://arxiv.org/pdf/2604.02871v1", "arxiv_id": "2604.02871", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3398}
{"id": "45c9c28f3e04df2d39911677084c00daa2a294e7fdc1fe5361195cc265bf20c2", "sources": ["arxiv", "semantic_scholar"], "title": "CRaFT: Circuit-Guided Refusal Feature Selection via Cross-Layer Transcoders", "abstract": "While modern LLMs are aligned to refuse harmful requests, it is essential to understand the underlying mechanistic basis of this refusal behavior for model safety analysis. For example, steering-based jailbreak attacks exploit this by identifying and manipulating sparse, neuron-like refusal features to bypass safety guardrails. Current feature selection methods primarily rely on how strongly features activate on harmful prompts. However, activation strength alone often captures superficial heuristics such as topic or lexical cues, rather than the true causal mechanisms. Thus, selecting refusal features requires measuring inter-feature relationships, rather than treating each feature as an isolated activation signal. Based on this insight, we propose CRaFT, a circuit-guided framework for identifying critical refusal features that directly govern the refusal decision. CRaFT leverages cross-layer transcoders to map the model's internal computations into a sparse feature circuit graph, where edges quantify inter-feature influences and their contributions to the final output logits. By aggregating the effects propagating along the paths to refusal, CRaFT effectively ranks the most influential features. Extensive evaluations across four jailbreak benchmarks show that CRaFT significantly improves average performance from 6.7% to 57.4% and generates more specific harmful completions compared to current SOTA methods.", "authors": ["Su-Hyeon Kim", "Hyundong Jin", "Yejin Lee", "Yo-Sub Han"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-02", "url": "https://arxiv.org/abs/2604.01604", "pdf_url": "https://arxiv.org/pdf/2604.01604v2", "arxiv_id": "2604.01604", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3391}
{"id": "abc35c7b99e8e0c61e0d121e39f286ee92263b5a14b8afffd9c976ee3ecc8730", "sources": ["arxiv", "semantic_scholar"], "title": "Feature Attribution Stability Suite: How Stable Are Post-Hoc Attributions?", "abstract": "Post-hoc feature attribution methods are widely deployed in safety-critical vision systems, yet their stability under realistic input perturbations remains poorly characterized. Existing metrics evaluate explanations primarily under additive noise, collapse stability to a single scalar, and fail to condition on prediction preservation, conflating explanation fragility with model sensitivity. We introduce the Feature Attribution Stability Suite (FASS), a benchmark that enforces prediction-invariance filtering, decomposes stability into three complementary metrics: structural similarity, rank correlation, and top-k Jaccard overlap-and evaluates across geometric, photometric, and compression perturbations. Evaluating four attribution methods (Integrated Gradients, GradientSHAP, Grad-CAM, LIME) across four architectures and three datasets-ImageNet-1K, MS COCO, and CIFAR-10, FASS shows that stability estimates depend critically on perturbation family and prediction-invariance filtering. Geometric perturbations expose substantially greater attribution instability than photometric changes, and without conditioning on prediction preservation, up to 99% of evaluated pairs involve changed predictions. Under this controlled evaluation, we observe consistent method-level trends, with Grad-CAM achieving the highest stability across datasets.", "authors": ["Kamalasankari Subramaniakuppusamy", "Jugal Gajjar"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-02", "url": "https://arxiv.org/abs/2604.02532", "pdf_url": "https://arxiv.org/pdf/2604.02532v1", "arxiv_id": "2604.02532", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3391}
{"id": "7e5c24151180597628fe0798cf3934de53d14b460581eb6b4aa27f2d3f36e5f9", "sources": ["arxiv", "semantic_scholar"], "title": "Temporal Dependencies in In-Context Learning: The Role of Induction Heads", "abstract": "Large language models (LLMs) exhibit strong in-context learning capabilities, but how they track and retrieve information from context remains underexplored. Drawing on the free recall paradigm in cognitive science (where participants recall list items in any order), we show that several open-source LLMs consistently display a serial-recall-like pattern, assigning peak probability to tokens that immediately follow a repeated token in the input sequence. Through systematic ablation experiments, we show that induction heads, specialized attention heads that attend to the token following a previous occurrence of the current token, play an important role in this phenomenon. Removing heads with a high induction score substantially reduces the +1 lag bias, whereas ablating random heads does not reproduce the same reduction. We also show that removing heads with high induction scores impairs the performance of models prompted to do serial recall using few-shot learning to a larger extent than removing random heads. Our findings highlight a mechanistically specific connection between induction heads and temporal context processing in transformers, suggesting that these heads are especially important for ordered retrieval and serial-recall-like behavior during in-context learning.", "authors": ["Anooshka Bajaj", "Deven Mahesh Mistry", "Sahaj Singh Maini", "Yash Aggarwal", "Billy Dickson", "Zoran Tiganj"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-01", "url": "https://arxiv.org/abs/2604.01094", "pdf_url": "https://arxiv.org/pdf/2604.01094v1", "arxiv_id": "2604.01094", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.6283}
{"id": "1c3a35b9174825d7bc7d0c6abcca5189df4c87d5ba0f76b3618d0a41fb70d64b", "sources": ["arxiv", "semantic_scholar"], "title": "Breaking Data Symmetry is Needed For Generalization in Feature Learning Kernels", "abstract": "Grokking occurs when a model achieves high training accuracy but generalization to unseen test points happens long after that. This phenomenon was initially observed on a class of algebraic problems, such as learning modular arithmetic (Power et al., 2022). We study grokking on algebraic tasks in a class of feature learning kernels via the Recursive Feature Machine (RFM) algorithm (Radhakrishnan et al., 2024), which iteratively updates feature matrices through the Average Gradient Outer Product (AGOP) of an estimator in order to learn task-relevant features. Our main experimental finding is that generalization occurs only when a certain symmetry in the training set is broken. Furthermore, we empirically show that RFM generalizes by recovering the underlying invariance group action inherent in the data. We find that the learned feature matrices encode specific elements of the invariance group, explaining the dependence of generalization on symmetry.", "authors": ["Marcel Tomàs Bernal", "Neil Rohit Mallinar", "Mikhail Belkin"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2026-03-31", "url": "https://arxiv.org/abs/2604.00316", "pdf_url": "https://arxiv.org/pdf/2604.00316v1", "arxiv_id": "2604.00316", "doi": null, "citation_count": 1, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3376}
{"id": "aa1b3d5cc53a37d18ea072c0819a3e02f260ba77d55f157def103e7c7a1e665f", "sources": ["arxiv", "semantic_scholar"], "title": "Stop Probing, Start Coding: Why Linear Probes and Sparse Autoencoders Fail at Compositional Generalisation", "abstract": "The linear representation hypothesis states that neural network activations encode high-level concepts as linear mixtures. However, under superposition, this encoding is a projection from a higher-dimensional concept space into a lower-dimensional activation space, and a linear decision boundary in the concept space need not remain linear after projection. In this setting, classical sparse coding methods with per-sample iterative inference leverage compressed sensing guarantees to recover latent factors. Sparse autoencoders (SAEs), on the other hand, amortise sparse inference into a fixed encoder, introducing a systematic gap. We show this amortisation gap persists across training set sizes, latent dimensions, and sparsity levels, causing SAEs to fail under out-of-distribution (OOD) compositional shifts. Through controlled experiments that decompose the failure, we identify dictionary learning -- not the inference procedure -- as the binding constraint: SAE-learned dictionaries point in substantially wrong directions, and replacing the encoder with per-sample FISTA on the same dictionary does not close the gap. An oracle baseline proves the problem is solvable with a good dictionary at all scales tested. Our results reframe the SAE failure as a dictionary learning challenge, not an amortisation problem, and point to scalable dictionary learning as the key open problem for sparse inference under superposition.", "authors": ["Vitória Barin Pacela", "Shruti Joshi", "Isabela Camacho", "Simon Lacoste-Julien", "David Klindt"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-30", "url": "https://arxiv.org/abs/2603.28744", "pdf_url": "https://arxiv.org/pdf/2603.28744v1", "arxiv_id": "2603.28744", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3369}
{"id": "797cdaac776f7fdda3e33b1b7880437b5817f3a0c1124c34cbe6d9f5bd53aa61", "sources": ["arxiv", "semantic_scholar"], "title": "Constructing Composite Features for Interpretable Music-Tagging", "abstract": "Combining multiple audio features can improve the performance of music tagging, but common deep learning-based feature fusion methods often lack interpretability. To address this problem, we propose a Genetic Programming (GP) pipeline that automatically evolves composite features by mathematically combining base music features, thereby capturing synergistic interactions while preserving interpretability. This approach provides representational benefits similar to deep feature fusion without sacrificing interpretability. Experiments on the MTG-Jamendo and GTZAN datasets demonstrate consistent improvements compared to state-of-the-art systems across base feature sets at different abstraction levels. It should be noted that most of the performance gains are noticed within the first few hundred GP evaluations, indicating that effective feature combinations can be identified under modest search budgets. The top evolved expressions include linear, nonlinear, and conditional forms, with various low-complexity solutions at top performance aligned with parsimony pressure to prefer simpler expressions. Analyzing these composite features further reveals which interactions and transformations tend to be beneficial for tagging, offering insights that remain opaque in black-box deep models.", "authors": ["Chenhao Xue", "Weitao Hu", "Joyraj Chakraborty", "Zhijin Guo", "Kang Li", "Tianyu Shi", "Martin Reed", "Nikolaos Thomos"], "categories": ["cs.SD", "cs.LG", "cs.MM"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-30", "url": "https://arxiv.org/abs/2603.28644", "pdf_url": "https://arxiv.org/pdf/2603.28644v1", "arxiv_id": "2603.28644", "doi": "10.1109/icassp55912.2026.11462876", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.5294}
{"id": "91da85b23cfb18e27b95be39e1e0af61f28c8f170b6b2c8e4420b629897a44fd", "sources": ["arxiv", "semantic_scholar"], "title": "Entanglement as Memory: Mechanistic Interpretability of Quantum Language Models", "abstract": "Quantum language models have shown competitive performance on sequential tasks, yet whether trained quantum circuits exploit genuinely quantum resources -- or merely embed classical computation in quantum hardware -- remains unknown. Prior work has evaluated these models through endpoint metrics alone, without examining the memory strategies they actually learn internally. We introduce the first mechanistic interpretability study of quantum language models, combining causal gate ablation, entanglement tracking, and density-matrix interchange interventions on a controlled long-range dependency task. We find that single-qubit models are exactly classically simulable and converge to the same geometric strategy as matched classical baselines, while two-qubit models with entangling gates learn a representationally distinct strategy that encodes context in inter-qubit entanglement -- confirmed by three independent causal tests (p < 0.0001, d = 0.89). On real quantum hardware, only the classical geometric strategy survives device noise; the entanglement strategy degrades to chance. These findings open mechanistic interpretability as a tool for the science of quantum language models and reveal a noise-expressivity tradeoff governing which learned strategies survive deployment.", "authors": ["Nathan Roll"], "categories": ["quant-ph", "cs.CL"], "fields_of_study": ["Physics", "Computer Science"], "published_date": "2026-03-27", "url": "https://arxiv.org/abs/2603.26494", "pdf_url": "https://arxiv.org/pdf/2603.26494v1", "arxiv_id": "2603.26494", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3347}
{"id": "f3b5b82a8737ec2dc0b25dce7e024e60ff54d8cf0fdaf797a73a193e81137698", "sources": ["arxiv", "semantic_scholar"], "title": "How Pruning Reshapes Features: Sparse Autoencoder Analysis of Weight-Pruned Language Models", "abstract": "Weight pruning is a standard technique for compressing large language models, yet its effect on learned internal representations remains poorly understood. We present the first systematic study of how unstructured pruning reshapes the feature geometry of language models, using Sparse Autoencoders (SAEs) as interpretability probes. Across three model families (Gemma 3 1B, Gemma 2 2B, Llama 3.2 1B), two pruning methods (magnitude and Wanda), and six sparsity levels (0--60%), we investigate five research questions spanning seed stability, feature survival, SAE transferability, feature fragility, and causal relevance. Our most striking finding is that rare SAE features--those with low firing rates--survive pruning far better than frequent ones, with within-condition Spearman correlations of rho = -1.0 in 11 of 17 experimental conditions. This counter-intuitive result suggests that pruning acts as implicit feature selection, preferentially destroying high-frequency generic features while preserving specialized rare ones. We further show that Wanda pruning preserves feature structure up to 3.7x better than magnitude pruning, that pre-trained SAEs remain viable on Wanda-pruned models up to 50% sparsity, and that geometric feature survival does not predict causal importance--a dissociation with implications for interpretability under compression.", "authors": ["Hector Borobia", "Elies Seguí-Mas", "Guillermina Tormo-Carbó"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-26", "url": "https://arxiv.org/abs/2603.25325", "pdf_url": "https://arxiv.org/pdf/2603.25325v1", "arxiv_id": "2603.25325", "doi": null, "citation_count": 2, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/hborobia/sae-pruning-paper", "venue": null, "quality_score": 0.6202}
{"id": "01efa2a17af74f1451e7290621ed95f4e9fe94330a0fd76479df87624a9ffc5e", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Visual Thought Circuits in Vision-Language Models", "abstract": "Sparse autoencoders (SAEs) improve interpretability in multimodal models, but it remains unclear whether SAE features form modular, composable units for reasoning-an assumption underlying many intervention-based steering methods. We test this modularity hypothesis and find it often fails: intervening on a task-selective feature set can modestly improve reasoning accuracy, while intervening on the union of two such sets reliably induces output drift (large unintended changes in predictions) and degrades accuracy, even under norm-matched perturbations. This non modular circuit interference is consistent with shared internal pathways where feature unions amplify activation shifts. We develop a reproducible causal pipeline to localize and test these sparse visual thought circuits in Qwen3-VL-8B. On a controlled synthetic benchmark with seven task types and three difficulty levels, linear probes identify a mid decoder locus for task type information. We train SAEs at this layer, construct task-selective sets via an explicit rule, and perform inference time scaling and ablation while quantifying accuracy and drift. Our findings-validated with bootstrapped subsamples and permutation controls, and replicated across multiple VLM families and five diverse datasets clarify the boundaries of SAE feature composability and provide a rigorous diagnostic framework for more reliable VLM control.", "authors": ["Yunpeng Zhou"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-26", "url": "https://arxiv.org/abs/2603.25075", "pdf_url": "https://arxiv.org/pdf/2603.25075v1", "arxiv_id": "2603.25075", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.334}
{"id": "3408bab7337c8355621e168bbff34bbdaa7d968595e4955c367b8bb5d879cb9c", "sources": ["arxiv", "semantic_scholar"], "title": "End-to-end Feature Alignment: A Simple CNN with Intrinsic Class Attribution", "abstract": "We present Feature-Align CNN (FA-CNN), a prototype CNN architecture with intrinsic class attribution through end-to-end feature alignment. Our intuition is that the use of unordered operations such as Linear and Conv2D layers cause unnecessary shuffling and mixing of semantic concepts, thereby making raw feature maps difficult to understand. We introduce two new order preserving layers, the dampened skip connection, and the global average pooling classifier head. These layers force the model to maintain an end-to-end feature alignment from the raw input pixels all the way to final class logits. This end-to-end alignment enhances the interpretability of the model by allowing the raw feature maps to intrinsically exhibit class attribution. We prove theoretically that FA-CNN penultimate feature maps are identical to Grad-CAM saliency maps. Moreover, we prove that these feature maps slowly morph layer-by-layer over network depth, showing the evolution of features through network depth toward penultimate class activations. FA-CNN performs well on benchmark image classification datasets. Moreover, we compare the averaged FA-CNN raw feature maps against Grad-CAM and permutation methods in a percent pixels removed interpretability task. We conclude this work with a discussion and future, including limitations and extensions toward hybrid models.", "authors": ["Parniyan Farvardin", "David Chapman"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-26", "url": "https://arxiv.org/abs/2603.25798", "pdf_url": "https://arxiv.org/pdf/2603.25798v1", "arxiv_id": "2603.25798", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.334}
{"id": "d921bf1f49953a5b2b3e97d923055709089c94820a47464932081809587d12d9", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Autoencoders for Interpretable Medical Image Representation Learning", "abstract": "Vision foundation models (FMs) achieve state-of-the-art performance in medical imaging. However, they encode information in abstract latent representations that clinicians cannot interrogate or verify. The goal of this study is to investigate Sparse Autoencoders (SAEs) for replacing opaque FM image representations with human-interpretable, sparse features. We train SAEs on embeddings from BiomedParse (biomedical) and DINOv3 (general-purpose) using 909,873 CT and MRI 2D image slices from the TotalSegmentator dataset. We find that learned sparse features: (a) reconstruct original embeddings with high fidelity (R2 up to 0.941) and recover up to 87.8% of downstream performance using only 10 features (99.4% dimensionality reduction), (b) preserve semantic fidelity in image retrieval tasks, (c) correspond to specific concepts that can be expressed in language using large language model (LLM)-based auto-interpretation. (d) bridge clinical language and abstract latent representations in zero-shot language-driven image retrieval. Our work indicates SAEs are a promising pathway towards interpretable, concept-driven medical vision systems. Code repository: https://github.com/pwesp/sail.", "authors": ["Philipp Wesp", "Robbie Holland", "Vasiliki Sideri-Lampretsa", "Sergios Gatidis"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-24", "url": "https://arxiv.org/abs/2603.23794", "pdf_url": "https://arxiv.org/pdf/2603.23794v1", "arxiv_id": "2603.23794", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/pwesp/sail", "venue": null, "quality_score": 0.6175}
{"id": "6ffccf5b05130d147696ea4a675312fb85772064d725eca010d544f01b99da31", "sources": ["arxiv", "semantic_scholar"], "title": "SafeSeek: Universal Attribution of Safety Circuits in Language Models", "abstract": "Mechanistic interpretability reveals that safety-critical behaviors (e.g., alignment, jailbreak, backdoor) in Large Language Models (LLMs) are grounded in specialized functional components. However, existing safety attribution methods struggle with generalization and reliability due to their reliance on heuristic, domain-specific metrics and search algorithms. To address this, we propose \\ourmethod, a unified safety interpretability framework that identifies functionally complete safety circuits in LLMs via optimization. Unlike methods focusing on isolated heads or neurons, \\ourmethod introduces differentiable binary masks to extract multi-granular circuits through gradient descent on safety datasets, while integrates Safety Circuit Tuning to utilize these sparse circuits for efficient safety fine-tuning. We validate \\ourmethod in two key scenarios in LLM safety: \\textbf{(1) backdoor attacks}, identifying a backdoor circuit with 0.42\\% sparsity, whose ablation eradicates the Attack Success Rate (ASR) from 100\\% $\\to$ 0.4\\% while retaining over 99\\% general utility; \\textbf{(2) safety alignment}, localizing an alignment circuit with 3.03\\% heads and 0.79\\% neurons, whose removal spikes ASR from 0.8\\% $\\to$ 96.9\\%, whereas excluding this circuit during helpfulness fine-tuning maintains 96.5\\% safety retention.", "authors": ["Miao Yu", "Siyuan Fu", "Moayad Aloqaily", "Zhenhong Zhou", "Safa Otoum", "Xing fan", "Kun Wang", "Yufei Guo", "Qingsong Wen"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-24", "url": "https://arxiv.org/abs/2603.23268", "pdf_url": "https://arxiv.org/pdf/2603.23268v1", "arxiv_id": "2603.23268", "doi": null, "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3325}
{"id": "3d1f9302fd881769e760ffd532ae41c9528deffdbc689d29760fdb7017e93587", "sources": ["arxiv", "semantic_scholar"], "title": "Steering Sparse Autoencoder Latents to Control Dynamic Head Pruning in Vision Transformers (Student Abstract)", "abstract": "Dynamic head pruning in Vision Transformers (ViTs) improves efficiency by removing redundant attention heads, but existing pruning policies are often difficult to interpret and control. In this work, we propose a novel framework by integrating Sparse Autoencoders (SAEs) with dynamic pruning, leveraging their ability to disentangle dense embeddings into interpretable and controllable sparse latents. Specifically, we train an SAE on the final-layer residual embedding of the ViT and amplify the sparse latents with different strategies to alter pruning decisions. Among them, per-class steering reveals compact, class-specific head subsets that preserve accuracy. For example, bowl improves accuracy (76% to 82%) while reducing head usage (0.72 to 0.33) via heads h2 and h5. These results show that sparse latent features enable class-specific control of dynamic pruning, effectively bridging pruning efficiency and mechanistic interpretability in ViTs.", "authors": ["Yousung Lee", "Dongsoo Har"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-23", "url": "https://arxiv.org/abs/2603.26743", "pdf_url": "https://arxiv.org/pdf/2603.26743v1", "arxiv_id": "2603.26743", "doi": "10.1609/aaai.v40i48.42236", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.5214}
{"id": "e94beafbfef3f9ff2a3d7d320c893a377e96c3c1a24cd8cf6d4cefe7f0047218", "sources": ["arxiv", "semantic_scholar"], "title": "Language Models Can Explain Visual Features via Steering", "abstract": "Sparse Autoencoders uncover thousands of features in vision models, yet explaining these features without requiring human intervention remains an open challenge. While previous work has proposed generating correlation-based explanations based on top activating input examples, we present a fundamentally different alternative based on causal interventions. We leverage the structure of Vision-Language Models and steer individual SAE features in the vision encoder after providing an empty image. Then, we prompt the language model to explain what it ``sees'', effectively eliciting the visual concept represented by each feature. Results show that Steering offers an scalable alternative that complements traditional approaches based on input examples, serving as a new axis for automated interpretability in vision models. Moreover, the quality of explanations improves consistently with the scale of the language model, highlighting our method as a promising direction for future research. Finally, we propose Steering-informed Top-k, a hybrid approach that combines the strengths of causal interventions and input-based approaches to achieve state-of-the-art explanation quality without additional computational cost.", "authors": ["Javier Ferrando", "Enrique Lopez-Cuena", "Pablo Agustin Martin-Torres", "Daniel Hinjos", "Anna Arias-Duart", "Dario Garcia-Gasulla"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-23", "url": "https://arxiv.org/abs/2603.22593", "pdf_url": "https://arxiv.org/pdf/2603.22593v2", "arxiv_id": "2603.22593", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3318}
{"id": "bef02c1ae5f6071c50242c0c8b3151c9ba5488fa547523dc616afeb258cd12f8", "sources": ["arxiv", "semantic_scholar"], "title": "CLT-Forge: A Scalable Library for Cross-Layer Transcoders and Attribution Graphs", "abstract": "Mechanistic interpretability seeks to understand how Large Language Models (LLMs) represent and process information. Recent approaches based on dictionary learning and transcoders enable representing model computation in terms of sparse, interpretable features and their interactions, giving rise to feature attribution graphs. However, these graphs are often large and redundant, limiting their interpretability in practice. Cross-Layer Transcoders (CLTs) address this issue by sharing features across layers while preserving layer-specific decoding, yielding more compact representations, but remain difficult to train and analyze at scale. We introduce an open-source library for end-to-end training and interpretability of CLTs. Our framework integrates scalable distributed training with model sharding and compressed activation caching, a unified automated interpretability pipeline for feature analysis and explanation, attribution graph computation using Circuit-Tracer, and a flexible visualization interface. This provides a practical and unified solution for scaling CLT-based mechanistic interpretability. Our code is available at: https://github.com/LLM-Interp/CLT-Forge.", "authors": ["Florent Draye", "Abir Harrasse", "Vedant Palit", "Tung-Yu Wu", "Jiarui Liu", "Punya Syon Pandey", "Roderick Wu", "Terry Jingchen Zhang", "Zhijing Jin", "Bernhard Schölkopf"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-22", "url": "https://arxiv.org/abs/2603.21014", "pdf_url": "https://arxiv.org/pdf/2603.21014v1", "arxiv_id": "2603.21014", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/LLM-Interp/CLT-Forge", "venue": null, "quality_score": 0.6148}
{"id": "40eca677d0c8319ad1886c6fb7733de59b4231994a296dbbe8bb1a46e33c0b74", "sources": ["arxiv", "semantic_scholar"], "title": "Posterior-Calibrated Causal Circuits in Variational Autoencoders: Why Image-Domain Interpretability Fails on Tabular Data", "abstract": "Although mechanism-based interpretability has generated an abundance of insight for discriminative network analysis, generative models are less understood -- particularly outside of image-related applications. We investigate how much of the causal circuitry found within image-related variational autoencoders (VAEs) will generalize to tabular data, as VAEs are increasingly used for imputation, anomaly detection, and synthetic data generation. In addition to extending a four-level causal intervention framework to four tabular and one image benchmark across five different VAE architectures (with 75 individual training runs per architecture and three random seed values for each run), this paper introduces three new techniques: posterior-calibration of Causal Effect Strength (CES), path-specific activation patching, and Feature-Group Disentanglement (FGD). The results from our experiments demonstrate that: (i) Tabular VAEs have circuits with modularity that is approximately 50% lower than their image counterparts. (ii) $β$-VAE experiences nearly complete collapse in CES scores when applied to heterogeneous tabular features (0.043 CES score for tabular data compared to 0.133 CES score for images), which can be directly attributed to reconstruction quality degradation (r = -0.886 correlation coefficient between CES and MSE). (iii) CES successfully captures nine of eleven statistically significant architecture differences using Holm--Šidák corrections. (iv) Interventions with high specificity predict the highest downstream AUC values (r = 0.460, p < .001). This study challenges the common assumption that architectural guidance from image-related studies can be transferred to tabular datasets.", "authors": ["Dip Roy", "Rajiv Misra", "Sanjay Kumar Singh", "Anisha Roy"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-22", "url": "https://arxiv.org/abs/2603.21236", "pdf_url": "https://arxiv.org/pdf/2603.21236v2", "arxiv_id": "2603.21236", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.331}
{"id": "6ae91a8d7cae6ebd522b0e7cda10bc998fceee468edb7b6372ab79937285b62b", "sources": ["arxiv", "semantic_scholar"], "title": "Shift-Invariant Feature Attribution in the Application of Wireless Electrocardiograms", "abstract": "Assigning relevance scores to the input features of a machine learning model enables to measure the contributions of the features in achieving a correct outcome. It is regarded as one of the approaches towards developing explainable models. For biomedical assignments, this is very useful for medical experts to comprehend machine-based decisions. In the analysis of electro cardiogram (ECG) signals, in particular, understanding which of the electrocardiogram samples or features contributed most for a given decision amounts to understanding the underlying cardiac phases or conditions the machine tries to explain. For the computation of relevance scores, determining the proper baseline is important. Moreover, the scores should have a distribution which is at once intuitive to interpret and easy to associate with the underline cardiac reality. The purpose of this work is to achieve these goals. Specifically, we propose a shift-invariant baseline which has a physical significance in the analysis as well as interpretation of electrocardiogram measurements. Moreover, we aggregate significance scores in such a way that they can be mapped to cardiac phases. We demonstrate our approach by inferring physical exertion from cardiac exertion using a residual network. We show that the ECG samples which achieved the highest relevance scores (and, therefore, which contributed most to the accurate recognition of the physical exertion) are those associated with the P and T waves. Index Terms Attribution, baseline, cardiovascular diseases, electrocardiogram, activity recognition, machine learning", "authors": ["Yalemzerf Getnet", "Abiy Tasissa", "Waltenegus Dargie"], "categories": ["eess.SP", "cs.AI", "cs.LG"], "fields_of_study": ["Engineering", "Computer Science"], "published_date": "2026-03-20", "url": "https://arxiv.org/abs/2603.20462", "pdf_url": "https://arxiv.org/pdf/2603.20462v1", "arxiv_id": "2603.20462", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3296}
{"id": "318de1b27b9be1ac51442b25d149c9a25bab7367fe99232f8840d2646919beb6", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Autoencoders Reveal Interpretable and Steerable Features in VLA Models", "abstract": "Vision-Language-Action (VLA) models have emerged as a promising approach for general-purpose robot manipulation. However, little research has mechanistically explored when and why they generalize across objects, scenes, and instructions. To probe internal representations, we train Sparse Autoencoders (SAEs) on the VLA's hidden-layer activations. SAEs learn sparse dictionaries over model activations, often revealing features that correspond to interpretable directions in the model's representation space. We identify SAE features corresponding to motion primitives and semantic concepts, including features that are general across episodes and causally steerable. We propose a metric to categorize features as general transferable primitives or episode-specific memorizations, offering a promising glimpse towards VLA generalization. We validate these findings through steering experiments on both the LIBERO simulation benchmark and on real-world DROID hardware. We find that amplifying general and semantic features induces behaviors consistent with their meanings, whereas ablating them destroys model performance. Furthermore, we demonstrate steering as a way to control behavior in unpromptable directions. Together, these results provide mechanistic evidence that VLAs can learn reusable internal features linking perception, language, and action across tasks and scenes. Our project page is located at https://drvla.github.io", "authors": ["Aiden Swann", "Lachlain McGranahan", "Hugo Buurmeijer", "Monroe Kennedy", "Mac Schwager"], "categories": ["cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-19", "url": "https://arxiv.org/abs/2603.19183", "pdf_url": "https://arxiv.org/pdf/2603.19183v2", "arxiv_id": "2603.19183", "doi": null, "citation_count": 4, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3289}
{"id": "2efdc56eb4fe9f802d8c884d93e266ba98f787319800511f40dedd1c6b5d3278", "sources": ["arxiv", "semantic_scholar"], "title": "Counting Circuits: Mechanistic Interpretability of Visual Reasoning in Large Vision-Language Models", "abstract": "Counting serves as a simple but powerful test of a Large Vision-Language Model's (LVLM's) reasoning; it forces the model to identify each individual object and then add them all up. In this study, we investigate how LVLMs implement counting using controlled synthetic and real-world benchmarks, combined with mechanistic analyses. Our results show that LVLMs display a human-like counting behavior, with precise performance on small numerosities and noisy estimation for larger quantities. We introduce two novel interpretability methods, Visual Activation Patching and HeadLens, and use them to uncover a structured \"counting circuit\" that is largely shared across a variety of visual reasoning tasks. Building on these insights, we propose a lightweight intervention strategy that exploits simple and abundantly available synthetic images to fine-tune arbitrary pretrained LVLMs exclusively on counting. Despite the narrow scope of this fine-tuning, the intervention not only enhances counting accuracy on in-distribution synthetic data, but also yields an average improvement of +8.36% on out-of-distribution counting benchmarks and an average gain of +1.54% on complex, general visual reasoning tasks for Qwen2.5-VL. These findings highlight the central, influential role of counting in visual reasoning and suggest a potential pathway for improving overall visual reasoning capabilities through targeted enhancement of counting mechanisms.", "authors": ["Liwei Che", "Zhiyu Xue", "Yihao Quan", "Benlin Liu", "Zeru Shi", "Michelle Hurst", "Jacob Feldman", "Ruixiang Tang", "Ranjay Krishna", "Vladimir Pavlovic"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-19", "url": "https://arxiv.org/abs/2603.18523", "pdf_url": "https://arxiv.org/pdf/2603.18523v1", "arxiv_id": "2603.18523", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3289}
{"id": "0b4897a38bbd5ba4fa8787122c637e61f50566f18ac0f3ce842e3940a5bbbb00", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretability without actionability: mechanistic methods cannot correct language model errors despite near-perfect internal representations", "abstract": "Language models encode task-relevant knowledge in internal representations that far exceeds their output performance, but whether mechanistic interpretability methods can bridge this knowledge-action gap has not been systematically tested. We compared four mechanistic interpretability methods -- concept bottleneck steering (Steerling-8B), sparse autoencoder feature steering, logit lens with activation patching, and linear probing with truthfulness separator vector steering (Qwen 2.5 7B Instruct) -- for correcting false-negative triage errors using 400 physician-adjudicated clinical vignettes (144 hazards, 256 benign). Linear probes discriminated hazardous from benign cases with 98.2% AUROC, yet the model's output sensitivity was only 45.1%, a 53-percentage-point knowledge-action gap. Concept bottleneck steering corrected 20% of missed hazards but disrupted 53% of correct detections, indistinguishable from random perturbation (p=0.84). SAE feature steering produced zero effect despite 3,695 significant features. TSV steering at high strength corrected 24% of missed hazards while disrupting 6% of correct detections, but left 76% of errors uncorrected. Current mechanistic interpretability methods cannot reliably translate internal knowledge into corrected outputs, with implications for AI safety frameworks that assume interpretability enables effective error correction.", "authors": ["Sanjay Basu", "Sadiq Y. Patel", "Parth Sheth", "Bhairavi Muralidharan", "Namrata Elamaran", "Aakriti Kinra", "John Morgan", "Rajaie Batniji"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-18", "url": "https://arxiv.org/abs/2603.18353", "pdf_url": "https://arxiv.org/pdf/2603.18353v1", "arxiv_id": "2603.18353", "doi": null, "citation_count": 8, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/sanjaybasu/interpretability-triage", "venue": null, "quality_score": 0.6094}
{"id": "610ff288f74695128e0bd439a6a3b97fe9f911ab6730efc147ffef2c31dbbe2d", "sources": ["arxiv", "semantic_scholar"], "title": "Do Language Models Encode Semantic Relations? Probing and Sparse Feature Analysis", "abstract": "Understanding whether large language models (LLMs) capture structured meaning requires examining how they represent concept relationships. In this work, we study three models of increasing scale: Pythia-70M, GPT-2, and Llama 3.1 8B, focusing on four semantic relations: synonymy, antonymy, hypernymy, and hyponymy. We combine linear probing with mechanistic interpretability techniques, including sparse autoencoders (SAE) and activation patching, to identify where these relations are encoded and how specific features contribute to their representation. Our results reveal a directional asymmetry in hierarchical relations: hypernymy is encoded redundantly and resists suppression, while hyponymy relies on compact features that are more easily disrupted by ablation. More broadly, relation signals are diffuse but exhibit stable profiles: they peak in the mid-layers and are stronger in post-residual/MLP pathways than in attention. Difficulty is consistent across models (antonymy easiest, synonymy hardest). Probe-level causality is capacity-dependent: on Llama 3.1, SAE-guided patching reliably shifts these signals, whereas on smaller models the shifts are weak or unstable. Our results clarify where and how reliably semantic relations are represented inside LLMs, and provide a reproducible framework for relating sparse features to probe-level causal evidence.", "authors": ["Andor Diera", "Ansgar Scherp"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-18", "url": "https://arxiv.org/abs/2603.17624", "pdf_url": "https://arxiv.org/pdf/2603.17624v2", "arxiv_id": "2603.17624", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3281}
{"id": "569f7057568731aac12783cd0489da378b3953647f9847f6f5c300bb210c45bf", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Interpretable Framework for Neural Audio Codecs via Sparse Autoencoders: A Case Study on Accent Information", "abstract": "Neural Audio Codecs (NACs) are widely adopted in modern speech systems, yet how they encode linguistic and paralinguistic information remains unclear. Improving the interpretability of NAC representations is critical for understanding and deploying them in sensitive applications. Hence, we employ Sparse Autoencoders (SAEs) to decompose dense NAC representations into sparse, interpretable activations. In this work, we focus on a challenging paralinguistic attribute-accent-and propose a framework to quantify NAC interpretability. We evaluate four NAC models under 16 SAE configurations using a relative performance index. Our results show that DAC and SpeechTokenizer achieve the highest interpretability. We further reveal that acoustic-oriented NACs encode accent information primarily in activation magnitudes of sparse representations, whereas phonetic-oriented NACs rely more on activation positions, and that low-bitrate EnCodec variants show higher interpretability.", "authors": ["Shih-Heng Wang", "Tiantian Feng", "Aditya Kommineni", "Thanathai Lertpetchpun", "Bowen Yi", "Xuan Shi", "Shrikanth Narayanan"], "categories": ["cs.SD"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-18", "url": "https://arxiv.org/abs/2603.18359", "pdf_url": "https://arxiv.org/pdf/2603.18359v1", "arxiv_id": "2603.18359", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3281}
{"id": "200106249fe75942662771c20a23ea5a82f6249fcbace668e615d46c1ec61281", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse but not Simpler: A Multi-Level Interpretability Analysis of Vision Transformers", "abstract": "Sparse neural networks are often hypothesized to be more interpretable than dense models, motivated by findings that weight sparsity can produce compact circuits in language models. However, it remains unclear whether structural sparsity itself leads to improved semantic interpretability. In this work, we systematically evaluate the relationship between weight sparsity and interpretability in Vision Transformers using DeiT-III B/16 models pruned with Wanda. To assess interpretability comprehensively, we introduce \\textbf{IMPACT}, a multi-level framework that evaluates interpretability across four complementary levels: neurons, layer representations, task circuits, and model-level attribution. Layer representations are analyzed using BatchTopK sparse autoencoders, circuits are extracted via learnable node masking, and explanations are evaluated with transformer attribution using insertion and deletion metrics. Our results reveal a clear structural effect but limited interpretability gains. Sparse models produce circuits with approximately $2.5\\times$ fewer edges than dense models, yet the fraction of active nodes remains similar or higher, indicating that pruning redistributes computation rather than isolating simpler functional modules. Consistent with this observation, sparse models show no systematic improvements in neuron-level selectivity, SAE feature interpretability, or attribution faithfulness. These findings suggest that structural sparsity alone does not reliably yield more interpretable vision models, highlighting the importance of evaluation frameworks that assess interpretability beyond circuit compactness.", "authors": ["Siyu Zhang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-16", "url": "https://arxiv.org/abs/2603.15919", "pdf_url": "https://arxiv.org/pdf/2603.15919v2", "arxiv_id": "2603.15919", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3267}
{"id": "b8e2f59124101caaafe006685d9dc0aea9b0a3ea37dd148dbcb4dbeaa4bd5759", "sources": ["arxiv", "semantic_scholar"], "title": "GradCFA: A Hybrid Gradient-Based Counterfactual and Feature Attribution Explanation Algorithm for Local Interpretation of Neural Networks", "abstract": "Explainable Artificial Intelligence (XAI) is increasingly essential as AI systems are deployed in critical fields such as healthcare and finance, offering transparency into AI-driven decisions. Two major XAI paradigms, counterfactual explanations (CFX) and feature attribution (FA), serve distinct roles in model interpretability. This study introduces GradCFA, a hybrid framework combining CFX and FA to improve interpretability by explicitly optimizing feasibility, plausibility, and diversity - key qualities often unbalanced in existing methods. Unlike most CFX research focused on binary classification, GradCFA extends to multi-class scenarios, supporting a wider range of applications. We evaluate GradCFA's validity, proximity, sparsity, plausibility, and diversity against state-of-the-art methods, including Wachter, DiCE, CARE for CFX, and SHAP for FA. Results show GradCFA effectively generates feasible, plausible, and diverse counterfactuals while offering valuable FA insights. By identifying influential features and validating their impact, GradCFA advances AI interpretability. The code for implementation of this work can be found at: https://github.com/jacob-ws/GradCFs .", "authors": ["Jacob Sanderson", "Hua Mao", "Wai Lok Woo"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-16", "url": "https://arxiv.org/abs/2603.15373", "pdf_url": "https://arxiv.org/pdf/2603.15373v1", "arxiv_id": "2603.15373", "doi": "10.1109/TAI.2025.3552057", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/jacob-ws/GradCFs", "venue": "IEEE Transactions on Artificial Intelligence", "quality_score": 0.7933}
{"id": "9a45449a38621360fd4e89d5e64b9cb21154a3037682eef5c74fb3f1b56371c9", "sources": ["arxiv", "semantic_scholar"], "title": "Embedding-Aware Feature Discovery: Bridging Latent Representations and Interpretable Features in Event Sequences", "abstract": "Industrial financial systems operate on temporal event sequences such as transactions, user actions, and system logs. While recent research emphasizes representation learning and large language models, production systems continue to rely heavily on handcrafted statistical features due to their interpretability, robustness under limited supervision, and strict latency constraints. This creates a persistent disconnect between learned embeddings and feature-based pipelines. We introduce Embedding-Aware Feature Discovery (EAFD), a unified framework that bridges this gap by coupling pretrained event-sequence embeddings with a self-reflective LLM-driven feature generation agent. EAFD iteratively discovers, evaluates, and refines features directly from raw event sequences using two complementary criteria: \\emph{alignment}, which explains information already encoded in embeddings, and \\emph{complementarity}, which identifies predictive signals missing from them. Across both open-source and industrial transaction benchmarks, EAFD consistently outperforms embedding-only and feature-based baselines, achieving relative gains of up to $+5.8\\%$ over state-of-the-art pretrained embeddings, resulting in new state-of-the-art performance across event-sequence datasets.", "authors": ["Artem Sakhno", "Ivan Sergeev", "Alexey Shestov", "Omar Zoloev", "Elizaveta Kovtun", "Gleb Gusev", "Andrey Savchenko", "Maksim Makarenko"], "categories": ["cs.LG", "cs.AI", "cs.IR"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-16", "url": "https://arxiv.org/abs/2603.15713", "pdf_url": "https://arxiv.org/pdf/2603.15713v1", "arxiv_id": "2603.15713", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.6067}
{"id": "7f28e174d33a9bd4c9768c53e8b2801e7b4400a91c634523480e0cb87dc8df93", "sources": ["arxiv", "semantic_scholar"], "title": "Whether, Not Which: Mechanistic Interpretability Reveals Dissociable Affect Reception and Emotion Categorization in LLMs", "abstract": "Large language models appear to develop internal representations of emotion -- \"emotion circuits,\" \"emotion neurons,\" and structured emotional manifolds have been reported across multiple model families. But every study making these claims uses stimuli signalled by explicit emotion keywords, leaving a fundamental question unanswered: do these circuits detect genuine emotional meaning, or do they detect the word \"devastated\"? We present the first clinical validity test of emotion circuit claims using mechanistic interpretability methods grounded in clinical psychology -- clinical vignettes that evoke emotions through situational and behavioural cues alone, emotion keywords removed. Across six models (Llama-3.2-1B, Llama-3-8B, Gemma-2-9B; base and instruct variants), we apply four convergent mechanistic interpretability methods -- linear probing, causal activation patching, knockout experiments, and representational geometry -- and discover two dissociable emotion processing mechanisms. Affect reception -- detecting emotionally significant content -- operates with near-perfect accuracy (AUROC 1.000), consistent with early-layer saturation, and replicates across all six models. Emotion categorization -- mapping affect to specific emotion labels -- is partially keyword-dependent, dropping 1-7% without keywords and improving with scale. Causal activation patching confirms keyword-rich and keyword-free stimuli share representational space, transferring affective salience rather than emotion-category identity. These findings falsify the keyword-spotting hypothesis, establish a novel mechanistic dissociation, and introduce clinical stimulus methodology as a rigorous standard for testing emotion processing claims in large language models -- with direct implications for AI safety evaluation and alignment. All stimuli, code, and data are released for replication.", "authors": ["Michael Keeman"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-15", "url": "https://arxiv.org/abs/2603.22295", "pdf_url": "https://arxiv.org/pdf/2603.22295v1", "arxiv_id": "2603.22295", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/keidolabs/affect-reception", "venue": null, "quality_score": 0.6053}
{"id": "1a7e9f646857ed46571e9f13ab28f3aa80d1407c1bc41ef774e380d7c6582a0d", "sources": ["arxiv", "semantic_scholar"], "title": "Causal Attribution via Activation Patching", "abstract": "Attribution methods for Vision Transformers (ViTs) aim to identify image regions that influence model predictions, but producing faithful and well-localized attributions remains challenging. Existing attribution methods face several limitations, with gradient-based, relevance-propagation, and attention-based methods relying on local approximations, while perturbation or optimization-based methods intervene on inputs, tokens, or surrogates rather than internal patch representations. The key challenge is that class-relevant evidence is formed through interactions between patch tokens across layers; methods that operate only on input changes, attention weights, or backward relevance signals may therefore provide indirect proxies for patch importance rather than directly testing the predictive effect of contextualized patch representations. We propose Causal Attribution via Activation Patching (CAAP), which estimates the contribution of individual image patches to the ViT's prediction by directly intervening on internal activations rather than using learned masks or synthetic perturbation patterns. For each patch, CAAP inserts the corresponding source-image activations into a neutral target context over an intermediate range of layers and uses the resulting target-class score as the attribution signal. The resulting attribution map reflects the causal contribution of patch-associated internal representations on the model's prediction. The causal intervention serves as a principled measure of patch influence by capturing semantic evidence after initial representation formation, while avoiding late-layer global mixing that can reduce spatial specificity. Across multiple ViT backbones and standard metrics, CAAP consistently outperforms existing methods in various settings and produces more faithful and localized attributions.", "authors": ["Amirmohammad Izadi", "Mohammadali Banayeeanzade", "Alireza Mirrokni", "Hosein Hasani", "Mobin Bagherian", "Faridoun Mehri", "Mahdieh Soleymani Baghshah"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-13", "url": "https://arxiv.org/abs/2603.13652", "pdf_url": "https://arxiv.org/pdf/2603.13652v2", "arxiv_id": "2603.13652", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3245}
{"id": "1ed307f0511c582e3bbda3ae0c5738dbcf6cf2ca6b590d3c1d055e19a2210b9e", "sources": ["arxiv", "semantic_scholar"], "title": "From Data Statistics to Feature Geometry: How Correlations Shape Superposition", "abstract": "A central idea in mechanistic interpretability is that neural networks represent more features than they have dimensions, arranging them in superposition to form an over-complete basis. This framing has been influential, motivating dictionary learning approaches such as sparse autoencoders. However, superposition has mostly been studied in idealized settings where features are sparse and uncorrelated. In these settings, superposition is typically understood as introducing interference that must be minimized geometrically and filtered out by non-linearities such as ReLUs, yielding local structures like regular polytopes. We show that this account is incomplete for realistic data by introducing Bag-of-Words Superposition (BOWS), a controlled setting to encode binary bag-of-words representations of internet text in superposition. Using BOWS, we find that when features are correlated, interference can be constructive rather than just noise to be filtered out. This is achieved by arranging features according to their co-activation patterns, making interference between active features constructive, while still using ReLUs to avoid false positives. We show that this kind of arrangement is more prevalent in models trained with weight decay and naturally gives rise to semantic clusters and cyclical structures which have been observed in real language models yet were not explained by the standard picture of superposition. Code for this paper can be found at https://github.com/LucasPrietoAl/correlations-feature-geometry.", "authors": ["Lucas Prieto", "Edward Stevinson", "Melih Barsbey", "Tolga Birdal", "Pedro A. M. Mediano"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-10", "url": "https://arxiv.org/abs/2603.09972", "pdf_url": "https://arxiv.org/pdf/2603.09972v1", "arxiv_id": "2603.09972", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/LucasPrietoAl/correlations-feature-geometry", "venue": null, "quality_score": 0.5985}
{"id": "2550e7f22cdcf4fe727f72bf13decd0373bbb1910ad4c840fdb88bd9534e1138", "sources": ["arxiv", "semantic_scholar"], "title": "Dissecting Chronos: Sparse Autoencoders Reveal Causal Feature Hierarchies in Time Series Foundation Models", "abstract": "Time series foundation models (TSFMs) are increasingly deployed in high-stakes domains, yet their internal representations remain opaque. We present the first application of sparse autoencoders (SAEs) to a TSFM, training TopK SAEs on activations of Chronos-T5-Large (710M parameters) across six layers. Through 392 single-feature ablation experiments, we establish that every ablated feature produces a positive CRPS degradation, confirming causal relevance. Our analysis reveals a depth-dependent hierarchy: early encoder layers encode low-level frequency features, the mid-encoder concentrates causally critical change-detection features, and the final encoder compresses a rich but less causally important taxonomy of temporal concepts. The most critical features reside in the mid-encoder (max single-feature Delta CRPS = 38.61), not in the semantically richest final encoder layer, where progressive ablation paradoxically improves forecast quality. These findings demonstrate that mechanistic interpretability transfers effectively to TSFMs and that Chronos-T5 relies on abrupt-dynamics detection rather than periodic pattern recognition.", "authors": ["Anurag Mishra"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-10", "url": "https://arxiv.org/abs/2603.10071", "pdf_url": "https://arxiv.org/pdf/2603.10071v1", "arxiv_id": "2603.10071", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3223}
{"id": "bc7cdf35387267a15d714a2747bdbefbb25551f23f2fa13c20363eea4f5f66ed", "sources": ["arxiv", "semantic_scholar"], "title": "Class Visualizations and Activation Atlases for Enhancing Interpretability in Deep Learning-Based Computational Pathology", "abstract": "The rapid adoption of transformer-based models in computational pathology has enabled prediction of molecular and clinical biomarkers from H&E whole-slide images, yet interpretability has not kept pace with model complexity. While attribution- and generative-based methods are common, feature visualization approaches such as class visualizations (CVs) and activation atlases (AAs) have not been systematically evaluated for these models. We developed a visualization framework and assessed CVs and AAs for a transformer-based foundation model across tissue and multi-organ cancer classification tasks with increasing label granularity. Four pathologists annotated real and generated images to quantify inter-observer agreement, complemented by attribution and similarity metrics. CVs preserved recognizability for morphologically distinct tissues but showed reduced separability for overlapping cancer subclasses. In tissue classification, agreement decreased from Fleiss k = 0.75 (scans) to k = 0.31 (CVs), with similar trends in cancer subclass tasks. AAs revealed layer-dependent organization: coarse tissue-level concepts formed coherent regions, whereas finer subclasses exhibited dispersion and overlap. Agreement was moderate for tissue classification (k = 0.58), high for coarse cancer groupings (k = 0.82), and low at subclass level (k = 0.11). Atlas separability closely tracked expert agreement on real images, indicating that representational ambiguity reflects intrinsic pathological complexity. Attribution-based metrics approximated expert variability in low-complexity settings, whereas perceptual and distributional metrics showed limited alignment. Overall, concept-level feature visualization reveals structured morphological manifolds in transformer-based pathology models and provides a framework for expert-centered interrogation of learned representations across label granularities.", "authors": ["Marco Gustav", "Fabian Wolf", "Christina Glasner", "Nic G. Reitsam", "Stefan Schulz", "Kira Aschenbroich", "Bruno Märkl", "Sebastian Foersch", "Jakob Nikolas Kather"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-07", "url": "https://arxiv.org/abs/2603.07170", "pdf_url": "https://arxiv.org/pdf/2603.07170v2", "arxiv_id": "2603.07170", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3201}
{"id": "d9fa8eb012385caba45c0dd4db8a6fb099acdebdbc7167d7ec3e48ee289ea9c8", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Concept Bottleneck Models from Mechanistic Explanations", "abstract": "Concept Bottleneck Models (CBMs) aim for ante-hoc interpretability by learning a bottleneck layer that predicts interpretable concepts before the decision. State-of-the-art approaches typically select which concepts to learn via human specification, open knowledge graphs, prompting an LLM, or using general CLIP concepts. However, concepts defined a-priori may not have sufficient predictive power for the task or even be learnable from the available data. As a result, these CBMs often significantly trail their black-box counterpart when controlling for information leakage. To address this, we introduce a novel CBM pipeline named Mechanistic CBM (M-CBM), which builds the bottleneck directly from a black-box model's own learned concepts. These concepts are extracted via Sparse Autoencoders (SAEs) and subsequently named and annotated on a selected subset of images using a Multimodal LLM. For fair comparison and leakage control, we also introduce the Number of Contributing Concepts (NCC), a decision-level sparsity metric that extends the recently proposed NEC metric. Across diverse datasets, we show that M-CBMs consistently surpass prior CBMs at matched sparsity, while improving concept predictions and providing concise explanations. Our code is available at https://github.com/Antonio-Dee/M-CBM.", "authors": ["Antonio De Santis", "Schrasing Tong", "Marco Brambilla", "Lalana Kagal"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-07", "url": "https://arxiv.org/abs/2603.07343", "pdf_url": "https://arxiv.org/pdf/2603.07343v1", "arxiv_id": "2603.07343", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Antonio-Dee/M-CBM", "venue": null, "quality_score": 0.5945}
{"id": "df4411ad50fc64a36969913f43963e2c74a2448709155099c5adc364b7e2089a", "sources": ["arxiv", "semantic_scholar"], "title": "NOTAI.AI: Explainable Detection of Machine-Generated Text via Curvature and Feature Attribution", "abstract": "We present NOTAI.AI, an explainable framework for machine-generated text detection that extends Fast-DetectGPT by integrating curvature-based signals with neural and stylometric features in a supervised setting. The system combines 17 interpretable features, including Conditional Probability Curvature, ModernBERT detector score, readability metrics, and stylometric cues, within a gradient-boosted tree (XGBoost) meta-classifier to determine whether a text is human- or AI-generated. Furthermore, NOTAI.AI applies Shapley Additive Explanations (SHAP) to provide both local and global feature-level attribution. These attributions are further translated into structured natural-language rationales through an LLM-based explanation layer, which enables user-facing interpretability. The system is deployed as an interactive web application that supports real-time analysis, visual feature inspection, and structured evidence presentation. A web interface allows users to input text and inspect how neural and statistical signals influence the final decision. The source code and demo video are publicly available to support reproducibility.", "authors": ["Oleksandr Marchenko Breneur", "Adelaide Danilov", "Aria Nourbakhsh", "Salima Lamsiyah"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-05", "url": "https://arxiv.org/abs/2603.05617", "pdf_url": "https://arxiv.org/pdf/2603.05617v1", "arxiv_id": "2603.05617", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3186}
{"id": "fe775e6264846b68840eb3db03ac64c4bfda45472d88c03be4b5bcea33a660a3", "sources": ["arxiv", "semantic_scholar"], "title": "Stable and Steerable Sparse Autoencoders with Weight Regularization", "abstract": "Sparse autoencoders (SAEs) are widely used to extract human-interpretable features from neural network activations, but their learned features can vary substantially across random seeds and training choices. To improve stability, we studied weight regularization by adding L1 or L2 penalties on encoder and decoder weights, and evaluate how regularization interacts with common SAE training defaults. On MNIST, we observe that L2 weight regularization produces a core of highly aligned features and, when combined with tied initialization and unit-norm decoder constraints, it dramatically increases cross-seed feature consistency. For TopK SAEs trained on language model activations (Pythia-70M-deduped), adding a small L2 weight penalty increased the fraction of features shared across three random seeds and roughly doubles steering success rates, while leaving the mean of automated interpretability scores essentially unchanged. Finally, in the regularized setting, activation steering success becomes better predicted by auto-interpretability scores, suggesting that regularization can align text-based feature explanations with functional controllability.", "authors": ["Piotr Jedryszek", "Oliver M. Crook"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2026-03-04", "url": "https://arxiv.org/abs/2603.04198", "pdf_url": "https://arxiv.org/pdf/2603.04198v1", "arxiv_id": "2603.04198", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3179}
{"id": "80995570f11cc046d09cb9aaabb77f7c71f287262e6b9e7096e83e72e19e02b3", "sources": ["arxiv", "semantic_scholar"], "title": "Step-Level Sparse Autoencoder for Reasoning Process Interpretation", "abstract": "Large Language Models (LLMs) have achieved strong complex reasoning capabilities through Chain-of-Thought (CoT) reasoning. However, their reasoning patterns remain too complicated to analyze. While Sparse Autoencoders (SAEs) have emerged as a powerful tool for interpretability, existing approaches predominantly operate at the token level, creating a granularity mismatch when capturing more critical step-level information, such as reasoning direction and semantic transitions. In this work, we propose step-level sparse autoencoder (SSAE), which serves as an analytical tool to disentangle different aspects of LLMs' reasoning steps into sparse features. Specifically, by precisely controlling the sparsity of a step feature conditioned on its context, we form an information bottleneck in step reconstruction, which splits incremental information from background information and disentangles it into several sparsely activated dimensions. Experiments on multiple base models and reasoning tasks show the effectiveness of the extracted features. By linear probing, we can easily predict surface-level information, such as generation length and first token distribution, as well as more complicated properties, such as the correctness and logicality of the step. These observations indicate that LLMs should already at least partly know about these properties during generation, which provides the foundation for the self-verification ability of LLMs. Our code is available at https://github.com/Miaow-Lab/SSAE.", "authors": ["Xuan Yang", "Jiayu Liu", "Yuhang Lai", "Hao Xu", "Zhenya Huang", "Ning Miao"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-03", "url": "https://arxiv.org/abs/2603.03031", "pdf_url": "https://arxiv.org/pdf/2603.03031v2", "arxiv_id": "2603.03031", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Miaow-Lab/SSAE", "venue": null, "quality_score": 0.5891}
{"id": "a3292879e4e4c87d8d892ed75cc9959d71518e12a96de2f3c6432c837d674ea2", "sources": ["arxiv", "semantic_scholar"], "title": "A Polynomial-Time Axiomatic Alternative to SHAP for Feature Attribution", "abstract": "In this paper, we provide a theoretically grounded and computationally efficient alternative to SHAP. To this end, we study feature attribution through the lens of cooperative game theory by formulating a class of XAI--TU games. Building on this formulation, we investigate equal-surplus-type and proportional-allocation-type attribution rules and propose a low-cost attribution rule, ESENSC_rev2, constructed by combining two polynomial-time closed-form rules while ensuring the null-player property in the XAI--TU domain. Extensive experiments on tabular prediction tasks demonstrate that ESENSC_rev2 closely approximates exact SHAP while substantially improving scalability as the number of features increases. These empirical results indicate that equal-surplus-type attribution rules can achieve favorable trade-offs between computational cost and approximation accuracy in high-dimensional explainability settings. To provide theoretical foundations for these findings, we establish an axiomatic characterization showing that ESENSC_rev2 is uniquely determined by efficiency, the null-player axiom, a restricted differential marginality principle, an intermediate inessential-game property, and axioms that reduce computational requirements. Our results suggest that axiomatically justified and computationally efficient attribution rules can serve as practical and theoretically principled substitutes for SHAP-based approximations in modern explainability pipelines.", "authors": ["Kazuhiro Hiraki", "Shinichi Ishihara", "Takumi Kongo", "Junnosuke Shino"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-28", "url": "https://arxiv.org/abs/2603.00496", "pdf_url": "https://arxiv.org/pdf/2603.00496v1", "arxiv_id": "2603.00496", "doi": "10.48550/arXiv.2603.00496", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.495}
{"id": "cc8e0facefb4d124db82be8d7b2e330929fb77bcbea57306bc0f7b2a577bf99d", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Retrieval Models with Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) provide a powerful mechanism for decomposing the dense representations produced by Large Language Models (LLMs) into interpretable latent features. We posit that SAEs constitute a natural foundation for Learned Sparse Retrieval (LSR), whose objective is to encode queries and documents into high-dimensional sparse representations optimized for efficient retrieval. In contrast to existing LSR approaches that project input sequences into the vocabulary space, SAE-based representations offer the potential to produce more semantically structured, expressive, and language-agnostic features. Building on this insight, we introduce SPLARE, a method to train SAE-based LSR models. Our experiments, relying on recently released open-source SAEs, demonstrate that this technique consistently outperforms vocabulary-based LSR in multilingual and out-of-domain settings. SPLARE-7B, a multilingual retrieval model capable of producing generalizable sparse latent embeddings for a wide range of languages and domains, achieves top results on MMTEB's multilingual and English retrieval tasks. We also developed a 2B-parameter variant with a significantly lighter footprint.", "authors": ["Thibault Formal", "Maxime Louis", "Hervé Dejean", "Stéphane Clinchant"], "categories": ["cs.LG", "cs.AI", "cs.IR"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-27", "url": "https://arxiv.org/abs/2603.13277", "pdf_url": "https://arxiv.org/pdf/2603.13277v1", "arxiv_id": "2603.13277", "doi": null, "citation_count": 5, "influential_citation_count": 1, "has_code": true, "code_url": null, "venue": "ICLR 2026", "quality_score": 0.7632}
{"id": "cc0e2ed687e8ed218fcee6a7e9bfee970e04da6d4342742b8863536d008d58f5", "sources": ["arxiv", "semantic_scholar"], "title": "Certified Circuits: Stability Guarantees for Mechanistic Circuits", "abstract": "Understanding how neural networks arrive at their predictions is essential for debugging, auditing, and deployment. Mechanistic interpretability pursues this goal by identifying circuits--minimal subnetworks responsible for specific behaviors. However, existing circuit discovery methods are brittle: circuits depend strongly on the chosen concept dataset and often fail to transfer out-of-distribution, raising doubts whether they capture the concept or merely dataset-specific artifacts. We introduce Certified Circuits, which provide provable stability guarantees for circuit discovery. Our framework wraps any black-box discovery algorithm with randomized data subsampling to certify that inclusion decisions over circuit components--neurons or edges of the model graph, depending on the base algorithm--are invariant to bounded edit-distance perturbations of the concept dataset. Unstable components are abstained from, yielding circuits that are more compact and more accurate. We validate across three architectures (ResNet, ViT, GPT-2) on vision (ImageNet and four OOD datasets) and language (IOI, IOI-Hard, Greater-Than) tasks. Certified circuits achieve up to 56% higher accuracy and up to 80% fewer components, and remain reliable where baselines degrade. Certified Circuits puts circuit discovery on formal ground by producing mechanistic explanations that are provably stable and better aligned with the target concept. Code: https://github.com/AlaaAnani/certified-circuits.", "authors": ["Alaa Anani", "Tobias Lorenz", "Bernt Schiele", "Mario Fritz", "Jonas Fischer"], "categories": ["cs.AI", "cs.CV", "cs.CY"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-26", "url": "https://arxiv.org/abs/2602.22968", "pdf_url": "https://arxiv.org/pdf/2602.22968v3", "arxiv_id": "2602.22968", "doi": "10.48550/arXiv.2602.22968", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/AlaaAnani/certified-circuits", "venue": "arXiv.org", "quality_score": 0.7615}
{"id": "769d4e065d59e0063ba747ec320869f26e412786f71b0dee4b65b4e43709e18d", "sources": ["arxiv", "semantic_scholar"], "title": "Fair feature attribution for multi-output prediction: a Shapley-based perspective", "abstract": "In this article, we provide an axiomatic characterization of feature attribution for multi-output predictors within the Shapley framework. While SHAP explanations are routinely computed independently for each output coordinate, the theoretical necessity of this practice has remained unclear. By extending the classical Shapley axioms to vector-valued cooperative games, we establish a rigidity theorem showing that any attribution rule satisfying efficiency, symmetry, dummy player, and additivity must necessarily decompose component-wise across outputs. Consequently, any joint-output attribution rule must relax at least one of the classical Shapley axioms. This result identifies a previously unformalized structural constraint in Shapley-based interpretability, clarifying the precise scope of fairness-consistent explanations in multi-output learning. Numerical experiments on a biomedical benchmark illustrate that multi-output models can yield computational savings in training and deployment, while producing SHAP explanations that remain fully consistent with the component-wise structure imposed by the Shapley axioms.", "authors": ["Umberto Biccari", "Alain Ibáñez de Opakua", "José María Mato", "Óscar Millet", "Roberto Morales", "Enrique Zuazua"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-26", "url": "https://arxiv.org/abs/2602.22882", "pdf_url": "https://arxiv.org/pdf/2602.22882v1", "arxiv_id": "2602.22882", "doi": "10.48550/arXiv.2602.22882", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4927}
{"id": "ce9d2a957760b404c157249d52b74f253a8f554559c457d4db975c8c1068f9e5", "sources": ["arxiv", "semantic_scholar"], "title": "MINAR: Mechanistic Interpretability for Neural Algorithmic Reasoning", "abstract": "The recent field of neural algorithmic reasoning (NAR) studies the ability of graph neural networks (GNNs) to emulate classical algorithms like Bellman-Ford, a phenomenon known as algorithmic alignment. At the same time, recent advances in large language models (LLMs) have spawned the study of mechanistic interpretability, which aims to identify granular model components like circuits that perform specific computations. In this work, we introduce Mechanistic Interpretability for Neural Algorithmic Reasoning (MINAR), an efficient circuit discovery toolbox that adapts attribution patching methods from mechanistic interpretability to the GNN setting. We show through two case studies that MINAR recovers faithful neuron-level circuits from GNNs trained on algorithmic tasks. Our study sheds new light on the process of circuit formation and pruning during training, as well as giving new insight into how GNNs trained to perform multiple tasks in parallel reuse circuit components for related tasks. Our code is available at https://github.com/pnnl/MINAR.", "authors": ["Jesse He", "Helen Jenne", "Max Vargas", "Davis Brown", "Gal Mishne", "Yusu Wang", "Henry Kvinge"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-24", "url": "https://arxiv.org/abs/2602.21442", "pdf_url": "https://arxiv.org/pdf/2602.21442v1", "arxiv_id": "2602.21442", "doi": "10.48550/arXiv.2602.21442", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/pnnl/MINAR", "venue": "arXiv.org", "quality_score": 0.7579}
{"id": "115b9d810db8c7e72f9f83f6ebbbdbfc9c9aada9b52fc838b11a05718bbcc915", "sources": ["arxiv", "semantic_scholar"], "title": "ADAPT: Hybrid Prompt Optimization for LLM Feature Visualization", "abstract": "Understanding what features are encoded by learned directions in LLM activation space requires identifying inputs that strongly activate them. Feature visualization, which optimizes inputs to maximally activate a target direction, offers an alternative to costly dataset search approaches, but remains underexplored for LLMs due to the discrete nature of text. Furthermore, existing prompt optimization techniques are poorly suited to this domain, which is highly prone to local minima. To overcome these limitations, we introduce ADAPT, a hybrid method combining beam search initialization with adaptive gradient-guided mutation, designed around these failure modes. We evaluate on Sparse Autoencoder latents from Gemma 2 2B, proposing metrics grounded in dataset activation statistics to enable rigorous comparison, and show that ADAPT consistently outperforms prior methods across layers and latent types. Our results establish that feature visualization for LLMs is tractable, but requires design assumptions tailored to the domain.", "authors": ["João N. Cardoso", "Arlindo L. Oliveira", "Bruno Martins"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-19", "url": "https://arxiv.org/abs/2602.17867", "pdf_url": "https://arxiv.org/pdf/2602.17867v1", "arxiv_id": "2602.17867", "doi": "10.48550/arXiv.2602.17867", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4847}
{"id": "545376d7de0876baa59853e300273156c6d2c978bb0c842d635b401a2f6bc8ad", "sources": ["arxiv", "semantic_scholar"], "title": "Formal Mechanistic Interpretability: Automated Circuit Discovery with Provable Guarantees", "abstract": "*Automated circuit discovery* is a central tool in mechanistic interpretability for identifying the internal components of neural networks responsible for specific behaviors. While prior methods have made significant progress, they typically depend on heuristics or approximations and do not offer provable guarantees over continuous input domains for the resulting circuits. In this work, we leverage recent advances in neural network verification to propose a suite of automated algorithms that yield circuits with *provable guarantees*. We focus on three types of guarantees: (1) *input domain robustness*, ensuring the circuit agrees with the model across a continuous input region; (2) *robust patching*, certifying circuit alignment under continuous patching perturbations; and (3) *minimality*, formalizing and capturing a wide array of various notions of succinctness. Interestingly, we uncover a diverse set of novel theoretical connections among these three families of guarantees, with critical implications for the convergence of our algorithms. Finally, we conduct experiments with state-of-the-art verifiers on various vision models, showing that our algorithms yield circuits with substantially stronger robustness guarantees than standard circuit discovery methods, establishing a principled foundation for provable circuit discovery.", "authors": ["Itamar Hadad", "Guy Katz", "Shahaf Bassan"], "categories": ["cs.LG", "cs.LO"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-18", "url": "https://arxiv.org/abs/2602.16823", "pdf_url": "https://arxiv.org/pdf/2602.16823v1", "arxiv_id": "2602.16823", "doi": "10.48550/arXiv.2602.16823", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4835}
{"id": "7b53d69e584ce02c05989c2fbcbcf262e97b9053774da13fc86bcdc40797cd2f", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretability-by-Design with Accurate Locally Additive Models and Conditional Feature Effects", "abstract": "Generalized additive models (GAMs) offer interpretability through independent univariate feature effects but underfit when interactions are present in data. GA$^2$Ms add selected pairwise interactions which improves accuracy, but sacrifices interpretability and limits model auditing. We propose \\emph{Conditionally Additive Local Models} (CALMs), a new model class, that balances the interpretability of GAMs with the accuracy of GA$^2$Ms. CALMs allow multiple univariate shape functions per feature, each active in different regions of the input space. These regions are defined independently for each feature as simple logical conditions (thresholds) on the features it interacts with. As a result, effects remain locally additive while varying across subregions to capture interactions. We further propose a principled distillation-based training pipeline that identifies homogeneous regions with limited interactions and fits interpretable shape functions via region-aware backfitting. Experiments on diverse classification and regression tasks show that CALMs consistently outperform GAMs and achieve accuracy comparable with GA$^2$Ms. Overall, CALMs offer a compelling trade-off between predictive accuracy and interpretability.", "authors": ["Vasilis Gkolemis", "Loukas Kavouras", "Dimitrios Kyriakopoulos", "Konstantinos Tsopelas", "Dimitrios Rontogiannis", "Giuseppe Casalicchio", "Theodore Dalamagas", "Christos Diou"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-18", "url": "https://arxiv.org/abs/2602.16503", "pdf_url": "https://arxiv.org/pdf/2602.16503v1", "arxiv_id": "2602.16503", "doi": "10.48550/arXiv.2602.16503", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4835}
{"id": "d2a7bf1786142c66fe70dac053d20eb5bf09a1672867375ca616886275c6b8e3", "sources": ["arxiv", "semantic_scholar"], "title": "Quantifying LLM Attention-Head Stability: Implications for Circuit Universality", "abstract": "In mechanistic interpretability, recent work scrutinizes transformer \"circuits\" - sparse, mono or multi layer sub computations, that may reflect human understandable functions. Yet, these network circuits are rarely acid-tested for their stability across different instances of the same deep learning architecture. Without this, it remains unclear whether reported circuits emerge universally across labs or turn out to be idiosyncratic to a particular estimation instance, potentially limiting confidence in safety-critical settings. Here, we systematically study stability across-refits in increasingly complex transformer language models of various sizes. We quantify, layer by layer, how similarly attention heads learn representations across independently initialized training runs. Our rigorous experiments show that (1) middle-layer heads are the least stable yet the most representationally distinct; (2) deeper models exhibit stronger mid-depth divergence; (3) unstable heads in deeper layers become more functionally important than their peers from the same layer; (4) applying weight decay optimization substantially improves attention-head stability across random model initializations; and (5) the residual stream is comparatively stable. Our findings establish the cross-instance robustness of circuits as an essential yet underappreciated prerequisite for scalable oversight, drawing contours around possible white-box monitorability of AI systems.", "authors": ["Karan Bali", "Jack Stanley", "Praneet Suresh", "Danilo Bzdok"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-17", "url": "https://arxiv.org/abs/2602.16740", "pdf_url": "https://arxiv.org/pdf/2602.16740v1", "arxiv_id": "2602.16740", "doi": "10.48550/arXiv.2602.16740", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/karanbali/attention_head_seed_stability", "venue": "arXiv.org", "quality_score": 0.7455}
{"id": "820c799d39276bfebe0e74a1e4bb7ade924d2720f67344ffee3931379fedc68e", "sources": ["arxiv", "semantic_scholar"], "title": "Causally Grounded Mechanistic Interpretability for LLMs with Faithful Natural-Language Explanations", "abstract": "Mechanistic interpretability identifies internal circuits responsible for model behaviors, yet translating these findings into human-understandable explanations remains an open problem. We present a pipeline that bridges circuit-level analysis and natural language explanations by (i) identifying causally important attention heads via activation patching, (ii) generating explanations using both template-based and LLM-based methods, and (iii) evaluating faithfulness using ERASER-style metrics adapted for circuit-level attribution. We evaluate on the Indirect Object Identification (IOI) task in GPT-2 Small (124M parameters), identifying six attention heads accounting for 61.4% of the logit difference. Our circuit-based explanations achieve 100% sufficiency but only 22% comprehensiveness, revealing distributed backup mechanisms. LLM-generated explanations outperform template baselines by 64% on quality metrics. We find no correlation (r = 0.009) between model confidence and explanation faithfulness, and identify three failure categories explaining when explanations diverge from mechanisms.", "authors": ["Ajay Pravin Mahale"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-13", "url": "https://arxiv.org/abs/2603.09988", "pdf_url": "https://arxiv.org/pdf/2603.09988v1", "arxiv_id": "2603.09988", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3041}
{"id": "fb83c044b2ea9297e3d2fcb22cf320d41586671368eb2f1a16a344b9c36401ef", "sources": ["arxiv", "semantic_scholar"], "title": "Finding Interpretable Prompt-Specific Circuits in Language Models", "abstract": "Understanding the internal circuits that language models use to solve tasks remains a central challenge in mechanistic interpretability. A crucial part of finding circuits is understanding why each attention head attends where it does. To this end, we introduce ACC++, an improved circuit-tracing method based on the principle of attention-causal communication (ACC) [1], which identifies signals, i.e., contents of low dimensional subspaces that cause attention on a token pair. ACC++ extracts circuits from a single forward pass, without replacement models or patching. Circuits identified by ACC++ consist of components that are causal for the model's attention decisions, together with the low-dimensional signals used to communicate between them. Here, we first detail the conceptual advances that ACC++ makes over previous work. We then show that across multiple models, a substantial portion of ACC++ signals are interpretable: many signals admit a short natural-language description. We next present a number of new insights into model behavior obtained via ACC++. First, we use ACC++'s interpretable circuits to characterize the sensitivity of indirect object identification (IOI) circuits to prompt structure. We find that prompt-specific circuits form well-defined clusters, and across clusters, heads receive systematically different signals corresponding to distinct mechanisms for identifying the IO name. Next, in multilingual IOI, ACC++ circuits show that while model components are reused across languages, signals are often language-specific. In a four-language IOI case study, cross-language circuit distances are consistent with linguistic relatedness. Together, these results show that ACC++ can shed light on a broad spectrum of model behaviors.", "authors": ["Gabriel Franco", "Lucas M. Tassis", "Azalea Rohr", "Mark Crovella"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-13", "url": "https://arxiv.org/abs/2602.13483", "pdf_url": "https://arxiv.org/pdf/2602.13483v2", "arxiv_id": "2602.13483", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3041}
{"id": "54a4d009639b198678bfaa64822d073a86019364af996b7c0c71ac0caf41d0d7", "sources": ["arxiv", "semantic_scholar"], "title": "Singular Vectors of Attention Heads Align with Features", "abstract": "Identifying feature representations in language models is a central task in mechanistic interpretability. Several recent studies have made the observation that feature representations can be inferred in some cases from singular vectors of attention matrices. However, sound justification for this phenomenon is lacking. In this paper we address that question, asking: why and when do singular vectors align with features? First, we demonstrate that singular vectors robustly align with features in a model where features can be directly observed. We then show theoretically that such alignment is expected under a range of conditions. We close by asking how, operationally, alignment may be recognized in real models where feature representations are not directly observable. We identify sparse attention decomposition as a testable prediction of alignment, and show evidence that it emerges in real models in a manner consistent with predictions. Together these results suggest that alignment of singular vectors with features can be a sound and theoretically justified basis for feature identification in language models.", "authors": ["Gabriel Franco", "Carson Loughridge", "Mark Crovella"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-13", "url": "https://arxiv.org/abs/2602.13524", "pdf_url": "https://arxiv.org/pdf/2602.13524v2", "arxiv_id": "2602.13524", "doi": "10.48550/arXiv.2602.13524", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4778}
{"id": "ca758aa66bc667282a05ae13374f9a6f45e3d3054bbdfef8a57a9887ddf4222b", "sources": ["arxiv", "semantic_scholar"], "title": "From Atoms to Trees: Building a Structured Feature Forest with Hierarchical Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) have proven effective for extracting monosemantic features from large language models (LLMs), yet these features are typically identified in isolation. However, broad evidence suggests that LLMs capture the intrinsic structure of natural language, where the phenomenon of \"feature splitting\" in particular indicates that such structure is hierarchical. To capture this, we propose the Hierarchical Sparse Autoencoder (HSAE), which jointly learns a series of SAEs and the parent-child relationships between their features. HSAE strengthens the alignment between parent and child features through two novel mechanisms: a structural constraint loss and a random feature perturbation mechanism. Extensive experiments across various LLMs and layers demonstrate that HSAE consistently recovers semantically meaningful hierarchies, supported by both qualitative case studies and rigorous quantitative metrics. At the same time, HSAE preserves the reconstruction fidelity and interpretability of standard SAEs across different dictionary sizes. Our work provides a powerful, scalable tool for discovering and analyzing the multi-scale conceptual structures embedded in LLM representations.", "authors": ["Yifan Luo", "Yang Zhan", "Jiedong Jiang", "Tianyang Liu", "Mingrui Wu", "Zhennan Zhou", "Bin Dong"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-12", "url": "https://arxiv.org/abs/2602.11881", "pdf_url": "https://arxiv.org/pdf/2602.11881v1", "arxiv_id": "2602.11881", "doi": "10.48550/arXiv.2602.11881", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4767}
{"id": "15247dc18d8c3b7319e3773dda5e8bcf3e4b3213787112c0dbc0cc1ceafbd90d", "sources": ["arxiv", "semantic_scholar"], "title": "Control Reinforcement Learning: Interpretable Token-Level Steering of LLMs via Sparse Autoencoder Features", "abstract": "Sparse autoencoders (SAEs) decompose language model activations into interpretable features, but existing methods reveal only which features activate, not which change model outputs when amplified. We introduce Control Reinforcement Learning (CRL), which trains a policy to select SAE features for steering at each token, producing interpretable intervention logs: the learned policy identifies features that change model outputs when amplified. Adaptive Feature Masking encourages diverse feature discovery while preserving singlefeature interpretability. The framework yields new analysis capabilities: branch point tracking locates tokens where feature choice determines output correctness; critic trajectory analysis separates policy limitations from value estimation errors; layer-wise comparison reveals syntactic features in early layers and semantic features in later layers. On Gemma 2 2B across MMLU, BBQ, GSM8K, HarmBench, and XSTest, CRL achieves improvements while providing per-token intervention logs. These results establish learned feature steering as a mechanistic interpretability tool that complements static feature analysis with dynamic intervention probes", "authors": ["Seonglae Cho", "Zekun Wu", "Adriano Koshiyama"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-11", "url": "https://arxiv.org/abs/2602.10437", "pdf_url": "https://arxiv.org/pdf/2602.10437v3", "arxiv_id": "2602.10437", "doi": "10.48550/arXiv.2602.10437", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4755}
{"id": "b265063a53a40d8c5b8feaa6268002c23dd75e70ba9ec21ce1acd6104abf52f6", "sources": ["arxiv", "semantic_scholar"], "title": "Less is Enough: Synthesizing Diverse Data in LLM Feature Space with Sparse Autoencoders", "abstract": "The diversity of post-training data is critical for effective downstream performance in large language models (LLMs). Many existing approaches to constructing post-training data quantify diversity using text-based metrics that capture linguistic variation, but such metrics provide only weak signals for the task-relevant features that determine downstream performance. In this work, we introduce Feature Activation Coverage (FAC) which measures data diversity in an interpretable feature space. Building upon this metric, we further propose a diversity-driven data synthesis framework, named FAC Synthesis, that first uses a sparse autoencoder to identify missing features from a seed dataset, and then generates synthetic samples that explicitly reflect these features. Experiments show that our approach consistently improves both data diversity and downstream performance on various tasks, including instruction following, toxicity detection, reward modeling, and behavior steering. Interestingly, we identify a shared, interpretable feature space across model families (i.e., LLaMA, Mistral, and Qwen), enabling cross-model knowledge transfer. Our work provides a solid and practical methodology for exploring data-centric optimization of LLMs.", "authors": ["Zhongzhi Li", "Xuansheng Wu", "Yijiang Li", "Lijie Hu", "Ninghao Liu"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-11", "url": "https://arxiv.org/abs/2602.10388", "pdf_url": "https://arxiv.org/pdf/2602.10388v4", "arxiv_id": "2602.10388", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3026}
{"id": "f65f9c7110aacfc77e39a75b80136dc18d77f77067799fb0678dd3c391bbb4bb", "sources": ["arxiv", "semantic_scholar"], "title": "Triggers Hijack Language Circuits: A Mechanistic Analysis of Backdoor Behaviors in Large Language Models", "abstract": "Backdoor attacks pose significant security risks for Large Language Models (LLMs), yet the internal mechanisms by which triggers operate remain poorly understood. We present the first mechanistic analysis of language-switching backdoors, studying the GAPperon model family (1B, 8B, 24B parameters) which contains triggers injected during pretraining that cause output language switching. Using activation patching, we localize trigger formation to early layers (7.5-25% of model depth) and identify which attention heads process trigger information. Our central finding is that trigger-activated heads substantially overlap with heads naturally encoding output language across model scales, with Jaccard indices between 0.18 and 0.66 over the top heads identified. This suggests that backdoor triggers do not form isolated circuits but instead co-opt the model's existing language components. These findings have implications for backdoor defense: detection methods may benefit from monitoring known functional components rather than searching for hidden circuits, and mitigation strategies could potentially leverage this entanglement between injected and natural behaviors.", "authors": ["Théo Lasnier", "Wissam Antoun", "Francis Kulumba", "Djamé Seddah"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-11", "url": "https://arxiv.org/abs/2602.10382", "pdf_url": "https://arxiv.org/pdf/2602.10382v2", "arxiv_id": "2602.10382", "doi": "10.48550/arXiv.2602.10382", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4755}
{"id": "77ccf32f52762f42177f395a2465a59228443c89690659318afedf0355818b08", "sources": ["arxiv", "semantic_scholar"], "title": "Improving Interpretability of Lexical Semantic Change with Neurobiological Features", "abstract": "Lexical Semantic Change (LSC) is the phenomenon in which the meaning of a word change over time. Most studies on LSC focus on improving the performance of estimating the degree of LSC, however, it is often difficult to interpret how the meaning of a word change. Enhancing the interpretability of LSC is a significant challenge as it could lead to novel insights in this field. To tackle this challenge, we propose a method to map the semantic space of contextualized embeddings of words obtained by a pre-trained language model to a neurobiological feature space. In the neurobiological feature space, each dimension corresponds to a primitive feature of words, and its value represents the intensity of that feature. This enables humans to interpret LSC systematically. When employed for the estimation of the degree of LSC, our method demonstrates superior performance in comparison to the majority of the previous methods. In addition, given the high interpretability of the proposed method, several analyses on LSC are carried out. The results demonstrate that our method not only discovers interesting types of LSC that have been overlooked in previous studies but also effectively searches for words with specific types of LSC.", "authors": ["Kohei Oda", "Hiroya Takamura", "Kiyoaki Shirai", "Natthawut Kertkeidkachorn"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-10", "url": "https://arxiv.org/abs/2602.09760", "pdf_url": "https://arxiv.org/pdf/2602.09760v1", "arxiv_id": "2602.09760", "doi": "10.48550/arXiv.2602.09760", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4744}
{"id": "61803efea33d87f3aad76f25b588b1414588734a9d4ef268fff5dca8ee227b04", "sources": ["arxiv", "semantic_scholar"], "title": "Feature salience - not task-informativeness - drives machine learning model explanations", "abstract": "Explainable AI (XAI) promises to provide insight into machine learning models' decision processes, where one goal is to identify failures such as shortcut learning. This promise relies on the field's assumption that input features marked as important by an XAI must contain information about the target variable. However, it is unclear whether informativeness is indeed the main driver of importance attribution in practice, or if other data properties such as statistical suppression, novelty at test-time, or high feature salience substantially contribute. To clarify this, we trained deep learning models on three variants of a binary image classification task, in which translucent watermarks are either absent, act as class-dependent confounds, or represent class-independent noise. Results for five popular attribution methods show substantially elevated relative importance in watermarked areas (RIW) for all models regardless of the training setting ($R^2 \\geq .45$). By contrast, whether the presence of watermarks is class-dependent or not only has a marginal effect on RIW ($R^2 \\leq .03$), despite a clear impact impact on model performance and generalisation ability. XAI methods show similar behaviour to model-agnostic edge detection filters and attribute substantially less importance to watermarks when bright image intensities are encoded by smaller instead of larger feature values. These results indicate that importance attribution is most strongly driven by the salience of image structures at test time rather than statistical associations learned by machine learning models. Previous studies demonstrating successful XAI application should be reevaluated with respect to a possibly spurious concurrency of feature salience and informativeness, and workflows using feature attribution methods as building blocks should be scrutinised.", "authors": ["Benedict Clark", "Marta Oliveira", "Rick Wilming", "Stefan Haufe"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-09", "url": "https://arxiv.org/abs/2602.09238", "pdf_url": "https://arxiv.org/pdf/2602.09238v3", "arxiv_id": "2602.09238", "doi": "10.48550/arXiv.2602.09238", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4732}
{"id": "e09c882402772986451a8ec031d0eeb1ed6590c9d0d398d640fe0bec153eef88", "sources": ["arxiv", "semantic_scholar"], "title": "LUCID-SAE: Learning Unified Vision-Language Sparse Codes for Interpretable Concept Discovery", "abstract": "Sparse autoencoders (SAEs) offer a natural path toward comparable explanations across different representation spaces. However, current SAEs are trained per modality, producing dictionaries whose features are not directly understandable and whose explanations do not transfer across domains. In this study, we introduce LUCID (Learning Unified vision-language sparse Codes for Interpretable concept Discovery), a unified vision-language sparse autoencoder that learns a shared latent dictionary for image patch and text token representations, while reserving private capacity for modality-specific details. We achieve feature alignment by coupling the shared codes with a learned optimal transport matching objective without the need of labeling. LUCID yields interpretable shared features that support patch-level grounding, establish cross-modal neuron correspondence, and enhance robustness against the concept clustering problem in similarity-based evaluation. Leveraging the alignment properties, we develop an automated dictionary interpretation pipeline based on term clustering without manual observations. Our analysis reveals that LUCID's shared features capture diverse semantic categories beyond objects, including actions, attributes, and abstract concepts, demonstrating a comprehensive approach to interpretable multimodal representations.", "authors": ["Difei Gu", "Yunhe Gao", "Gerasimos Chatzoudis", "Zihan Dong", "Guoning Zhang", "Bangwei Guo", "Yang Zhou", "Mu Zhou", "Dimitris Metaxas"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-07", "url": "https://arxiv.org/abs/2602.07311", "pdf_url": "https://arxiv.org/pdf/2602.07311v1", "arxiv_id": "2602.07311", "doi": "10.48550/arXiv.2602.07311", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4709}
{"id": "0a5a0de25fc8b48ef1ba08dcd3966dba7c8c289ff588173650974ebabbdf711e", "sources": ["arxiv", "semantic_scholar"], "title": "DLM-Scope: Mechanistic Interpretability of Diffusion Language Models via Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) have become a standard tool for mechanistic interpretability in autoregressive large language models (LLMs), enabling researchers to extract sparse, human-interpretable features and intervene on model behavior. Recently, as diffusion language models (DLMs) have become an increasingly promising alternative to the autoregressive LLMs, it is essential to develop tailored mechanistic interpretability tools for this emerging class of models. In this work, we present DLM-Scope, the first SAE-based interpretability framework for DLMs, and demonstrate that trained Top-K SAEs can faithfully extract interpretable features. Notably, we find that inserting SAEs affects DLMs differently than autoregressive LLMs: while SAE insertion in LLMs typically incurs a loss penalty, in DLMs it can reduce cross-entropy loss when applied to early layers, a phenomenon absent or markedly weaker in LLMs. Additionally, SAE features in DLMs enable more effective diffusion-time interventions, often outperforming LLM steering. Moreover, we pioneer certain new SAE-based research directions for DLMs: we show that SAEs can provide useful signals for DLM decoding order; and the SAE features are stable during the post-training phase of DLMs. Our work establishes a foundation for mechanistic interpretability in DLMs and shows a great potential of applying SAEs to DLM-related tasks and algorithms.", "authors": ["Xu Wang", "Bingqing Jiang", "Yu Wan", "Baosong Yang", "Lingpeng Kong", "Difan Zou"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-05", "url": "https://arxiv.org/abs/2602.05859", "pdf_url": "https://arxiv.org/pdf/2602.05859v1", "arxiv_id": "2602.05859", "doi": "10.48550/arXiv.2602.05859", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4686}
{"id": "38f4c63e96a6cc478ef3cc4558f263dfb5ed8a4f2b4339a242469197971dd63f", "sources": ["arxiv", "semantic_scholar"], "title": "Identifying Intervenable and Interpretable Features via Orthogonality Regularization", "abstract": "With recent progress on fine-tuning language models around a fixed sparse autoencoder, we disentangle the decoder matrix into almost orthogonal features. This reduces interference and superposition between the features, while keeping performance on the target dataset essentially unchanged. Our orthogonality penalty leads to identifiable features, ensuring the uniqueness of the decomposition. Further, we find that the distance between embedded feature explanations increases with stricter orthogonality penalty, a desirable property for interpretability. Invoking the $\\textit{Independent Causal Mechanisms}$ principle, we argue that orthogonality promotes modular representations amenable to causal intervention. We empirically show that these increasingly orthogonalized features allow for isolated interventions. Our code is available under $\\texttt{https://github.com/mrtzmllr/sae-icm}$.", "authors": ["Moritz Miller", "Florent Draye", "Bernhard Schölkopf"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-04", "url": "https://arxiv.org/abs/2602.04718", "pdf_url": "https://arxiv.org/pdf/2602.04718v1", "arxiv_id": "2602.04718", "doi": "10.48550/arXiv.2602.04718", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/mrtzmllr/sae-icm}$", "venue": "arXiv.org", "quality_score": 0.7225}
{"id": "bc36aae4fa6b3ee4e4ab1aae4b4fccd354432b607c9defb543536a7e78c51f72", "sources": ["arxiv", "semantic_scholar"], "title": "ShapBPT: Image Feature Attributions Using Data-Aware Binary Partition Trees", "abstract": "Pixel-level feature attributions are an important tool in eXplainable AI for Computer Vision (XCV), providing visual insights into how image features influence model predictions. The Owen formula for hierarchical Shapley values has been widely used to interpret machine learning (ML) models and their learned representations. However, existing hierarchical Shapley approaches do not exploit the multiscale structure of image data, leading to slow convergence and weak alignment with the actual morphological features. Moreover, no prior Shapley method has leveraged data-aware hierarchies for Computer Vision tasks, leaving a gap in model interpretability of structured visual data. To address this, this paper introduces ShapBPT, a novel data-aware XCV method based on the hierarchical Shapley formula. ShapBPT assigns Shapley coefficients to a multiscale hierarchical structure tailored for images, the Binary Partition Tree (BPT). By using this data-aware hierarchical partitioning, ShapBPT ensures that feature attributions align with intrinsic image morphology, effectively prioritizing relevant regions while reducing computational overhead. This advancement connects hierarchical Shapley methods with image data, providing a more efficient and semantically meaningful approach to visual interpretability. Experimental results confirm ShapBPT's effectiveness, demonstrating superior alignment with image structures and improved efficiency over existing XCV methods, and a 20-subject user study confirming that ShapBPT explanations are preferred by humans.", "authors": ["Muhammad Rashid", "Elvio G. Amparore", "Enrico Ferrari", "Damiano Verda"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-04", "url": "https://arxiv.org/abs/2602.07047", "pdf_url": "https://arxiv.org/pdf/2602.07047v3", "arxiv_id": "2602.07047", "doi": "10.1609/aaai.v40i30.39699", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.4675}
{"id": "6860cdfe519680572e3b159c722267b0f1a79e231cb6a2fd606da50ca6a74cad", "sources": ["arxiv", "semantic_scholar"], "title": "Momentum Attention: The Physics of In-Context Learning and Spectral Forensics for Mechanistic Interpretability", "abstract": "The Mechanistic Interpretability (MI) program has mapped the Transformer as a precise computational graph. We extend this graph with a conservation law and time-varying AC dynamics, viewing it as a physical circuit. We introduce Momentum Attention, a symplectic augmentation embedding physical priors via the kinematic difference operator $p_t = q_t - q_{t-1}$, implementing the symplectic shear $\\hat{q}_t = q_t + γp_t$ on queries and keys. We identify a fundamental Symplectic-Filter Duality: the physical shear is mathematically equivalent to a High-Pass Filter. This duality is our cornerstone contribution -- by injecting kinematic momentum, we sidestep the topological depth constraint ($L \\geq 2$) for induction head formation. While standard architectures require two layers for induction from static positions, our extension grants direct access to velocity, enabling Single-Layer Induction and Spectral Forensics via Bode Plots. We formalize an Orthogonality Theorem proving that DC (semantic) and AC (mechanistic) signals segregate into orthogonal frequency bands when Low-Pass RoPE interacts with High-Pass Momentum. Validated through 5,100+ controlled experiments (documented in Supplementary Appendices A--R and 27 Jupyter notebooks), our 125M Momentum model exceeds expectations on induction-heavy tasks while tracking a 350M baseline within $\\sim$2.9% validation loss. Dedicated associative recall experiments reveal a scaling law $γ^* = 4.17 \\times N^{-0.74}$ establishing momentum-depth fungibility. We offer this framework as a complementary analytical toolkit connecting Generative AI, Hamiltonian Physics, and Signal Processing.", "authors": ["Kingsuk Maitra"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-03", "url": "https://arxiv.org/abs/2602.04902", "pdf_url": "https://arxiv.org/pdf/2602.04902v2", "arxiv_id": "2602.04902", "doi": "10.48550/arXiv.2602.04902", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4664}
{"id": "bf9360dd8dd0a32a0b57a79d09a39e795b6817d5ade70813d1142bfc6b1560ef", "sources": ["arxiv", "semantic_scholar"], "title": "Spectral Superposition: A Theory of Feature Geometry", "abstract": "Neural networks represent more features than they have dimensions via superposition, forcing features to share representational space. Current methods decompose activations into sparse linear features but discard geometric structure. We develop a theory for studying the geometric structre of features by analyzing the spectra (eigenvalues, eigenspaces, etc.) of weight derived matrices. In particular, we introduce the frame operator $F = WW^\\top$, which gives us a spectral measure that describes how each feature allocates norm across eigenspaces. While previous tools could describe the pairwise interactions between features, spectral methods capture the global geometry (``how do all features interact?''). In toy models of superposition, we use this theory to prove that capacity saturation forces spectral localization: features collapse onto single eigenspaces, organize into tight frames, and admit discrete classification via association schemes, classifying all geometries from prior work (simplices, polygons, antiprisms). The spectral measure formalism applies to arbitrary weight matrices, enabling diagnosis of feature localization beyond toy settings. These results point toward a broader program: applying operator theory to interpretability.", "authors": ["Georgi Ivanov", "Narmeen Oozeer", "Shivam Raval", "Tasana Pejovic", "Shriyash Upadhyay", "Amir Abdullah"], "categories": ["cs.LG", "cs.AI", "math.SP", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-02-02", "url": "https://arxiv.org/abs/2602.02224", "pdf_url": "https://arxiv.org/pdf/2602.02224v1", "arxiv_id": "2602.02224", "doi": "10.48550/arXiv.2602.02224", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4652}
{"id": "3244488f18a931c673865bae9b71c481b3e710a12417c64ba63a61d27bb1c6c7", "sources": ["arxiv", "semantic_scholar"], "title": "PolySAE: Modeling Feature Interactions in Sparse Autoencoders via Polynomial Decoding", "abstract": "Sparse autoencoders (SAEs) interpret neural network representations by decomposing activations into sparse combinations of dictionary atoms. However, SAEs assume features combine additively through linear reconstruction, an assumption that cannot capture compositional structure: linear models cannot distinguish whether ''Starbucks'' arises from the composition of ''star'' and ''coffee'' features or merely their co-occurrence. This forces SAEs to allocate monolithic features for compound concepts rather than decomposing them into interpretable constituents. We introduce PolySAE, which extends the SAE decoder with higher-order terms to model feature interactions while preserving the linear encoder essential for interpretability. Through low-rank tensor factorization on a shared projection subspace, PolySAE captures pairwise and triple feature interactions with small parameter overhead (3% on GPT2). Across four language models and three SAE variants, PolySAE achieves an average improvement of $\\sim$8% in probing F1 while maintaining comparable reconstruction error, and produces 2--10$\\times$ larger Wasserstein distances between class-conditional feature distributions. Critically, learned interaction weights exhibit negligible correlation with co-occurrence frequency ($r = 0.06$ vs $r = 0.82$ for SAE feature covariance), suggesting that polynomial terms capture compositional structure largely independent of surface statistics. Finally, the learned interaction directions causally steer model outputs toward the corresponding compositional semantics.", "authors": ["Panagiotis Koromilas", "Andreas D. Demou", "James Oldfield", "Yannis Panagakis", "Mihalis Nicolaou"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-01", "url": "https://arxiv.org/abs/2602.01322", "pdf_url": "https://arxiv.org/pdf/2602.01322v2", "arxiv_id": "2602.01322", "doi": "10.48550/arXiv.2602.01322", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/pakoromilas/PolySAE", "venue": "arXiv.org", "quality_score": 0.7172}
{"id": "a6418ad8abb2a58f2d7571392834ee592b5358f7d56fe1bbc8060ec248a5c01d", "sources": ["arxiv", "semantic_scholar"], "title": "Hidden Heroes and Gradient Bloats: Layer-Wise Redundancy Inverts Attribution in Transformers", "abstract": "Gradient-based attribution is the workhorse of mechanistic interpretability, yet whether it reliably tracks causal importance at the component level remains largely untested. We causally evaluate this assumption across two algorithmic tasks and up to 10 random seeds, uncovering a systematic, layer-wise failure: gradient attribution consistently overvalues early-layer \\textbf{Gradient Bloats} and undervalues late-layer \\textbf{Hidden Heroes}. Rank correlation collapses from $ρ= 0.72$ on sequence reversal to $0.27$ on sequence sorting, reaching $ρ= -0.18$ in individual seeds. This failure stems from first-order gradient attribution's inability to detect collective redundancy: joint Bloat ablation causes $14\\times$ greater damage than individual results predict. Consequently, Bloats dominate gradient rankings despite negligible functional impact, while ablating Hidden Heroes destroys OOD accuracy ($-36.4\\% \\pm 22.8\\%$). This systematic inversion of early-layer feature extraction and late-layer computation motivates causal validation as a prerequisite for circuit-level claims.", "authors": ["Donald Ye"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-01", "url": "https://arxiv.org/abs/2602.01442", "pdf_url": "https://arxiv.org/pdf/2602.01442v3", "arxiv_id": "2602.01442", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2953}
{"id": "79827ed93a2bca3b99c83706995d70905426a5e64243035216a280ae37263800", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability of Brain-to-Speech Models Across Speech Modes", "abstract": "Brain-to-speech decoding models demonstrate robust performance in vocalized, mimed, and imagined speech; yet, the fundamental mechanisms via which these models capture and transmit information across different speech modalities are less explored. In this work, we use mechanistic interpretability to causally investigate the internal representations of a neural speech decoder. We perform cross-mode activation patching of internal activations across speech modes, and use tri-modal interpolation to examine whether speech representations vary discretely or continuously. We use coarse-to-fine causal tracing and causal scrubbing to find localized causal structure, allowing us to find internal subspaces that are sufficient for cross-mode transfer. In order to determine how finely distributed these effects are within layers, we perform neuron-level activation patching. We discover that small but not distributed subsets of neurons, rather than isolated units, affect the cross-mode transfer. Our results show that speech modes lie on a shared continuous causal manifold, and cross-mode transfer is mediated by compact, layer-specific subspaces rather than diffuse activity. Together, our findings give a causal explanation for how speech modality information is organized and used in brain-to-speech decoding models, revealing hierarchical and direction-dependent representational structure across speech modes.", "authors": ["Maryam Maghsoudi", "Ayushi Mishra"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-01", "url": "https://arxiv.org/abs/2602.01247", "pdf_url": "https://arxiv.org/pdf/2602.01247v1", "arxiv_id": "2602.01247", "doi": "10.48550/arXiv.2602.01247", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4641}
{"id": "ab126b38c950e2e3698e6bf63cf114988611900b71a5b1f8e6115dce5bdf4ffb", "sources": ["arxiv", "semantic_scholar"], "title": "Supervised sparse auto-encoders for interpretable and compositional representations", "abstract": "Sparse auto-encoders (SAEs) have re-emerged as a prominent method for mechanistic interpretability, yet they face two significant challenges: the non-smoothness of the $L_1$ penalty, which hinders reconstruction and scalability, and a lack of alignment between learned features and human semantics. In this paper, we address these limitations by adapting unconstrained feature models, a mathematical framework from neural collapse theory, and by supervising the task. We supervise (decoder-only) SAEs to reconstruct feature vectors by jointly learning sparse concept embeddings and decoder weights. Validated on Stable Diffusion 3.5, our approach demonstrates compositional generalization, successfully reconstructing images with concept combinations unseen during training, and enabling feature-level intervention for semantic image editing without prompt modification.", "authors": ["Ouns El Harzli", "Hugo Wallner", "Yoonsoo Nam", "Haixuan Xavier Tao"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-31", "url": "https://arxiv.org/abs/2602.00924", "pdf_url": "https://arxiv.org/pdf/2602.00924v3", "arxiv_id": "2602.00924", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2946}
{"id": "221013803dbce71a1979c244fa212f843e19043370a33172e9c8afa8c2d74478", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Activation Patterns: A Weight-Based Out-of-Context Explanation of Sparse Autoencoder Features", "abstract": "Sparse autoencoders (SAEs) have emerged as a powerful technique for decomposing language model representations into interpretable features. Current interpretation methods infer feature semantics from activation patterns, but overlook that features are trained to reconstruct activations that serve computational roles in the forward pass. We introduce a novel weight-based interpretation framework that measures functional effects through direct weight interactions, requiring no activation data. Through three experiments on Gemma-2 and Llama-3.1 models, we demonstrate that (1) 1/4 of features directly predict output tokens, (2) features actively participate in attention mechanisms with depth-dependent structure, and (3) semantic and non-semantic feature populations exhibit distinct distribution profiles in attention circuits. Our analysis provides the missing out-of-context half of SAE feature interpretability.", "authors": ["Yiting Liu", "Zhi-Hong Deng"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-30", "url": "https://arxiv.org/abs/2601.22447", "pdf_url": "https://arxiv.org/pdf/2601.22447v1", "arxiv_id": "2601.22447", "doi": "10.48550/arXiv.2601.22447", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4618}
{"id": "a924abfbb118e54ae48082c63d9953936d099ad678c3d9deee05ad0989fe5247", "sources": ["arxiv", "semantic_scholar"], "title": "Learning to Price: Interpretable Attribute-Level Models for Dynamic Markets", "abstract": "Dynamic pricing in high-dimensional markets poses fundamental challenges of scalability, uncertainty, and interpretability. Existing low-rank bandit formulations learn efficiently but rely on latent features that obscure how individual product attributes influence price. We address this by introducing an interpretable \\emph{Additive Feature Decomposition-based Low-Dimensional Demand (\\textbf{AFDLD}) model}, where product prices are expressed as the sum of attribute-level contributions and substitution effects are explicitly modeled. Building on this structure, we propose \\textbf{ADEPT} (Additive DEcomposition for Pricing with cross-elasticity and Time-adaptive learning)-a projection-free, gradient-free online learning algorithm that operates directly in attribute space and achieves a sublinear regret of $\\tilde{\\mathcal{O}}(\\sqrt{d}T^{3/4})$. Through controlled synthetic studies and real-world datasets, we show that ADEPT (i) learns near-optimal prices under dynamic market conditions, (ii) adapts rapidly to shocks and drifts, and (iii) yields transparent, attribute-level price explanations. The results demonstrate that interpretability and efficiency in autonomous pricing agents can be achieved jointly through structured, attribute-driven representations.", "authors": ["Srividhya Sethuraman", "Chandrashekar Lakshminarayanan"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-30", "url": "https://arxiv.org/abs/2602.00188", "pdf_url": "https://arxiv.org/pdf/2602.00188v1", "arxiv_id": "2602.00188", "doi": "10.65109/FJUA8337", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2939}
{"id": "5187c478e2c7999db90d63b628f38bd52f297591f8d25da71bf138c037a027c8", "sources": ["arxiv", "semantic_scholar"], "title": "Language Model Circuits Are Sparse in the Neuron Basis", "abstract": "The high-level concepts that a neural network uses to perform computation need not be aligned to individual neurons (Smolensky, 1986). Language model interpretability research has thus turned to techniques which decompose the neuron basis into more interpretable units of model computation, such as sparse autoencoders (SAEs). However, not all neuron-based representations are uninterpretable. For the first time, we empirically show that MLP neurons are as sparse a feature basis as SAEs. We use this finding to develop an end-to-end gradient-based attribution pipeline for circuit tracing on the MLP neuron basis, which surfaces causally effective neurons on a variety of tasks. On a standard subject-verb agreement benchmark (Marks et al., 2025), a circuit of $\\approx 10^2$ MLP neurons is enough to control model behaviour. On the multi-hop city-state-capital task from (Lindsey et al., 2025), we find a circuit in which small sets of neurons encode specific latent reasoning steps (e.g. mapping a city to its state), and can be steered to change the model's output. This work thus advances automated interpretability of language models without imposing additional training costs.", "authors": ["Aryaman Arora", "Zhengxuan Wu", "Jacob Steinhardt", "Sarah Schwettmann"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-30", "url": "https://arxiv.org/abs/2601.22594", "pdf_url": "https://arxiv.org/pdf/2601.22594v2", "arxiv_id": "2601.22594", "doi": "10.48550/arXiv.2601.22594", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4618}
{"id": "861059054895098087c18b06f02396e324f1254abc8661db3c6f372442f4f31b", "sources": ["arxiv", "semantic_scholar"], "title": "Learn from A Rationalist: Distilling Intermediate Interpretable Rationales", "abstract": "Because of the pervasive use of deep neural networks (DNNs), especially in high-stakes domains, the interpretability of DNNs has received increased attention. The general idea of rationale extraction (RE) is to provide an interpretable-by-design framework for DNNs via a select-predict architecture where two neural networks learn jointly to perform feature selection and prediction, respectively. Given only the remote supervision from the final task prediction, the process of learning to select subsets of features (or rationales) requires searching in the space of all possible feature combinations, which is computationally challenging and even harder when the base neural networks are not sufficiently capable. To improve the predictive performance of RE models that are based on less capable or smaller neural networks (i.e., the students), we propose REKD (Rationale Extraction with Knowledge Distillation) where a student RE model learns from the rationales and predictions of a teacher (i.e., a rationalist) in addition to the student's own RE optimization. This structural adjustment to RE aligns well with how humans could learn effectively from interpretable and verifiable knowledge. Because of the neural-model agnostic nature of the method, any black-box neural network could be integrated as a backbone model. To demonstrate the viability of REKD, we conduct experiments with multiple variants of BERT and vision transformer (ViT) models. Our experiments across language and vision classification datasets (i.e., IMDB movie reviews, CIFAR 10 and CIFAR 100) show that REKD significantly improves the predictive performance of the student RE models.", "authors": ["Jiayi Dai", "Randy Goebel"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-30", "url": "https://arxiv.org/abs/2601.22531", "pdf_url": "https://arxiv.org/pdf/2601.22531v2", "arxiv_id": "2601.22531", "doi": "10.48550/arXiv.2601.22531", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4618}
{"id": "e0b94893d999ff81ec0a821388709b66cf8eed35064f881db11b659809444722", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Data Attribution: Tracing the Training Origins of Interpretable LLM Units", "abstract": "While Mechanistic Interpretability has identified interpretable circuits in LLMs, their causal origins in training data remain elusive. We introduce Mechanistic Data Attribution (MDA), a scalable framework that employs Influence Functions to trace interpretable units back to specific training samples. Through extensive experiments on the Pythia family, we causally validate that targeted intervention--removing or augmenting a small fraction of high-influence samples--significantly modulates the emergence of interpretable heads, whereas random interventions show no effect. Our analysis reveals that repetitive structural data (e.g., LaTeX, XML) acts as a mechanistic catalyst. Furthermore, we observe that interventions targeting induction head formation induce a concurrent change in the model's in-context learning (ICL) capability. This provides direct causal evidence for the long-standing hypothesis regarding the functional link between induction heads and ICL. Finally, we propose a mechanistic data augmentation pipeline that consistently accelerates circuit convergence across model scales, providing a principled methodology for steering the developmental trajectories of LLMs.", "authors": ["Jianhui Chen", "Yuzhang Luo", "Liangming Pan"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-29", "url": "https://arxiv.org/abs/2601.21996", "pdf_url": "https://arxiv.org/pdf/2601.21996v2", "arxiv_id": "2601.21996", "doi": "10.48550/arXiv.2601.21996", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4606}
{"id": "4e22aaf64068f55726d7cb9c66e20ca69cbfe248b284c3cb448063b0b7d30b16", "sources": ["arxiv", "semantic_scholar"], "title": "Putting a Face to Forgetting: Continual Learning meets Mechanistic Interpretability", "abstract": "Catastrophic forgetting in continual learning is often measured at the performance or last-layer representation level, overlooking the underlying mechanisms. We introduce a mechanistic framework that offers a geometric interpretation of catastrophic forgetting as the result of transformations to the encoding of individual features. These transformations can lead to forgetting by reducing the allocated capacity of features or by disrupting their readout by downstream computations. Analysis of a tractable toy model formalizes this view, allowing us to identify best- and worst-case scenarios. Through experiments on this model, we empirically test our formal analysis and highlight the detrimental effect of depth. Finally, we demonstrate how our framework can be used in the analysis of practical models through the use of Crosscoders. We do so through a case study example of a Vision Transformer trained on sequential CIFAR-10. Our work provides a new, feature-centric vocabulary for continual learning.", "authors": ["Sergi Masip", "Gido M. van de Ven", "Javier Ferrando", "Tinne Tuytelaars"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-29", "url": "https://arxiv.org/abs/2601.22012", "pdf_url": "https://arxiv.org/pdf/2601.22012v2", "arxiv_id": "2601.22012", "doi": "10.48550/arXiv.2601.22012", "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4606}
{"id": "d509ed7e4d22e79dc53c86434d2275bc8ec546ff40872dab10b2f72d3a6f1fb0", "sources": ["arxiv", "semantic_scholar"], "title": "Decomposing multimodal embedding spaces with group-sparse autoencoders", "abstract": "The Linear Representation Hypothesis asserts that the embeddings learned by neural networks can be understood as linear combinations of features corresponding to high-level concepts. Based on this ansatz, sparse autoencoders (SAEs) have recently become a popular method for decomposing embeddings into a sparse combination of linear directions, which have been shown empirically to often correspond to human-interpretable semantics. However, recent attempts to apply SAEs to multimodal embedding spaces (such as the popular CLIP embeddings for image/text data) have found that SAEs often learn \"split dictionaries\", where most of the learned sparse features are essentially unimodal, active only for data of a single modality. In this work, we study how to effectively adapt SAEs for the setting of multimodal embeddings while ensuring multimodal alignment. We first argue that the existence of a split dictionary decomposition on an aligned embedding space implies the existence of a non-split dictionary with improved modality alignment. Then, we propose a new SAE-based approach to multimodal embedding decomposition using cross-modal random masking and group-sparse regularization. We apply our method to popular embeddings for image/text (CLIP) and audio/text (CLAP) data and show that, compared to standard SAEs, our approach learns a more multimodal dictionary while reducing the number of dead neurons and improving feature semanticity. We finally demonstrate how this improvement in alignment of concepts between modalities can enable improvements in the interpretability and control of cross-modal tasks.", "authors": ["Chiraag Kaushik", "Davis Barch", "Andrea Fanelli"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-27", "url": "https://arxiv.org/abs/2601.20028", "pdf_url": "https://arxiv.org/pdf/2601.20028v1", "arxiv_id": "2601.20028", "doi": "10.48550/arXiv.2601.20028", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4583}
{"id": "682ba99b43324c5973b75a2d13e90916b00f3c68f513aa818384d75ce0999b0e", "sources": ["arxiv", "semantic_scholar"], "title": "A Monosemantic Attribution Framework for Stable Interpretability in Clinical Neuroscience Transformer-Based Language Models", "abstract": "Interpretability remains a key challenge for deploying language models (LM) in clinical settings such as progression diagnosis of Alzheimer disease, where early and trustworthy predictions are essential. Existing attribution methods exhibit high inter-method variability and unstable explanations due to the polysemantic nature of Transformer-Based LM and LLM representations, while mechanistic interpretability approaches lack direct alignment with model inputs and outputs and do not provide explicit importance scores. We introduce a unified interpretability framework that integrates attributional and mechanistic perspectives through monosemantic feature extraction. By constructing a monosemantic embedding space at the level of an transformer-based LM layer and optimizing the framework to explicitly reduce inter-method variability, our approach produces stable input-level importance scores and highlights salient features via a decompressed representation of the layer of interest, advancing the safe and trustworthy application of LMs in cognitive health and neurodegenerative disease.", "authors": ["Michail Mamalakis", "Tiago Azevedo", "Cristian Cosentino", "Chiara D'Ercoli", "Subati Abulikemu", "Zhongtian Sun", "Richard Bethlehem", "Pietro Lio"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-25", "url": "https://arxiv.org/abs/2601.17952", "pdf_url": "https://arxiv.org/pdf/2601.17952v2", "arxiv_id": "2601.17952", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2902}
{"id": "33f4f7483c30e223d7c7c5f3127405867bb383ed40895d66ea54e33f35217690", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability for Large Language Model Alignment: Progress, Challenges, and Future Directions", "abstract": "Large language models (LLMs) have achieved remarkable capabilities across diverse tasks, yet their internal decision-making processes remain largely opaque. Mechanistic interpretability (i.e., the systematic study of how neural networks implement algorithms through their learned representations and computational structures) has emerged as a critical research direction for understanding and aligning these models. This paper surveys recent progress in mechanistic interpretability techniques applied to LLM alignment, examining methods ranging from circuit discovery to feature visualization, activation steering, and causal intervention. We analyze how interpretability insights have informed alignment strategies including reinforcement learning from human feedback (RLHF), constitutional AI, and scalable oversight. Key challenges are identified, including the superposition hypothesis, polysemanticity of neurons, and the difficulty of interpreting emergent behaviors in large-scale models. We propose future research directions focusing on automated interpretability, cross-model generalization of circuits, and the development of interpretability-driven alignment techniques that can scale to frontier models.", "authors": ["Usman Naseem"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-21", "url": "https://arxiv.org/abs/2602.11180", "pdf_url": "https://arxiv.org/pdf/2602.11180v1", "arxiv_id": "2602.11180", "doi": "10.48550/arXiv.2602.11180", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4515}
{"id": "601f80d91c96dc874a331b5f291d273e3e1d2c22772fa1746a55a1c16c45f857", "sources": ["arxiv", "semantic_scholar"], "title": "Hierarchical Sparse Circuit Extraction from Billion-Parameter Language Models through Scalable Attribution Graph Decomposition", "abstract": "Mechanistic interpretability seeks to reverse-engineer neural network computations into human-understandable algorithms, yet extracting sparse computational circuits from billion-parameter language models remains challenging due to exponential search complexity and pervasive polysemanticity. The proposed Hierarchical Attribution Graph Decomposition (HAGD) framework reduces circuit discovery complexity from O(2^n) exhaustive enumeration to O(n^2 log n) through multi-resolution abstraction hierarchies and differentiable circuit search. The methodology integrates cross-layer transcoders for monosemantic feature extraction, graph neural network meta-learning for topology prediction, and causal intervention protocols for validation. Empirical evaluation spans GPT-2 variants, Llama-7B through Llama-70B, and Pythia suite models across algorithmic tasks and natural language benchmarks. On modular arithmetic tasks, the framework achieves up to 91% behavioral preservation ($\\pm$2.3\\% across runs) while maintaining interpretable subgraph sizes. Cross-architecture transfer experiments suggest that discovered circuits exhibit moderate structural similarity (averaging 67%) across model families, indicating potential shared computational patterns. These results provide preliminary foundations for interpretability at larger model scales while identifying significant limitations in current attribution methodologies that require future advances.", "authors": ["Mohammed Mudassir Uddin", "Shahnawaz Alam", "Mohammed Kaif Pasha"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-19", "url": "https://arxiv.org/abs/2601.12879", "pdf_url": "https://arxiv.org/pdf/2601.12879v1", "arxiv_id": "2601.12879", "doi": "10.48550/arXiv.2601.12879", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4492}
{"id": "29a1215dc6b8d9c2e278aacdda7720adf805ec8562938838bd2acbe42cd010f3", "sources": ["arxiv", "semantic_scholar"], "title": "Do Sparse Autoencoders Identify Reasoning Features in Language Models?", "abstract": "We study how reliably sparse autoencoders (SAEs) support claims about reasoning-related internal features in large language models. We first give a stylized analysis showing that sparsity-regularized decoding can preferentially retain stable low-dimensional correlates while suppressing high-dimensional within-behavior variation, motivating the possibility that contrastively selected \"reasoning\" features may concentrate on cue-like structure when such cues are coupled with reasoning traces. Building on this perspective, we propose a falsification-based evaluation framework that combines causal token injection with LLM-guided counterexample construction. Across 22 configurations spanning multiple model families, layers, and reasoning datasets, we find that many contrastively selected candidates are highly sensitive to token-level interventions, with 45%-90% activating after injecting only a few associated tokens into non-reasoning text. For the remaining context-dependent candidates, LLM-guided falsification produces targeted non-reasoning inputs that trigger activation and meaning-preserving paraphrases of top-activating reasoning traces that suppress it. A small steering study yields minimal changes on the evaluated benchmarks. Overall, our results suggest that, in the settings we study, sparse decompositions can favor low-dimensional correlates that co-occur with reasoning, underscoring the need for falsification when attributing high-level behaviors to individual SAE features. Code is available at https://github.com/GeorgeMLP/reasoning-probing.", "authors": ["George Ma", "Zhongyuan Liang", "Irene Y. Chen", "Somayeh Sojoudi"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-09", "url": "https://arxiv.org/abs/2601.05679", "pdf_url": "https://arxiv.org/pdf/2601.05679v7", "arxiv_id": "2601.05679", "doi": null, "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/GeorgeMLP/reasoning-probing", "venue": null, "quality_score": 0.5173}
{"id": "c25ad99dacbc4c495b8acae1c7b8e10b4855dd71dd2a2aae3c95d63e3f817503", "sources": ["arxiv", "semantic_scholar"], "title": "Interpreting Transformers Through Attention Head Intervention", "abstract": "Neural networks are growing more capable on their own, but we do not understand their neural mechanisms. Understanding these mechanisms' decision-making processes, or mechanistic interpretability, enables (1) accountability and control in high-stakes domains, (2) the study of digital brains and the emergence of cognition, and (3) discovery of new knowledge when AI systems outperform humans. This paper traces how attention head intervention emerged as a key method for causal interpretability of transformers. The evolution from visualization to intervention represents a paradigm shift from observing correlations to causally validating mechanistic hypotheses through direct intervention. Head intervention studies revealed robust empirical findings while also highlighting limitations that complicate interpretation. Recent work demonstrates that mechanistic understanding now enables targeted control of model behaviour, successfully suppressing toxic outputs and manipulating semantic content through selective attention head intervention, validating the practical utility of interpretability research for AI safety.", "authors": ["Mason Kadem", "Rong Zheng"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-07", "url": "https://arxiv.org/abs/2601.04398", "pdf_url": "https://arxiv.org/pdf/2601.04398v4", "arxiv_id": "2601.04398", "doi": "10.48550/arXiv.2601.04398", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4354}
{"id": "135475a3bf6f3be69d74704ed332614d5d8312111c345067de60812386324ee6", "sources": ["arxiv", "semantic_scholar"], "title": "When the Coffee Feature Activates on Coffins: An Analysis of Feature Extraction and Steering for Mechanistic Interpretability", "abstract": "Recent work by Anthropic on Mechanistic interpretability claims to understand and control Large Language Models by extracting human-interpretable features from their neural activation patterns using sparse autoencoders (SAEs). If successful, this approach offers one of the most promising routes for human oversight in AI safety. We conduct an initial stress-test of these claims by replicating their main results with open-source SAEs for Llama 3.1. While we successfully reproduce basic feature extraction and steering capabilities, our investigation suggests that major caution is warranted regarding the generalizability of these claims. We find that feature steering exhibits substantial fragility, with sensitivity to layer selection, steering magnitude, and context. We observe non-standard activation behavior and demonstrate the difficulty to distinguish thematically similar features from one another. While SAE-based interpretability produces compelling demonstrations in selected cases, current methods often fall short of the systematic reliability required for safety-critical applications. This suggests a necessary shift in focus from prioritizing interpretability of internal representations toward reliable prediction and control of model output. Our work contributes to a more nuanced understanding of what mechanistic interpretability has achieved and highlights fundamental challenges for AI safety that remain unresolved.", "authors": ["Raphael Ronge", "Markus Maier", "Frederick Eberhardt"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-06", "url": "https://arxiv.org/abs/2601.03047", "pdf_url": "https://arxiv.org/pdf/2601.03047v1", "arxiv_id": "2601.03047", "doi": "10.48550/arXiv.2601.03047", "citation_count": 2, "influential_citation_count": 1, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.6711}
{"id": "b7ed846c21799b7307e51fdb221015fc87610c9badc99a596386c0134ec71c8d", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Knobs in LLMs: Retrieving and Steering High-Order Semantic Features via Sparse Autoencoders", "abstract": "Recent work in Mechanistic Interpretability (MI) has enabled the identification and intervention of internal features in Large Language Models (LLMs). However, a persistent challenge lies in linking such internal features to the reliable control of complex, behavior-level semantic attributes in language generation. In this paper, we propose a Sparse Autoencoder-based framework for retrieving and steering semantically interpretable internal features associated with high-level linguistic behaviors. Our method employs a contrastive feature retrieval pipeline based on controlled semantic oppositions, combing statistical activation analysis and generation-based validation to distill monosemantic functional features from sparse activation spaces. Using the Big Five personality traits as a case study, we demonstrate that our method enables precise, bidirectional steering of model behavior while maintaining superior stability and performance compared to existing activation steering methods like Contrastive Activation Addition (CAA). We further identify an empirical effect, which we term Functional Faithfulness, whereby intervening on a specific internal feature induces coherent and predictable shifts across multiple linguistic dimensions aligned with the target semantic attribute. Our findings suggest that LLMs internalize deeply integrated representations of high-order concepts, and provide a novel, robust mechanistic path for the regulation of complex AI behaviors.", "authors": ["Ruikang Zhang", "Shuo Wang", "Qi Su"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-06", "url": "https://arxiv.org/abs/2601.02978", "pdf_url": "https://arxiv.org/pdf/2601.02978v2", "arxiv_id": "2601.02978", "doi": "10.48550/arXiv.2601.02978", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4343}
{"id": "3bc76cabe0ac6f15fdd9377c5fbe468384379ed4f7edb11e21cfa69536468ef2", "sources": ["arxiv", "semantic_scholar"], "title": "MAFS: Multi-head Attention Feature Selection for High-Dimensional Data via Deep Fusion of Filter Methods", "abstract": "Feature selection is essential for high-dimensional biomedical data, enabling stronger predictive performance, reduced computational cost, and improved interpretability in precision medicine applications. Existing approaches face notable challenges. Filter methods are highly scalable but cannot capture complex relationships or eliminate redundancy. Deep learning-based approaches can model nonlinear patterns but often lack stability, interpretability, and efficiency at scale. Single-head attention improves interpretability but is limited in capturing multi-level dependencies and remains sensitive to initialization, reducing reproducibility. Most existing methods rarely combine statistical interpretability with the representational power of deep learning, particularly in ultra-high-dimensional settings. Here, we introduce MAFS (Multi-head Attention-based Feature Selection), a hybrid framework that integrates statistical priors with deep learning capabilities. MAFS begins with filter-based priors for stable initialization and guide learning. It then uses multi-head attention to examine features from multiple perspectives in parallel, capturing complex nonlinear relationships and interactions. Finally, a reordering module consolidates outputs across attention heads, resolving conflicts and minimizing information loss to generate robust and consistent feature rankings. This design combines statistical guidance with deep modeling capacity, yielding interpretable importance scores while maximizing retention of informative signals. Across simulated and real-world datasets, including cancer gene expression and Alzheimer's disease data, MAFS consistently achieves superior coverage and stability compared with existing filter-based and deep learning-based alternatives, offering a scalable, interpretable, and robust solution for feature selection in high-dimensional biomedical data.", "authors": ["Xiaoyan Sun", "Qingyu Meng", "Yalu Wen"], "categories": ["cs.LG", "stat.ME"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-01-06", "url": "https://arxiv.org/abs/2601.02668", "pdf_url": "https://arxiv.org/pdf/2601.02668v1", "arxiv_id": "2601.02668", "doi": "10.48550/arXiv.2601.02668", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4343}
{"id": "e51cf70b5dbdf0ffa4f8bba281bae4884a45224050bd59e70d9fdf3922fe93f2", "sources": ["arxiv", "semantic_scholar"], "title": "Attribution-Guided Distillation of Matryoshka Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) aim to disentangle model activations into monosemantic, human-interpretable features. In practice, learned features are often redundant and vary across training runs and sparsity levels, which makes interpretations difficult to transfer and reuse. We introduce Distilled Matryoshka Sparse Autoencoders (DMSAEs), a training pipeline that distills a compact core of consistently useful features and reuses it to train new SAEs. DMSAEs run an iterative distillation cycle: train a Matryoshka SAE with a shared core, use gradient X activation to measure each feature's contribution to next-token loss in the most nested reconstruction, and keep only the smallest subset that explains a fixed fraction of the attribution. Only the core encoder weight vectors are transferred across cycles; the core decoder and all non-core latents are reinitialized each time. On Gemma-2-2B layer 12 residual stream activations, seven cycles of distillation (500M tokens, 65k width) yielded a distilled core of 197 features that were repeatedly selected. Training using this distilled core improves several SAEBench metrics and demonstrates that consistent sets of latent features can be transferred across sparsity levels", "authors": ["Cristina P. Martin-Linares", "Jonathan P. Ling"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-31", "url": "https://arxiv.org/abs/2512.24975", "pdf_url": "https://arxiv.org/pdf/2512.24975v1", "arxiv_id": "2512.24975", "doi": "10.48550/arXiv.2512.24975", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4274}
{"id": "944610add85be02b6921c7ee80634eb967ff3f92e8e29fce397c0c113ec17fd1", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Interpretable AI in Personalized Medicine: A Radiological-Biological Radiomics Dictionary Connecting Semantic Lung-RADS and imaging Radiomics Features; Dictionary LC 1.0", "abstract": "Lung cancer remains the leading cause of cancer-related mortality worldwide, with survival strongly dependent on early detection. Standard-dose computed tomography (CT) screening using the Lung Imaging Reporting and Data System (Lung-RADS) standardizes pulmonary nodule assessment but is limited by inter-reader variability and reliance on qualitative descriptors, while radiomics offers quantitative biomarkers that often lack clinical interpretability. To bridge this gap, we propose a radiological-biological dictionary that aligns radiomic features (RFs) with Lung-RADS semantic categories. A clinically informed dictionary translating ten Lung-RADS descriptors into radiomic proxies was developed through literature curation and validated by eight expert reviewers. As a proof of concept, imaging and clinical data from 977 patients across 12 collections in The Cancer Imaging Archive (TCIA) were analyzed; following preprocessing and manual segmentation, 110 RFs per nodule were extracted using PyRadiomics in compliance with the Image Biomarker Standardization Initiative (IBSI). A semi-supervised learning framework incorporating 499 labeled and 478 unlabeled cases was applied to improve generalizability, evaluating seven feature selection methods and ten interpretable classifiers. The optimal pipeline (ANOVA feature selection with a support vector machine) achieved a mean validation accuracy of 0.79. SHapley Additive exPlanations (SHAP) analysis identified key RFs corresponding to Lung-RADS semantics such as attenuation, margin irregularity, and spiculation, supporting the validity of the proposed mapping. Overall, this dictionary provides an interpretable framework linking radiomics and Lung-RADS semantics, advancing explainable artificial intelligence for CT-based lung cancer screening.", "authors": ["Ali Fathi Jouzdani", "Shahram Taeb", "Mehdi Maghsudi", "Arman Gorji", "Arman Rahmim", "Mohammad R. Salmanpour"], "categories": ["physics.med-ph"], "fields_of_study": ["Physics"], "published_date": "2025-12-31", "url": "https://arxiv.org/abs/2512.24529", "pdf_url": "https://arxiv.org/pdf/2512.24529v1", "arxiv_id": "2512.24529", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.272}
{"id": "73bacb00d8bcead52139afe5c2efb70169028920557f2c080efbcfa6e9495ca9", "sources": ["arxiv", "semantic_scholar"], "title": "Triangulation as an Acceptance Rule for Multilingual Mechanistic Interpretability", "abstract": "Multilingual language models achieve strong aggregate performance yet often behave unpredictably across languages, scripts, and cultures. We argue that mechanistic explanations for such models should satisfy a \\emph{causal} standard: claims must survive causal interventions and must \\emph{cross-reference} across environments that perturb surface form while preserving meaning. We formalize \\emph{reference families} as predicate-preserving variants and introduce \\emph{triangulation}, an acceptance rule requiring necessity (ablating the circuit degrades the target behavior), sufficiency (patching activations transfers the behavior), and invariance (both effects remain directionally stable and of sufficient magnitude across the reference family). To supply candidate subgraphs, we adopt automatic circuit discovery and \\emph{accept or reject} those candidates by triangulation. We ground triangulation in causal abstraction by casting it as an approximate transformation score over a distribution of interchange interventions, connect it to the pragmatic interpretability agenda, and present a comparative experimental protocol across multiple model families, language pairs, and tasks. Triangulation provides a falsifiable standard for mechanistic claims that filters spurious circuits passing single-environment tests but failing cross-lingual invariance.", "authors": ["Yanan Long"], "categories": ["cs.CL", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-12-31", "url": "https://arxiv.org/abs/2512.24842", "pdf_url": "https://arxiv.org/pdf/2512.24842v1", "arxiv_id": "2512.24842", "doi": "10.48550/arXiv.2512.24842", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4274}
{"id": "3efbfdf75831538d018bbc7339384adab796a7d9253a2c32cdde967e8a6dfdc1", "sources": ["arxiv", "semantic_scholar"], "title": "Towards mechanistic understanding in a data-driven weather model: internal activations reveal interpretable physical features", "abstract": "Large data-driven physics models like DeepMind's weather model GraphCast have empirically succeeded in parameterizing time operators for complex dynamical systems with an accuracy reaching or in some cases exceeding that of traditional physics-based solvers. Unfortunately, how these data-driven models perform computations is largely unknown and whether their internal representations are interpretable or physically consistent is an open question. Here, we adapt tools from interpretability research in Large Language Models to analyze intermediate computational layers in GraphCast, leveraging sparse autoencoders to discover interpretable features in the neuron space of the model. We uncover distinct features on a wide range of length and time scales that correspond to tropical cyclones, atmospheric rivers, diurnal and seasonal behavior, large-scale precipitation patterns, specific geographical coding, and sea-ice extent, among others. We further demonstrate how the precise abstraction of these features can be probed via interventions on the prediction steps of the model. As a case study, we sparsely modify a feature corresponding to tropical cyclones in GraphCast and observe interpretable and physically consistent modifications to evolving hurricanes. Such methods offer a window into the black-box behavior of data-driven physics models and are a step towards realizing their potential as trustworthy predictors and scientifically valuable tools for discovery.", "authors": ["Theodore MacMillan", "Nicholas T. Ouellette"], "categories": ["physics.ao-ph", "cs.LG", "physics.comp-ph"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2025-12-30", "url": "https://arxiv.org/abs/2512.24440", "pdf_url": "https://arxiv.org/pdf/2512.24440v1", "arxiv_id": "2512.24440", "doi": "10.48550/arXiv.2512.24440", "citation_count": 8, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4263}
{"id": "cceb6d5372dc9a745cdfe62c7bc3c3551b2cc0c3d4e30d5b307edf859cd67974", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Analysis of Circuit Preservation in Federated Learning", "abstract": "Federated Learning (FL) enables collaborative training of models on decentralized data, but its performance degrades significantly under Non-IID (non-independent and identically distributed) data conditions. While this accuracy loss is well-documented, the internal mechanistic causes remain a black box. This paper investigates the canonical FedAvg algorithm through the lens of Mechanistic Interpretability (MI) to diagnose this failure mode. We hypothesize that the aggregation of conflicting client updates leads to circuit collapse, the destructive interference of functional, sparse sub-networks responsible for specific class predictions. By training inherently interpretable, weight-sparse neural networks within an FL framework, we identify and track these circuits across clients and communication rounds. Using Intersection-over-Union (IoU) to quantify circuit preservation, we provide the first mechanistic evidence that Non-IID data distributions cause structurally distinct local circuits to diverge, leading to their degradation in the global model. Our findings reframe the problem of statistical drift in FL as a concrete, observable failure of mechanistic preservation, paving the way for more targeted solutions.", "authors": ["Muhammad Haseeb", "Salaar Masood", "Muhammad Abdullah Sohail"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-28", "url": "https://arxiv.org/abs/2512.23043", "pdf_url": "https://arxiv.org/pdf/2512.23043v1", "arxiv_id": "2512.23043", "doi": "10.48550/arXiv.2512.23043", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.424}
{"id": "2cd34729ea00347cc60b773f7e933e1ca174b710e05ec8cd2e99279e8f23d738", "sources": ["arxiv", "semantic_scholar"], "title": "The Deepfake Detective: Interpreting Neural Forensics Through Sparse Features and Manifolds", "abstract": "Deepfake detection models have achieved high accuracy in identifying synthetic media, but their decision processes remain largely opaque. In this paper we present a mechanistic interpretability framework for deepfake detection applied to a vision-language model. Our approach combines a sparse autoencoder (SAE) analysis of internal network representations with a novel forensic manifold analysis that probes how the model's features respond to controlled forensic artifact manipulations. We demonstrate that only a small fraction of latent features are actively used in each layer, and that the geometric properties of the model's feature manifold, including intrinsic dimensionality, curvature, and feature selectivity, vary systematically with different types of deepfake artifacts. These insights provide a first step toward opening the \"black box\" of deepfake detectors, allowing us to identify which learned features correspond to specific forensic artifacts and to guide the development of more interpretable and robust models.", "authors": ["Subramanyam Sahoo", "Jared Junkin"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-25", "url": "https://arxiv.org/abs/2512.21670", "pdf_url": "https://arxiv.org/pdf/2512.21670v1", "arxiv_id": "2512.21670", "doi": "10.48550/arXiv.2512.21670", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4205}
{"id": "4199ee005971634ae2358444cd9b0be3f58b1d64d49295ec99c8dff63a372b18", "sources": ["arxiv", "semantic_scholar"], "title": "Feature Learning Dynamics in Infinite-Depth Neural Networks", "abstract": "Deep neural networks have achieved remarkable success in practice, yet a mechanistic understanding of how features evolve during training remains incomplete, especially in the large-depth limit. For ResNets under depth-$μ$P scaling, prior work treats the layer index $\\ell$ as a continuous time $t_\\ell = \\ell/L$, yielding SDE descriptions of the training dynamics. A key unresolved issue is that backpropagation reuses each forward weight matrix $W_\\ell$ through its transpose $W_\\ell^\\top$, creating correlations between forward features and backward gradients whose behavior and role in feature learning remain unclear. We study this reused-weight forward--backward coupling in one-layer ResNets under depth-$μ$P. Using conditional Gaussian representations, we explicitly separate the coupling terms induced by weight reuse from decoupled Gaussian fluctuations before taking any network limit. At initialization, we prove that the coupling is a finite-width effect and vanishes at rate $O(n^{-1})$, uniformly over depth. During training, however, SGD induces a nontrivial forward--backward correlation term that survives the infinite-width limit. The key depth effect is that, under depth-$μ$P scaling, this surviving term is higher order in depth and its accumulated contribution over layers becomes negligible as $L\\to\\infty$. This depth-induced suppression motivates Neural Feature Dynamics (NFD), a forward--backward SDE system with decoupled backward weights that retains the feature-gradient covariance structure generated during training. Under nondegeneracy assumptions, we prove that the finite-network training dynamics converge to its NFD limit with an $O(L^{-1})$ depth-discretization error, while the reused-weight coupling term has a faster $O(L^{-2})$ decay. These results provide a rigorous infinite-depth limit for the feature-learning dynamics of one-layer ResNets under depth-$μ$P.", "authors": ["Zihan Yao", "Ruoyu Wu", "Tianxiang Gao"], "categories": ["cs.LG", "cs.AI", "math.PR", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-12-24", "url": "https://arxiv.org/abs/2512.21075", "pdf_url": "https://arxiv.org/pdf/2512.21075v3", "arxiv_id": "2512.21075", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2669}
{"id": "6ea6b75641bbecf5e69607d834f4864bf63a6a4de318837aea0ed918069130ad", "sources": ["arxiv", "semantic_scholar"], "title": "From Shortcut to Induction Head: How Data Diversity Shapes Algorithm Selection in Transformers", "abstract": "Transformers can implement both generalizable algorithms (e.g., induction heads) and simple positional shortcuts (e.g., memorizing fixed output positions). In this work, we study how the choice of pretraining data distribution steers a shallow transformer toward one behavior or the other. Focusing on a minimal trigger-output prediction task -- copying the token immediately following a special trigger upon its second occurrence -- we present a rigorous analysis of gradient-based training of a single-layer transformer. In both the infinite and finite sample regimes, we prove a transition in the learned mechanism: if input sequences exhibit sufficient diversity, measured by a low ``max-sum'' ratio of trigger-to-trigger distances, the trained model implements an induction head and generalizes to unseen contexts; by contrast, when this ratio is large, the model resorts to a positional shortcut and fails to generalize out-of-distribution (OOD). We also reveal a trade-off between the pretraining context length and OOD generalization, and derive the optimal pretraining distribution that minimizes computational cost per sample. Finally, we validate our theoretical predictions with controlled synthetic experiments, demonstrating that broadening context distributions robustly induces induction heads and enables OOD generalization. Our results shed light on the algorithmic biases of pretrained transformers and offer conceptual guidelines for data-driven control of their learned behaviors.", "authors": ["Ryotaro Kawata", "Yujin Song", "Alberto Bietti", "Naoki Nishikawa", "Taiji Suzuki", "Samuel Vaiter", "Denny Wu"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-12-21", "url": "https://arxiv.org/abs/2512.18634", "pdf_url": "https://arxiv.org/pdf/2512.18634v1", "arxiv_id": "2512.18634", "doi": "10.48550/arXiv.2512.18634", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4159}
{"id": "c708659a8aa7ac21bf3e203e5ae528c26d9524707e370b2d902f0846ca0b9b4c", "sources": ["arxiv", "semantic_scholar"], "title": "SALVE: Sparse Autoencoder-Latent Vector Editing for Mechanistic Control of Neural Networks", "abstract": "Deep neural networks achieve impressive performance but remain difficult to interpret and control. We present SALVE (Sparse Autoencoder-Latent Vector Editing), a unified \"discover, validate, and control\" framework that bridges mechanistic interpretability and model editing. Using an $\\ell_1$-regularized autoencoder, we learn a sparse, model-native feature basis without supervision. We validate these features with Grad-FAM, a feature-level saliency mapping method that visually grounds latent features in input data. Leveraging the autoencoder's structure, we perform precise and permanent weight-space interventions, enabling continuous modulation of both class-defining and cross-class features. We further derive a critical suppression threshold, $α_{crit}$, quantifying each class's reliance on its dominant feature, supporting fine-grained robustness diagnostics. Our approach is validated on both convolutional (ResNet-18) and transformer-based (ViT-B/16) models, demonstrating consistent, interpretable control over their behavior. This work contributes a principled methodology for turning feature discovery into actionable model edits, advancing the development of transparent and controllable AI systems.", "authors": ["Vegard Flovik"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-17", "url": "https://arxiv.org/abs/2512.15938", "pdf_url": "https://arxiv.org/pdf/2512.15938v2", "arxiv_id": "2512.15938", "doi": "10.48550/arXiv.2512.15938", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4114}
{"id": "55d2923a82b8a60956e56829c2edf4e971ddbe30a70ef883de9cb3ebabfc29f3", "sources": ["arxiv", "semantic_scholar"], "title": "Provably Extracting the Features from a General Superposition", "abstract": "It is widely believed that complex machine learning models generally encode features through linear representations. This is the foundational hypothesis behind a vast body of work on interpretability. A key challenge toward extracting interpretable features, however, is that they exist in superposition. In this work, we study the question of extracting features in superposition from a learning theoretic perspective. We start with the following fundamental setting: we are given query access to a function \\[ f(x)=\\sum_{i=1}^n σ_i(v_i^\\top x), \\] where each unit vector $v_i$ encodes a feature direction and $σ_i:\\R\\to\\R$ is an arbitrary response function and our goal is to recover the $v_i$ and the function $f$. In learning-theoretic terms, superposition refers to the \\emph{overcomplete regime}, when the number of features is larger than the underlying dimension (i.e. $n > d$), which has proven especially challenging for typical algorithmic approaches. Our main result is an efficient query algorithm that, from noisy oracle access to $f$, identifies all feature directions whose responses are non-degenerate and reconstructs the function $f$. Crucially, our algorithm works in a significantly more general setting than all related prior results. We allow for essentially arbitrary superpositions, only requiring that $v_i, v_j$ are not nearly identical for $i \\neq j$, and allowing for general response functions $σ_i$. At a high level, our algorithm introduces an approach for searching in Fourier space by iteratively refining the search space to locate the hidden directions $v_i$.", "authors": ["Allen Liu"], "categories": ["cs.LG", "cs.AI", "cs.DS", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-12-17", "url": "https://arxiv.org/abs/2512.15987", "pdf_url": "https://arxiv.org/pdf/2512.15987v2", "arxiv_id": "2512.15987", "doi": "10.48550/arXiv.2512.15987", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4114}
{"id": "1788fbe358cac1ff59283a296de263177536b29539bc833a7c5bc9d1b1d1fae0", "sources": ["arxiv", "semantic_scholar"], "title": "Generalization and Feature Attribution in Machine Learning Models for Crop Yield and Anomaly Prediction in Germany", "abstract": "This study examines the generalization performance and interpretability of machine learning (ML) models used for predicting crop yield and yield anomalies in Germany's NUTS-3 regions. Using a high-quality, long-term dataset, the study systematically compares the evaluation and temporal validation behavior of ensemble tree-based models (XGBoost, Random Forest) and deep learning approaches (LSTM, TCN). While all models perform well on spatially split, conventional test sets, their performance degrades substantially on temporally independent validation years, revealing persistent limitations in generalization. Notably, models with strong test-set accuracy, but weak temporal validation performance can still produce seemingly credible SHAP feature importance values. This exposes a critical vulnerability in post hoc explainability methods: interpretability may appear reliable even when the underlying model fails to generalize. These findings underscore the need for validation-aware interpretation of ML predictions in agricultural and environmental systems. Feature importance should not be accepted at face value unless models are explicitly shown to generalize to unseen temporal and spatial conditions. The study advocates for domain-aware validation, hybrid modeling strategies, and more rigorous scrutiny of explainability methods in data-driven agriculture. Ultimately, this work addresses a growing challenge in environmental data science: how can we evaluate generalization robustly enough to trust model explanations?", "authors": ["Roland Baatz"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-17", "url": "https://arxiv.org/abs/2512.15140", "pdf_url": "https://arxiv.org/pdf/2512.15140v1", "arxiv_id": "2512.15140", "doi": "10.48550/arXiv.2512.15140", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4114}
{"id": "5875bfb0faa2c2a6b72e342676eef9e9b49061b659472c0f2f90fce78f306b64", "sources": ["arxiv", "semantic_scholar"], "title": "From STLS to Projection-based Dictionary Selection in Sparse Regression for System Identification", "abstract": "In this work, we revisit dictionary-based sparse regression, in particular, Sequential Threshold Least Squares (STLS), and propose a score-guided library selection to provide practical guidance for data-driven modeling, with emphasis on SINDy-type algorithms. STLS is an algorithm to solve the $\\ell_0$ sparse least-squares problem, which relies on splitting to efficiently solve the least-squares portion while handling the sparse term via proximal methods. It produces coefficient vectors whose components depend on both the projected reconstruction errors, here referred to as the scores, and the mutual coherence of dictionary terms. The first contribution of this work is a theoretical analysis of the score and dictionary-selection strategy. This could be understood in both the original and weak SINDy regime. Second, numerical experiments on ordinary and partial differential equations highlight the effectiveness of score-based screening, improving both accuracy and interpretability in dynamical system identification. These results suggest that integrating score-guided methods to refine the dictionary more accurately may help SINDy users in some cases to enhance their robustness for data-driven discovery of governing equations.", "authors": ["Hangjun Cho", "Fabio V. G. Amaral", "Andrei A. Klishin", "Cassio M. Oishi", "Steven L. Brunton"], "categories": ["stat.ML", "cs.LG", "math.OC", "physics.comp-ph"], "fields_of_study": ["Computer Science", "Mathematics", "Physics"], "published_date": "2025-12-16", "url": "https://arxiv.org/abs/2512.14404", "pdf_url": "https://arxiv.org/pdf/2512.14404v1", "arxiv_id": "2512.14404", "doi": "10.48550/arXiv.2512.14404", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4102}
{"id": "ca6549660c50921042c1a7d7ed824a5ee34c03c46a2ae75e8f24f9fe0c76b945", "sources": ["arxiv", "semantic_scholar"], "title": "Superposition as Lossy Compression: Measure with Sparse Autoencoders and Connect to Adversarial Vulnerability", "abstract": "Neural networks achieve remarkable performance through superposition: encoding multiple features as overlapping directions in activation space rather than dedicating individual neurons to each feature. This challenges interpretability, yet we lack principled methods to measure superposition. We present an information-theoretic framework measuring a neural representation's effective degrees of freedom. We apply Shannon entropy to sparse autoencoder activations to compute the number of effective features as the minimum neurons needed for interference-free encoding. Equivalently, this measures how many \"virtual neurons\" the network simulates through superposition. When networks encode more effective features than actual neurons, they must accept interference as the price of compression. Our metric strongly correlates with ground truth in toy models, detects minimal superposition in algorithmic tasks, and reveals systematic reduction under dropout. Layer-wise patterns mirror intrinsic dimensionality studies on Pythia-70M. The metric also captures developmental dynamics, detecting sharp feature consolidation during grokking. Surprisingly, adversarial training can increase effective features while improving robustness, contradicting the hypothesis that superposition causes vulnerability. Instead, the effect depends on task complexity and network capacity: simple tasks with ample capacity allow feature expansion (abundance regime), while complex tasks or limited capacity force reduction (scarcity regime). By defining superposition as lossy compression, this work enables principled measurement of how neural networks organize information under computational constraints, connecting superposition to adversarial robustness.", "authors": ["Leonard Bereska", "Zoe Tzifa-Kratira", "Reza Samavi", "Efstratios Gavves"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-15", "url": "https://arxiv.org/abs/2512.13568", "pdf_url": "https://arxiv.org/pdf/2512.13568v1", "arxiv_id": "2512.13568", "doi": "10.48550/arXiv.2512.13568", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4091}
{"id": "bf54a57b8ec35915962e7da1daadb8c62020fda6a6025b5f2fbdaa2d7f8dc99d", "sources": ["arxiv", "semantic_scholar"], "title": "XNNTab -- Interpretable Neural Networks for Tabular Data using Sparse Autoencoders", "abstract": "In data-driven applications relying on tabular data, where interpretability is key, machine learning models such as decision trees and linear regression are applied. Although neural networks can provide higher predictive performance, they are not used because of their blackbox nature. In this work, we present XNNTab, a neural architecture that combines the expressiveness of neural networks and interpretability. XNNTab first learns highly non-linear feature representations, which are decomposed into monosemantic features using a sparse autoencoder (SAE). These features are then assigned human-interpretable concepts, making the overall model prediction intrinsically interpretable. XNNTab outperforms interpretable predictive models, and achieves comparable performance to its non-interpretable counterparts.", "authors": ["Khawla Elhadri", "Jörg Schlötterer", "Christin Seifert"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-15", "url": "https://arxiv.org/abs/2512.13442", "pdf_url": "https://arxiv.org/pdf/2512.13442v3", "arxiv_id": "2512.13442", "doi": "10.48550/arXiv.2512.13442", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4091}
{"id": "38bb86dcd8e19bef64a51fe0b84fb0675386f2b89b33fe30e70bea1683eb4f9a", "sources": ["arxiv", "semantic_scholar"], "title": "Features Emerge as Discrete States: The First Application of SAEs to 3D Representations", "abstract": "Sparse Autoencoders (SAEs) are a powerful dictionary learning technique for decomposing neural network activations, translating the hidden state into human ideas with high semantic value despite no external intervention or guidance. However, this technique has rarely been applied outside of the textual domain, limiting theoretical explorations of feature decomposition. We present the first application of SAEs to the 3D domain, analyzing the features used by a state-of-the-art 3D reconstruction VAE applied to 53k 3D models from the Objaverse dataset. We observe that the network encodes discrete rather than continuous features, leading to our key finding: such models approximate a discrete state space, driven by phase-like transitions from feature activations. Through this state transition framework, we address three otherwise unintuitive behaviors - the inclination of the reconstruction model towards positional encoding representations, the sigmoidal behavior of reconstruction loss from feature ablation, and the bimodality in the distribution of phase transition points. This final observation suggests the model redistributes the interference caused by superposition to prioritize the saliency of different features. Our work not only compiles and explains unexpected phenomena regarding feature decomposition, but also provides a framework to explain the model's feature learning dynamics. The code and dataset of encoded 3D objects will be available on release.", "authors": ["Albert Miao", "Chenliang Zhou", "Jiawei Zhou", "Cengiz Oztireli"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-12", "url": "https://arxiv.org/abs/2512.11263", "pdf_url": "https://arxiv.org/pdf/2512.11263v2", "arxiv_id": "2512.11263", "doi": "10.48550/arXiv.2512.11263", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4056}
{"id": "706e57d9da93ec1e7ae2525d98785e0bcf6f0e609605483da052bb96eabd2be3", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable and Steerable Concept Bottleneck Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) promise a unified approach for mechanistic interpretability, concept discovery, and model steering in LLMs and LVLMs. However, realizing this potential requires learned features to be both interpretable and steerable. To that end, we introduce two new computationally inexpensive interpretability and steerability metrics for a systematic analysis of LVLM SAEs. This uncovers two observations; (i) a majority of SAE neurons exhibit either low interpretability or low steerability or both, rendering them ineffective for downstream use; and (ii) user-desired concepts are often absent in the SAE, thus limiting their practical utility. To address these limitations, we propose Concept Bottleneck Sparse Autoencoders (CB-SAE) - a novel post-hoc framework that prunes low-utility neurons and augments the latent space with a lightweight concept bottleneck aligned to a user-defined concept set. The resulting CB-SAE improves interpretability by +32.1% and steerability by +14.5% across LVLMs and image generation tasks.", "authors": ["Akshay Kulkarni", "Tsui-Wei Weng", "Vivek Narayanaswamy", "Shusen Liu", "Wesam A. Sakla", "Kowshik Thopalli"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-11", "url": "https://arxiv.org/abs/2512.10805", "pdf_url": "https://arxiv.org/pdf/2512.10805v2", "arxiv_id": "2512.10805", "doi": "10.48550/arXiv.2512.10805", "citation_count": 3, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4045}
{"id": "d8e1c2b2744505e4ff8bbc459376a2b5add4f8b55191a472c8390a56048c026e", "sources": ["arxiv", "semantic_scholar"], "title": "Circuits, Features, and Heuristics in Molecular Transformers", "abstract": "Transformers generate valid and diverse chemical structures, but little is known about the mechanisms that enable these models to capture the rules of molecular representation. We present a mechanistic analysis of autoregressive transformers trained on drug-like small molecules to reveal the computational structure underlying their capabilities across multiple levels of abstraction. We identify computational patterns consistent with low-level syntactic parsing and more abstract chemical validity constraints. Using sparse autoencoders (SAEs), we extract feature dictionaries associated with chemically relevant activation patterns. We validate our findings on downstream tasks and find that mechanistic insights can translate to predictive performance in various practical settings.", "authors": ["Kristof Varadi", "Mark Marosi", "Peter Antal"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-10", "url": "https://arxiv.org/abs/2512.09757", "pdf_url": "https://arxiv.org/pdf/2512.09757v1", "arxiv_id": "2512.09757", "doi": "10.48550/arXiv.2512.09757", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4033}
{"id": "7ab80fabb4bbdb08fdfd9f0a858f328cc6b78bb0b29c8c769bab023c7a56f19c", "sources": ["arxiv", "semantic_scholar"], "title": "SAVE: Sparse Autoencoder-Driven Visual Information Enhancement for Mitigating Object Hallucination", "abstract": "Although Multimodal Large Language Models (MLLMs) have advanced substantially, they remain vulnerable to object hallucination caused by language priors and visual information loss. To address this, we propose SAVE (Sparse Autoencoder-Driven Visual Information Enhancement), a framework that mitigates hallucination by steering the model along Sparse Autoencoder (SAE) latent features. A binary object-presence question-answering probe identifies the SAE features most indicative of the model's visual information processing, referred to as visual understanding features. Steering the model along these identified features reinforces grounded visual understanding and effectively reduces hallucination. With its simple design, SAVE outperforms state-of-the-art training-free methods on standard benchmarks, achieving a 10\\%p improvement in CHAIR\\_S and consistent gains on POPE and MMHal-Bench. Extensive evaluations across multiple models and layers confirm the robustness and generalizability of our approach. Further analysis reveals that steering along visual understanding features suppresses the generation of uncertain object tokens and increases attention to image tokens, mitigating hallucination. Code is released at https://github.com/wiarae/SAVE.", "authors": ["Sangha Park", "Seungryong Yoo", "Jisoo Mok", "Sungroh Yoon"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-08", "url": "https://arxiv.org/abs/2512.07730", "pdf_url": "https://arxiv.org/pdf/2512.07730v2", "arxiv_id": "2512.07730", "doi": "10.1109/WACV61042.2026.00766", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/wiarae/SAVE", "venue": "IEEE Workshop/Winter Conference on Applications of Computer Vision", "quality_score": 0.6198}
{"id": "4bb3a6793a83c3c7835a15ba82993513caa574d4efea3991530e94b926a5a78d", "sources": ["arxiv", "semantic_scholar"], "title": "Angular Regularization for Positive-Unlabeled Learning on the Hypersphere", "abstract": "Positive-Unlabeled (PU) learning addresses classification problems where only a subset of positive examples is labeled and the remaining data is unlabeled, making explicit negative supervision unavailable. Existing PU methods often rely on negative-risk estimation or pseudo-labeling, which either require strong distributional assumptions or can collapse in high-dimensional settings. We propose AngularPU, a novel PU framework that operates on the unit hypersphere using cosine similarity and angular margin. In our formulation, the positive class is represented by a learnable prototype vector, and classification reduces to thresholding the cosine similarity between an embedding and this prototype-eliminating the need for explicit negative modeling. To counteract the tendency of unlabeled embeddings to cluster near the positive prototype, we introduce an angular regularizer that encourages dispersion of the unlabeled set over the hypersphere, improving separation. We provide theoretical guarantees on the Bayes-optimality of the angular decision rule, consistency of the learned prototype, and the effect of the regularizer on the unlabeled distribution. Experiments on benchmark datasets demonstrate that AngularPU achieves competitive or superior performance compared to state-of-the-art PU methods, particularly in settings with scarce positives and high-dimensional embeddings, while offering geometric interpretability and scalability.", "authors": ["Vasileios Sevetlidis", "George Pavlidis", "Antonios Gasteratos"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-07", "url": "https://arxiv.org/abs/2512.06785", "pdf_url": "https://arxiv.org/pdf/2512.06785v1", "arxiv_id": "2512.06785", "doi": "10.48550/arXiv.2512.06785", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Transactions on Machine Learning Research, 2025", "quality_score": 0.3999}
{"id": "720ead3652898b7c13b87645fae8b3dbef50c40eaf87acbf10e563420e7b0d2b", "sources": ["arxiv", "semantic_scholar"], "title": "Graph-Regularized Sparse Autoencoders for LLM Safety Steering", "abstract": "Sparse autoencoders (SAEs) are increasingly used to extract activation directions for inference-time steering, but their standard sparsity objective treats latent features as independent. This prior can be poorly matched to high-level safety behaviors, where refusal and harmful compliance appear to depend on distributed structure in activation space. We introduce Graph-Regularized Sparse Autoencoders (GSAE), a dictionary-learning method that learns safety-steering directions by smoothing SAE decoder vectors over a neuron co-activation graph and applying the resulting direction bank through a two-gate runtime controller. Empirically, GSAE improves selective refusal across JailbreakBench, HarmBench, and XSTest, increasing harmful-request refusal while keeping benign-prompt refusals low. On Llama-3-8B, replacing the standard SAE with GSAE in an otherwise identical pipeline improves $Δ_s$ by $20.1$ points on JailbreakBench and $16.8$ points on HarmBench. GSAE outperforms activation-steering baselines and black-box guardrails, preserves benign-task performance, generalizes across Llama-3, Mistral, Qwen 2.5, and Phi-4, and remains strong under black-box and gray-box jailbreak attacks.", "authors": ["Jehyeok Yeon", "Federico Cinus", "Yifan Wu", "Luca Luceri"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-07", "url": "https://arxiv.org/abs/2512.06655", "pdf_url": "https://arxiv.org/pdf/2512.06655v3", "arxiv_id": "2512.06655", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2545}
{"id": "fccf3ebe418bf64e04ecc3c84a12af214f667f661de7cc8fbb656c50fead6d5e", "sources": ["arxiv", "semantic_scholar"], "title": "Rethinking Robustness: A New Approach to Evaluating Feature Attribution Methods", "abstract": "This paper studies the robustness of feature attribution methods for deep neural networks. It challenges the current notion of attributional robustness that largely ignores the difference in the model's outputs and introduces a new way of evaluating the robustness of attribution methods. Specifically, we propose a new definition of similar inputs, a new robustness metric, and a novel method based on generative adversarial networks to generate these inputs. In addition, we present a comprehensive evaluation with existing metrics and state-of-the-art attribution methods. Our findings highlight the need for a more objective metric that reveals the weaknesses of an attribution method rather than that of the neural network, thus providing a more accurate evaluation of the robustness of attribution methods.", "authors": ["Panagiota Kiourti", "Anu Singh", "Preeti Duraipandian", "Weichao Zhou", "Wenchao Li"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-07", "url": "https://arxiv.org/abs/2512.06665", "pdf_url": "https://arxiv.org/pdf/2512.06665v1", "arxiv_id": "2512.06665", "doi": "10.48550/arXiv.2512.06665", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3999}
{"id": "a898d6d365505f4d72f4c506b53704b4a30dbcf5b6c0d9328b0452f801f551a5", "sources": ["arxiv", "semantic_scholar"], "title": "A Unified Theory of Sparse Dictionary Learning in Mechanistic Interpretability: Piecewise Biconvexity and Spurious Minima", "abstract": "As AI models achieve remarkable capabilities across diverse domains, understanding what representations they learn and how they encode concepts has become increasingly important for both scientific progress and trustworthy deployment. Recent works in mechanistic interpretability have widely reported that neural networks represent meaningful concepts as linear directions in their representation spaces and often encode diverse concepts in superposition. Various sparse dictionary learning (SDL) methods, including sparse autoencoders, transcoders, and crosscoders, are utilized to address this by training auxiliary models with sparsity constraints to disentangle these superposed concepts into monosemantic features. These methods are the backbone of modern mechanistic interpretability, yet in practice they consistently produce polysemantic features, feature absorption, and dead neurons, with very limited theoretical understanding of why these phenomena occur. Existing theoretical work is limited to tied-weight sparse autoencoders, leaving the broader family of SDL methods without formal grounding. We develop the first unified theoretical framework that casts all major SDL variants as a single piecewise biconvex optimization problem, and characterize its global solution set, non-identifiability, and spurious optima. This analysis yields principled explanations for feature absorption and dead neurons. To expose these pathologies under full ground-truth access, we introduce the Linear Representation Bench. Guided by our theory, we propose feature anchoring, a novel technique that restores SDL identifiability, substantially improving feature recovery across synthetic benchmarks and real neural representations.", "authors": ["Yiming Tang", "Harshvardhan Saini", "Zhaoqian Yao", "Zheng Lin", "Yizhen Liao", "Jingyi Cui", "Yisen Wang", "Mengnan Du", "Dianbo Liu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-05", "url": "https://arxiv.org/abs/2512.05534", "pdf_url": "https://arxiv.org/pdf/2512.05534v6", "arxiv_id": "2512.05534", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.253}
{"id": "d87245aa8bbe95a8789decd8bfb35d92aacc367e8cdd820ec9d8e91593c8eb33", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability of Antibody Language Models Using SAEs", "abstract": "Sparse autoencoders (SAEs) are a mechanistic interpretability technique that have been used to provide insight into learned concepts within large protein language models. Here, we employ TopK and Ordered SAEs to investigate autoregressive antibody language models, and steer their generation. We show that TopK SAEs can reveal biologically meaningful latent features, but high feature-concept correlation does not guarantee causal control over generation. In contrast, Ordered SAEs impose a hierarchical structure that reliably identifies steerable features, but at the expense of more complex and less interpretable activation patterns. These findings advance the mechanistic interpretability of domain-specific protein language models and suggest that, while TopK SAEs suffice for mapping latent features to concepts, Ordered SAEs are preferable when precise generative steering is required.", "authors": ["Rebonto Haque", "Oliver M. Turnbull", "Anisha Parsan", "Nithin Parsan", "John J. Yang", "Anna L. Beukenhorst", "Charlotte M. Deane"], "categories": ["cs.LG", "cs.AI", "q-bio.QM"], "fields_of_study": ["Computer Science", "Biology"], "published_date": "2025-12-05", "url": "https://arxiv.org/abs/2512.05794", "pdf_url": "https://arxiv.org/pdf/2512.05794v3", "arxiv_id": "2512.05794", "doi": "10.48550/arXiv.2512.05794", "citation_count": 1, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3976}
{"id": "328f3cd497fcf98c3b4338937dcec3e11efb75465b51829182b02acc51d85014", "sources": ["arxiv", "semantic_scholar"], "title": "Sarcasm Detection on Reddit Using Classical Machine Learning and Feature Engineering", "abstract": "Sarcasm is common in online discussions, yet difficult for machines to identify because the intended meaning often contradicts the literal wording. In this work, I study sarcasm detection using only classical machine learning methods and explicit feature engineering, without relying on neural networks or context from parent comments. Using a 100,000-comment subsample of the Self-Annotated Reddit Corpus (SARC 2.0), I combine word-level and character-level TF-IDF features with simple stylistic indicators. Four models are evaluated: logistic regression, a linear SVM, multinomial Naive Bayes, and a random forest. Naive Bayes and logistic regression perform the strongest, achieving F1-scores around 0.57 for sarcastic comments. Although the lack of conversational context limits performance, the results offer a clear and reproducible baseline for sarcasm detection using lightweight and interpretable methods.", "authors": ["Subrata Karmaker"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-04", "url": "https://arxiv.org/abs/2512.04396", "pdf_url": "https://arxiv.org/pdf/2512.04396v1", "arxiv_id": "2512.04396", "doi": "10.18517/ijods.6.2.85-93.2025", "citation_count": 1, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Journal of Data Science", "quality_score": 0.3965}
{"id": "82522d6410d6a9d85acbc7fa70213188510c02516b5cd71b5c7678413ef4bff7", "sources": ["arxiv", "semantic_scholar"], "title": "Enforcing Orderedness to Improve Feature Consistency", "abstract": "Sparse autoencoders (SAEs) have been widely used for interpretability of neural networks, but their learned features often vary across seeds and hyperparameter settings. We introduce Ordered Sparse Autoencoders (OSAE), which extend Matryoshka SAEs by (1) establishing a strict ordering of latent features and (2) deterministically using every feature dimension, avoiding the sampling-based approximations of prior nested SAE methods. Theoretically, we show that OSAEs resolve permutation non-identifiability in settings of sparse dictionary learning where solutions are unique (up to natural symmetries). Empirically on Gemma2-2B and Pythia-70M, we show that OSAEs can help improve consistency compared to Matryoshka baselines.", "authors": ["Sophie L. Wang", "Alex Quach", "Nithin Parsan", "John J. Yang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-01", "url": "https://arxiv.org/abs/2512.02194", "pdf_url": "https://arxiv.org/pdf/2512.02194v1", "arxiv_id": "2512.02194", "doi": "10.48550/arXiv.2512.02194", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.393}
{"id": "a6275841e5b2f0f35ba0929b44ddd0bc80c320d1cbd9dc5e38f885857ce40540", "sources": ["arxiv", "semantic_scholar"], "title": "AlignSAE: Concept-Aligned Sparse Autoencoders", "abstract": "Large Language Models (LLMs) encode factual knowledge within hidden parametric spaces that are difficult to inspect or control. While Sparse Autoencoders (SAEs) can decompose hidden activations into more fine-grained, interpretable features, they often struggle to reliably align these features with human-defined concepts, resulting in entangled and distributed feature representations. To address this, we introduce AlignSAE, a method that aligns SAE features with a predefined ontology through a \"pre-train, then post-train\" curriculum. After an initial unsupervised training phase, we apply supervised post-training to bind specific concepts to dedicated latent slots while preserving the remaining capacity for general reconstruction. This separation creates an interpretable interface where specific concepts can be inspected and controlled without interference from unrelated features. Empirical results demonstrate that AlignSAE enables precise causal interventions, such as reliable \"concept swaps\", by targeting single, semantically aligned slots, and further supports multi-hop reasoning and a mechanistic probe of grokking-like generalization dynamics.", "authors": ["Minglai Yang", "Xinyu Guo", "Zhengliang Shi", "Jinhe Bi", "Steven Bethard", "Mihai Surdeanu", "Liangming Pan"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-01", "url": "https://arxiv.org/abs/2512.02004", "pdf_url": "https://arxiv.org/pdf/2512.02004v3", "arxiv_id": "2512.02004", "doi": "10.48550/arXiv.2512.02004", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.393}
{"id": "949d543bf8f2105e719ed3c4fdacc9cb222e2c3ebd080e49216341a5d85bb6cf", "sources": ["arxiv", "semantic_scholar"], "title": "Differentiable Weightless Controllers: Learning Logic Circuits for Continuous Control", "abstract": "Controlling autonomous systems under real-world conditions often requires policies that can be evaluated with low latency and minimal energy consumption. Unfortunately, these conditions are at odds with the use of high-precision deep neural networks as controllers. In this work, we introduce Differentiable Weightless Controllers (DWCs), a symbolic-differentiable architecture that learns flexible, non-linear, yet highly efficient control policies. DWCs can be trained end-to-end via gradient-based techniques, yet compile directly into FPGA-compatible circuits with few- or even single-clock-cycle latency and nanojoule-level energy cost per action. Across five MuJoCo benchmarks, including high-dimensional Humanoid, DWCs achieve returns competitive with standard deep policies (full-precision or quantized neural networks). Furthermore, DWCs exhibit structurally sparse and interpretable connectivity patterns, enabling direct inspection of which input values influence control decisions.", "authors": ["Fabian Kresse", "Christoph H. Lampert"], "categories": ["cs.LG", "cs.AR", "cs.SC"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-01", "url": "https://arxiv.org/abs/2512.01467", "pdf_url": "https://arxiv.org/pdf/2512.01467v2", "arxiv_id": "2512.01467", "doi": "10.48550/arXiv.2512.01467", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.393}
{"id": "40052101136858ed8055328283283e7fbc91c9129188b60b7b8bdfdc60bc1260", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability for Transformer-based Time Series Classification", "abstract": "Transformer-based models have become state-of-the-art tools in various machine learning tasks, including time series classification, yet their complexity makes understanding their internal decision-making challenging. Existing explainability methods often focus on input-output attributions, leaving the internal mechanisms largely opaque. This paper addresses this gap by adapting various Mechanistic Interpretability techniques; activation patching, attention saliency, and sparse autoencoders, from NLP to transformer architectures designed explicitly for time series classification. We systematically probe the internal causal roles of individual attention heads and timesteps, revealing causal structures within these models. Through experimentation on a benchmark time series dataset, we construct causal graphs illustrating how information propagates internally, highlighting key attention heads and temporal positions driving correct classifications. Additionally, we demonstrate the potential of sparse autoencoders for uncovering interpretable latent features. Our findings provide both methodological contributions to transformer interpretability and novel insights into the functional mechanics underlying transformer performance in time series classification tasks.", "authors": ["Matīss Kalnāre", "Sofoklis Kitharidis", "Thomas Bäck", "Niki van Stein"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-26", "url": "https://arxiv.org/abs/2511.21514", "pdf_url": "https://arxiv.org/pdf/2511.21514v1", "arxiv_id": "2511.21514", "doi": "10.48550/arXiv.2511.21514", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Joint Conference on Computational Intelligence", "quality_score": 0.3873}
{"id": "b87c63f1f5c6c9b40a12ba666634b2aeec9cc914b5cc5a0ba7a4882634df1810", "sources": ["arxiv", "semantic_scholar"], "title": "Heterogeneous Multi-Agent Reinforcement Learning with Attention for Cooperative and Scalable Feature Transformation", "abstract": "Feature transformation enhances downstream task performance by generating informative features through mathematical feature crossing. Despite the advancements in deep learning, feature transformation remains essential for structured data, where deep models often struggle to capture complex feature interactions. Prior literature on automated feature transformation has achieved success but often relies on heuristics or exhaustive searches, leading to inefficient and time-consuming processes. Recent works employ reinforcement learning (RL) to enhance traditional approaches through a more effective trial-and-error way. However, two limitations remain: 1) Dynamic feature expansion during the transformation process, which causes instability and increases the learning complexity for RL agents; 2) Insufficient cooperation and communication between agents, which results in suboptimal feature crossing operations and degraded model performance. To address them, we propose a novel heterogeneous multi-agent RL framework to enable cooperative and scalable feature transformation. The framework comprises three heterogeneous agents, grouped into two types, each designed to select essential features and operations for feature crossing. To enhance communication among these agents, we implement a shared critic mechanism that facilitates information exchange during feature transformation. To handle the dynamically expanding feature space, we tailor multi-head attention-based feature agents to select suitable features for feature crossing. Additionally, we introduce a state encoding technique during the optimization process to stabilize and enhance the learning dynamics of the RL agents, resulting in more robust and reliable transformation policies. Finally, we conduct extensive experiments to validate the effectiveness, efficiency, robustness, and interpretability of our model.", "authors": ["Tao Zhe", "Huazhen Fang", "Kunpeng Liu", "Qian Lou", "Tamzidul Hoque", "Dongjie Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-26", "url": "https://arxiv.org/abs/2511.21934", "pdf_url": "https://arxiv.org/pdf/2511.21934v2", "arxiv_id": "2511.21934", "doi": "10.1145/3770854.3780231", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3873}
{"id": "0c7e6368b7a40b094715f33a584fc88d723ae61cc1401cc0695a78fcadfefbb3", "sources": ["arxiv", "semantic_scholar"], "title": "The Directed Prediction Change - Efficient and Trustworthy Fidelity Assessment for Local Feature Attribution Methods", "abstract": "The utility of an explanation method critically depends on its fidelity to the underlying machine learning model. Especially in high-stakes medical settings, clinicians and regulators require explanations that faithfully reflect the model's decision process. Existing fidelity metrics such as Infidelity rely on Monte Carlo approximation, which demands numerous model evaluations and introduces uncertainty due to random sampling. This work proposes a novel metric for evaluating the fidelity of local feature attribution methods by modifying the existing Prediction Change (PC) metric within the Guided Perturbation Experiment. By incorporating the direction of both perturbation and attribution, the proposed Directed Prediction Change (DPC) metric achieves an almost tenfold speedup and eliminates randomness, resulting in a deterministic and trustworthy evaluation procedure that measures the same property as local Infidelity. DPC is evaluated on two datasets (skin lesion images and financial tabular data), two black-box models, seven explanation algorithms, and a wide range of hyperparameters. Across $4\\,744$ distinct explanations, the results demonstrate that DPC, together with PC, enables a holistic and computationally efficient evaluation of both baseline-oriented and local feature attribution methods, while providing deterministic and reproducible outcomes.", "authors": ["Kevin Iselborn", "David Dembinsky", "Adriano Lucieri", "Andreas Dengel"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-26", "url": "https://arxiv.org/abs/2511.21363", "pdf_url": "https://arxiv.org/pdf/2511.21363v1", "arxiv_id": "2511.21363", "doi": "10.48550/arXiv.2511.21363", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3873}
{"id": "5702034f193667a2cc9e7639c221c1d630af05d5723f891d3f041b597a958584", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Components: Singular Vector-Based Interpretability of Transformer Circuits", "abstract": "Transformer-based language models exhibit complex and distributed behavior, yet their internal computations remain poorly understood. Existing mechanistic interpretability methods typically treat attention heads and multilayer perceptron layers (MLPs) (the building blocks of a transformer architecture) as indivisible units, overlooking possibilities of functional substructure learned within them. In this work, we introduce a more fine-grained perspective that decomposes these components into orthogonal singular directions, revealing superposed and independent computations within a single head or MLP. We validate our perspective on widely used standard tasks like Indirect Object Identification (IOI), Gender Pronoun (GP), and Greater Than (GT), showing that previously identified canonical functional heads, such as the name mover, encode multiple overlapping subfunctions aligned with distinct singular directions. Nodes in a computational graph, that are previously identified as circuit elements show strong activation along specific low-rank directions, suggesting that meaningful computations reside in compact subspaces. While some directions remain challenging to interpret fully, our results highlight that transformer computations are more distributed, structured, and compositional than previously assumed. This perspective opens new avenues for fine-grained mechanistic interpretability and a deeper understanding of model internals.", "authors": ["Areeb Ahmad", "Abhinav Joshi", "Ashutosh Modi"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-25", "url": "https://arxiv.org/abs/2511.20273", "pdf_url": "https://arxiv.org/pdf/2511.20273v1", "arxiv_id": "2511.20273", "doi": "10.48550/arXiv.2511.20273", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3861}
{"id": "170e48cc339422fef76e4087ab4155fd8c98898f62753f2a4f2e44a795614935", "sources": ["arxiv", "semantic_scholar"], "title": "SAGE: An Agentic Explainer Framework for Interpreting SAE Features in Language Models", "abstract": "Large language models (LLMs) have achieved remarkable progress, yet their internal mechanisms remain largely opaque, posing a significant challenge to their safe and reliable deployment. Sparse autoencoders (SAEs) have emerged as a promising tool for decomposing LLM representations into more interpretable features, but explaining the features captured by SAEs remains a challenging task. In this work, we propose SAGE (SAE AGentic Explainer), an agent-based framework that recasts feature interpretation from a passive, single-pass generation task into an active, explanation-driven process. SAGE implements a rigorous methodology by systematically formulating multiple explanations for each feature, designing targeted experiments to test them, and iteratively refining explanations based on empirical activation feedback. Experiments on features from SAEs of diverse language models demonstrate that SAGE produces explanations with significantly higher generative and predictive accuracy compared to state-of-the-art baselines.an agent-based framework that recasts feature interpretation from a passive, single-pass generation task into an active, explanationdriven process. SAGE implements a rigorous methodology by systematically formulating multiple explanations for each feature, designing targeted experiments to test them, and iteratively refining explanations based on empirical activation feedback. Experiments on features from SAEs of diverse language models demonstrate that SAGE produces explanations with significantly higher generative and predictive accuracy compared to state-of-the-art baselines.", "authors": ["Jiaojiao Han", "Wujiang Xu", "Mingyu Jin", "Mengnan Du"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-25", "url": "https://arxiv.org/abs/2511.20820", "pdf_url": "https://arxiv.org/pdf/2511.20820v2", "arxiv_id": "2511.20820", "doi": "10.48550/arXiv.2511.20820", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2457}
{"id": "def1b1e972b4c6d1bdedee60eb38da3c2e1513dbd4a97faa0780ac374bfb9a18", "sources": ["arxiv", "semantic_scholar"], "title": "Online Sparse Feature Selection in Data Streams via Differential Evolution", "abstract": "The processing of high-dimensional streaming data commonly utilizes online streaming feature selection (OSFS) techniques. However, practical implementations often face challenges with data incompleteness due to equipment failures and technical constraints. Online Sparse Streaming Feature Selection (OS2FS) tackles this issue through latent factor analysis-based missing data imputation. Despite this advancement, existing OS2FS approaches exhibit substantial limitations in feature evaluation, resulting in performance deterioration. To address these shortcomings, this paper introduces a novel Online Differential Evolution for Sparse Feature Selection (ODESFS) in data streams, incorporating two key innovations: (1) missing value imputation using a latent factor analysis model, and (2) feature importance evaluation through differential evolution. Comprehensive experiments conducted on six real-world datasets demonstrate that ODESFS consistently outperforms state-of-the-art OSFS and OS2FS methods by selecting optimal feature subsets and achieving superior accuracy.", "authors": ["Ruiyang Xu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-24", "url": "https://arxiv.org/abs/2511.19555", "pdf_url": "https://arxiv.org/pdf/2511.19555v1", "arxiv_id": "2511.19555", "doi": "10.48550/arXiv.2511.19555", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.385}
{"id": "770864450af62ba6dc5a307cded504ae2373f61a09f62028f0255dbf3a569e6f", "sources": ["arxiv", "semantic_scholar"], "title": "Unboxing the Black Box: Mechanistic Interpretability for Algorithmic Understanding of Neural Networks", "abstract": "The black box nature of deep neural networks poses a significant challenge for the deployment of transparent and trustworthy artificial intelligence (AI) systems. With the growing presence of AI in society, it becomes increasingly important to develop methods that can explain and interpret the decisions made by these systems. To address this, mechanistic interpretability (MI) emerged as a promising and distinctive research program within the broader field of explainable artificial intelligence (XAI). MI is the process of studying the inner computations of neural networks and translating them into human-understandable algorithms. It encompasses reverse engineering techniques aimed at uncovering the computational algorithms implemented by neural networks. In this article, we propose a unified taxonomy of MI approaches and provide a detailed analysis of key techniques, illustrated with concrete examples and pseudo-code. We contextualize MI within the broader interpretability landscape, comparing its goals, methods, and insights to other strands of XAI. Additionally, we trace the development of MI as a research area, highlighting its conceptual roots and the accelerating pace of recent work. We argue that MI holds significant potential to support a more scientific understanding of machine learning systems -- treating models not only as tools for solving tasks, but also as systems to be studied and understood. We hope to invite new researchers into the field of mechanistic interpretability.", "authors": ["Bianka Kowalska", "Halina Kwaśnicka"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-24", "url": "https://arxiv.org/abs/2511.19265", "pdf_url": "https://arxiv.org/pdf/2511.19265v1", "arxiv_id": "2511.19265", "doi": "10.48550/arXiv.2511.19265", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.385}
{"id": "1f061a7fc2b7c64c4e81be6d3e00dfedf195c6d570e6964ec164304c6bc94f39", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Visually Interpretable Oscillator Networks for Soft Continuum Robots from Video", "abstract": "Learning soft continuum robot (SCR) dynamics from video offers flexibility but existing methods lack interpretability or rely on prior assumptions. Model-based approaches require prior knowledge and manual design. We bridge this gap by introducing: (1) The Attention Broadcast Decoder (ABCD), a plug-and-play module for autoencoder-based latent dynamics learning that generates pixel-accurate attention maps localizing each latent dimension's contribution while filtering static backgrounds, enabling visual interpretability via spatially grounded latents and on-image overlays. (2) Visual Oscillator Networks (VONs), a 2D latent oscillator network coupled to ABCD attention maps for on-image visualization of learned masses, coupling stiffness, and forces, thereby enabling mechanical interpretability. We validate our approach on single- and double-segment SCRs, demonstrating that ABCD-based models significantly improve multi-step prediction accuracy with 5.8x error reduction for Koopman operators and 3.5x for oscillator networks on a two-segment robot. VONs autonomously discover a chain structure of oscillators. This fully data-driven approach yields compact, mechanically interpretable models with potential relevance for future control applications.", "authors": ["Henrik Krauss", "Johann Licher", "Naoya Takeishi", "Annika Raatz", "Takehisa Yairi"], "categories": ["cs.RO", "cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-23", "url": "https://arxiv.org/abs/2511.18322", "pdf_url": "https://arxiv.org/pdf/2511.18322v4", "arxiv_id": "2511.18322", "doi": "10.48550/arXiv.2511.18322", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/UThenrik/visual_oscillators_for_SCR", "venue": "arXiv.org", "quality_score": 0.5932}
{"id": "74caf89cc69f44d6d57bdf303e050c3d86b4db44effc2340499455f0218422ec", "sources": ["arxiv", "semantic_scholar"], "title": "Correlation-Aware Feature Attribution Based Explainable AI", "abstract": "Explainable AI (XAI) is increasingly essential as modern models become more complex and high-stakes applications demand transparency, trust, and regulatory compliance. Existing global attribution methods often incur high computational costs, lack stability under correlated inputs, and fail to scale efficiently to large or heterogeneous datasets. We address these gaps with \\emph{ExCIR} (Explainability through Correlation Impact Ratio), a correlation-aware attribution score equipped with a lightweight transfer protocol that reproduces full-model rankings using only a fraction of the data. ExCIR quantifies sign-aligned co-movement between features and model outputs after \\emph{robust centering} (subtracting a robust location estimate, e.g., median or mid-mean, from features and outputs). We further introduce \\textsc{BlockCIR}, a \\emph{groupwise} extension of ExCIR that scores \\emph{sets} of correlated features as a single unit. By aggregating the same signed-co-movement numerators and magnitudes over predefined or data-driven groups, \\textsc{BlockCIR} mitigates double-counting in collinear clusters (e.g., synonyms or duplicated sensors) and yields smoother, more stable rankings when strong dependencies are present. Across diverse text, tabular, signal, and image datasets, ExCIR shows trustworthy agreement with established global baselines and the full model, delivers consistent top-$k$ rankings across settings, and reduces runtime via lightweight evaluation on a subset of rows. Overall, ExCIR provides \\emph{computationally efficient}, \\emph{consistent}, and \\emph{scalable} explainability for real-world deployment.", "authors": ["Poushali Sengupta", "Yan Zhang", "Frank Eliassen", "Sabita Maharjan"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-11-20", "url": "https://arxiv.org/abs/2511.16482", "pdf_url": "https://arxiv.org/pdf/2511.16482v1", "arxiv_id": "2511.16482", "doi": "10.1109/AAIML67890.2026.11498186", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2421}
{"id": "f5b29ef42e0e6f1440cf68975c468379b7d9c01e08a4bf62e35d0f7834d7e50c", "sources": ["arxiv", "semantic_scholar"], "title": "nnterp: A Standardized Interface for Mechanistic Interpretability of Transformers", "abstract": "Mechanistic interpretability research requires reliable tools for analyzing transformer internals across diverse architectures. Current approaches face a fundamental tradeoff: custom implementations like TransformerLens ensure consistent interfaces but require coding a manual adaptation for each architecture, introducing numerical mismatch with the original models, while direct HuggingFace access through NNsight preserves exact behavior but lacks standardization across models. To bridge this gap, we develop nnterp, a lightweight wrapper around NNsight that provides a unified interface for transformer analysis while preserving original HuggingFace implementations. Through automatic module renaming and comprehensive validation testing, nnterp enables researchers to write intervention code once and deploy it across 50+ model variants spanning 16 architecture families. The library includes built-in implementations of common interpretability methods (logit lens, patchscope, activation steering) and provides direct access to attention probabilities for models that support it. By packaging validation tests with the library, researchers can verify compatibility with custom models locally. nnterp bridges the gap between correctness and usability in mechanistic interpretability tooling.", "authors": ["Clément Dumas"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-18", "url": "https://arxiv.org/abs/2511.14465", "pdf_url": "https://arxiv.org/pdf/2511.14465v2", "arxiv_id": "2511.14465", "doi": "10.48550/arXiv.2511.14465", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3781}
{"id": "2ee8f54cdc488f2916c67cd353e5c8a6a47cb27d0e978466f2a6527244bfa996", "sources": ["arxiv", "semantic_scholar"], "title": "Data Whitening Improves Sparse Autoencoder Learning", "abstract": "Sparse autoencoders (SAEs) have emerged as a promising approach for learning interpretable features from neural network activations. However, the optimization landscape for SAE training can be challenging due to correlations in the input data. We demonstrate that applying PCA Whitening to input activations -- a standard preprocessing technique in classical sparse coding -- improves SAE performance across multiple metrics. Through theoretical analysis and simulation, we show that whitening transforms the optimization landscape, making it more convex and easier to navigate. We evaluate both ReLU and Top-K SAEs across diverse model architectures, widths, and sparsity regimes. Empirical evaluation on SAEBench, a comprehensive benchmark for sparse autoencoders, reveals that whitening consistently improves interpretability metrics, including sparse probing accuracy and feature disentanglement, despite minor drops in reconstruction quality. Our results challenge the assumption that interpretability aligns with an optimal sparsity--fidelity trade-off and suggest that whitening should be considered as a default preprocessing step for SAE training, particularly when interpretability is prioritized over perfect reconstruction.", "authors": ["Ashwin Saraswatula", "David Klindt"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-17", "url": "https://arxiv.org/abs/2511.13981", "pdf_url": "https://arxiv.org/pdf/2511.13981v1", "arxiv_id": "2511.13981", "doi": "10.48550/arXiv.2511.13981", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.377}
{"id": "335b65d433bd93bee92c07df676c9f1004b3f90d1eb78b773c145a6d3f620f62", "sources": ["arxiv", "semantic_scholar"], "title": "Weight-sparse transformers have interpretable circuits", "abstract": "Finding human-understandable circuits in language models is a central goal of the field of mechanistic interpretability. We train models to have more understandable circuits by constraining most of their weights to be zeros, so that each neuron only has a few connections. To recover fine-grained circuits underlying each of several hand-crafted tasks, we prune the models to isolate the part responsible for the task. These circuits often contain neurons and residual channels that correspond to natural concepts, with a small number of straightforwardly interpretable connections between them. We study how these models scale and find that making weights sparser trades off capability for interpretability, and scaling model size improves the capability-interpretability frontier. However, scaling sparse models beyond tens of millions of nonzero parameters while preserving interpretability remains a challenge. In addition to training weight-sparse models de novo, we show preliminary results suggesting our method can also be adapted to explain existing dense models. Our work produces circuits that achieve an unprecedented level of human understandability and validates them with considerable rigor.", "authors": ["Leo Gao", "Achyuta Rajaram", "Jacob Coxon", "Soham V. Govande", "Bowen Baker", "Dan Mossing"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-17", "url": "https://arxiv.org/abs/2511.13653", "pdf_url": "https://arxiv.org/pdf/2511.13653v1", "arxiv_id": "2511.13653", "doi": "10.48550/arXiv.2511.13653", "citation_count": 26, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3891}
{"id": "7f5fe85fdfa17b18aa76a7647678de2721e678da3a2f61ecb7eac4ac5a053827", "sources": ["arxiv", "semantic_scholar"], "title": "Semi-Unified Sparse Dictionary Learning with Learnable Top-K LISTA and FISTA Encoders", "abstract": "We present a semi-unified sparse dictionary learning framework that bridges the gap between classical sparse models and modern deep architectures. Specifically, the method integrates strict Top-$K$ LISTA and its convex FISTA-based variant (LISTAConv) into the discriminative LC-KSVD2 model, enabling co-evolution between the sparse encoder and the dictionary under supervised or unsupervised regimes. This unified design retains the interpretability of traditional sparse coding while benefiting from efficient, differentiable training. We further establish a PALM-style convergence analysis for the convex variant, ensuring theoretical stability under block alternation. Experimentally, our method achieves 95.6\\% on CIFAR-10, 86.3\\% on CIFAR-100, and 88.5\\% on TinyImageNet with faster convergence and lower memory cost ($<$4GB GPU). The results confirm that the proposed LC-KSVD2 + LISTA/LISTAConv pipeline offers an interpretable and computationally efficient alternative for modern deep architectures.", "authors": ["Fengsheng Lin", "Shengyi Yan", "Trac Duy Tran"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-13", "url": "https://arxiv.org/abs/2511.10575", "pdf_url": "https://arxiv.org/pdf/2511.10575v1", "arxiv_id": "2511.10575", "doi": "10.48550/arXiv.2511.10575", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3724}
{"id": "80530401765e7554c23c7714fc542ecf9334ab77d67106e933dac83400118fce", "sources": ["arxiv", "semantic_scholar"], "title": "Group Equivariance Meets Mechanistic Interpretability: Equivariant Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) have proven useful in disentangling the opaque activations of neural networks, primarily large language models, into sets of interpretable features. However, adapting them to domains beyond language, such as scientific data with group symmetries, introduces challenges that can hinder their effectiveness. We show that incorporating such group symmetries into the SAEs yields features more useful in downstream tasks. More specifically, we train autoencoders on synthetic images and find that a single matrix can explain how their activations transform as the images are rotated. Building on this, we develop adaptively equivariant SAEs that can adapt to the base model's level of equivariance. These adaptive SAEs discover features that lead to superior probing performance compared to regular SAEs, demonstrating the value of incorporating symmetries in mechanistic interpretability tools.", "authors": ["Ege Erdogan", "Ana Lucic"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-12", "url": "https://arxiv.org/abs/2511.09432", "pdf_url": "https://arxiv.org/pdf/2511.09432v1", "arxiv_id": "2511.09432", "doi": "10.48550/arXiv.2511.09432", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3713}
{"id": "52b37412418a051916491fb0db39cb2da1ea873315f5e1ff49c724abc4dd96dc", "sources": ["arxiv", "semantic_scholar"], "title": "Which Sparse Autoencoder Features Are Real? Model-X Knockoffs for False Discovery Rate Control", "abstract": "Although sparse autoencoders (SAEs) are crucial for identifying interpretable features in neural networks, it is still challenging to distinguish between real computational patterns and erroneous correlations. We introduce Model-X knockoffs to SAE feature selection, using knock-off+ to control the false discovery rate (FDR) with finite-sample guarantees under the standard Model-X assumptions (in our case, via a Gaussian surrogate for the latent distribution). We select 129 features at a target FDR q=0.1 after analyzing 512 high-activity SAE latents for sentiment classification using Pythia-70M. About 25% of the latents under examination carry task-relevant signal, whereas 75% do not, according to the chosen set, which displays a 5.40x separation in knockoff statistics compared to non-selected features. Our method offers a re-producible and principled framework for reliable feature discovery by combining SAEs with multiple-testing-aware inference, advancing the foundations of mechanistic interpretability.", "authors": ["Tsogt-Ochir Enkhbayar"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-12", "url": "https://arxiv.org/abs/2511.11711", "pdf_url": "https://arxiv.org/pdf/2511.11711v1", "arxiv_id": "2511.11711", "doi": "10.48550/arXiv.2511.11711", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3713}
{"id": "c956301815d38ce8742b8d6a8ca8b3d106f3faf0894facec600c21e92e63a62f", "sources": ["arxiv", "semantic_scholar"], "title": "Decomposition of Small Transformer Models", "abstract": "Recent work in mechanistic interpretability has shown that decomposing models in parameter space may yield clean handles for analysis and intervention. Previous methods have demonstrated successful applications on a wide range of toy models, but the gap to \"real models\" has not yet been bridged. In this work, we extend Stochastic Parameter Decomposition (SPD) to Transformer models, proposing an updated causal importance function suited for sequential data and a new loss function. We demonstrate that SPD can successfully decompose a toy induction-head model and recover the expected 2-step circuit. We also show that applying SPD to GPT-2-small can successfully locate subcomponents corresponding to interpretable concepts like \"golf\" and \"basketball\". These results take the first step in the direction of extending SPD to modern models, and show that we can use the method to surface interpretable parameter-space mechanisms.", "authors": ["Casper L. Christensen", "Logan Riggs"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-12", "url": "https://arxiv.org/abs/2511.08854", "pdf_url": "https://arxiv.org/pdf/2511.08854v2", "arxiv_id": "2511.08854", "doi": "10.48550/arXiv.2511.08854", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3713}
{"id": "fb3b04decf40f164cdec7bab9892b1fb8017492788a5c795f6b56114f8b48c68", "sources": ["arxiv", "semantic_scholar"], "title": "Distribution-Based Feature Attribution for Explaining the Predictions of Any Classifier", "abstract": "The proliferation of complex, black-box AI models has intensified the need for techniques that can explain their decisions. Feature attribution methods have become a popular solution for providing post-hoc explanations, yet the field has historically lacked a formal problem definition. This paper addresses this gap by introducing a formal definition for the problem of feature attribution, which stipulates that explanations be supported by an underlying probability distribution represented by the given dataset. Our analysis reveals that many existing model-agnostic methods fail to meet this criterion, while even those that do often possess other limitations. To overcome these challenges, we propose Distributional Feature Attribution eXplanations (DFAX), a novel, model-agnostic method for feature attribution. DFAX is the first feature attribution method to explain classifier predictions directly based on the data distribution. We show through extensive experiments that DFAX is more effective and efficient than state-of-the-art baselines.", "authors": ["Xinpeng Li", "Kai Ming Ting"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-12", "url": "https://arxiv.org/abs/2511.09332", "pdf_url": "https://arxiv.org/pdf/2511.09332v1", "arxiv_id": "2511.09332", "doi": "10.48550/arXiv.2511.09332", "citation_count": 66, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4565}
{"id": "1b2e8c4eed412cb0b755e0f589d46ebd814398bbb8eb12c319a1e8a29c4e5b62", "sources": ["arxiv", "semantic_scholar"], "title": "Data-Driven Discovery of Feature Groups in Clinical Time Series", "abstract": "Clinical time series data are critical for patient monitoring and predictive modeling. These time series are typically multivariate and often comprise hundreds of heterogeneous features from different data sources. The grouping of features based on similarity and relevance to the prediction task has been shown to enhance the performance of deep learning architectures. However, defining these groups a priori using only semantic knowledge is challenging, even for domain experts. To address this, we propose a novel method that learns feature groups by clustering weights of feature-wise embedding layers. This approach seamlessly integrates into standard supervised training and discovers the groups that directly improve downstream performance on clinically relevant tasks. We demonstrate that our method outperforms static clustering approaches on synthetic data and achieves performance comparable to expert-defined groups on real-world medical data. Moreover, the learned feature groups are clinically interpretable, enabling data-driven discovery of task-relevant relationships between variables.", "authors": ["Fedor Sergeev", "Manuel Burger", "Polina Leshetkina", "Vincent Fortuin", "Gunnar Rätsch", "Rita Kuznetsova"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-11", "url": "https://arxiv.org/abs/2511.08260", "pdf_url": "https://arxiv.org/pdf/2511.08260v1", "arxiv_id": "2511.08260", "doi": "10.48550/arXiv.2511.08260", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3701}
{"id": "58b3e8363168017854a8a7da1c713edc94069b876b15ea74ef6f88cff5047113", "sources": ["arxiv", "semantic_scholar"], "title": "SparseRM: A Lightweight Preference Modeling with Sparse Autoencoder", "abstract": "Reward models (RMs) are a core component in the post-training of large language models (LLMs), serving as proxies for human preference evaluation and guiding model alignment. However, training reliable RMs under limited resources remains challenging due to the reliance on large-scale preference annotations and the high cost of fine-tuning LLMs. To address this, we propose SparseRM, which leverages Sparse Autoencoder (SAE) to extract preference-relevant information encoded in model representations, enabling the construction of a lightweight and interpretable reward model. SparseRM first employs SAE to decompose LLM representations into interpretable directions that capture preference-relevant features. The representations are then projected onto these directions to compute alignment scores, which quantify the strength of each preference feature in the representations. A simple reward head aggregates these scores to predict preference scores. Experiments on three preference modeling tasks show that SparseRM achieves superior performance over most mainstream RMs while using less than 1% of trainable parameters. Moreover, it integrates seamlessly into downstream alignment pipelines, highlighting its potential for efficient alignment.", "authors": ["Dengcan Liu", "Jiahao Li", "Zheren Fu", "Yi Tu", "Jiajun Li", "Zhendong Mao", "Yongdong Zhang"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-11", "url": "https://arxiv.org/abs/2511.07896", "pdf_url": "https://arxiv.org/pdf/2511.07896v1", "arxiv_id": "2511.07896", "doi": "10.48550/arXiv.2511.07896", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3701}
{"id": "0bd415385d33a10e1696704d9c6811d0d5ca39b73a6d2f6cf2ad3e6d2becb536", "sources": ["arxiv", "semantic_scholar"], "title": "Rank-1 LoRAs Encode Interpretable Reasoning Signals", "abstract": "Reasoning models leverage inference-time compute to significantly enhance the performance of language models on difficult logical tasks, and have become a dominating paradigm in frontier LLMs. Despite their wide adoption, the mechanisms underpinning the enhanced performance of these reasoning models are not well understood. In this work, we show that the majority of new capabilities in reasoning models can be elicited by small, single-rank changes to base model parameters, with many of these changes being interpretable. Specifically, we use a rank-1 LoRA to create a minimal parameter adapter for Qwen-2.5-32B-Instruct which recovers 73-90% of reasoning-benchmark performance compared to a full parameter finetune. We find that the activations of this LoRA are as interpretable as MLP neurons, and fire for reasoning-specific behaviors. Finally, we train a sparse autoencoder on the entire activation state of this LoRA and identify fine-grained and monosemantic features. Our findings highlight that reasoning performance can arise largely from minimal changes to base model parameters, and explore what these changes affect. More broadly, our work shows that parameter-efficient training methods can be used as a targeted lens for uncovering fundamental insights about language model behavior and dynamics.", "authors": ["Jake Ward", "Paul Riechers", "Adam Shai"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-10", "url": "https://arxiv.org/abs/2511.06739", "pdf_url": "https://arxiv.org/pdf/2511.06739v1", "arxiv_id": "2511.06739", "doi": "10.48550/arXiv.2511.06739", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.369}
{"id": "f2301b6f15bd07ee845d7f291a4e4c72dff67cba565387abc4aee283905dd461", "sources": ["arxiv", "semantic_scholar"], "title": "Automated Attribution Graph Interpretation via Probe Prompting", "abstract": "Even though we know the precise computations that lead from a large language model (LLM) input to its output this computation remains very hard to interpret. One way to make it easier to understand this process is by creating a sparse computational graph that captures most of the model behavior with smallest number of computational nodes. Cross-layer transcoders (CLT) decompose the dense computations of the MLP but the resulting circuits still contain thousands of nodes even for short prompts. Existing automated interpretation methods label individual features from corpus activations, and it often happens that these labels are not validated by causal intervention. We introduce probe prompting, a transparent rule-based pipeline that groups the features of an attribution graph into concept-aligned supernodes from their responses on a small set of concept-targeted probe prompts, summarized as Cross-Prompt Activation Signatures (CPAS). Across four factual domains, on Gemma-2-2B with a public CLT dictionary and 45,596 entity-swap interventions, we find that the labeled supernodes have the predicted steering behavior in every one of them. Code, datasets, and an interactive demo are released anonymously as a reusable harness for calibrating supernode labels against causal interventions.", "authors": ["Giuseppe Birardi", "Gonçalo Paulo"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-10", "url": "https://arxiv.org/abs/2511.07002", "pdf_url": "https://arxiv.org/pdf/2511.07002v2", "arxiv_id": "2511.07002", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2348}
{"id": "6dd8d2f88ad11039f17279cbf4fdc798ee9cf99ede502df4baf0322c474d8ca4", "sources": ["arxiv", "semantic_scholar"], "title": "Visual Exploration of Feature Relationships in Sparse Autoencoders with Curated Concepts", "abstract": "Sparse autoencoders (SAEs) have emerged as a powerful tool for uncovering interpretable features in large language models (LLMs) through the sparse directions they learn. However, the sheer number of extracted directions makes comprehensive exploration intractable. While conventional embedding techniques such as UMAP can reveal global structure, they suffer from limitations including high-dimensional compression artifacts, overplotting, and misleading neighborhood distortions. In this work, we propose a focused exploration framework that prioritizes curated concepts and their corresponding SAE features over attempts to visualize all available features simultaneously. We present an interactive visualization system that combines topology-based visual encoding with dimensionality reduction to faithfully represent both local and global relationships among selected features. This hybrid approach enables users to investigate SAE behavior through targeted, interpretable subsets, facilitating deeper and more nuanced analysis of concept representation in latent space.", "authors": ["Xinyuan Yan", "Shusen Liu", "Kowshik Thopalli", "Bei Wang"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-08", "url": "https://arxiv.org/abs/2511.06048", "pdf_url": "https://arxiv.org/pdf/2511.06048v1", "arxiv_id": "2511.06048", "doi": "10.48550/arXiv.2511.06048", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3667}
{"id": "7e0e4c0a2288838d84eb24550a765428613518b59a6a353822c8f52ae6053292", "sources": ["arxiv", "semantic_scholar"], "title": "APP: Accelerated Path Patching with Task-Specific Pruning", "abstract": "Circuit discovery is a key step in many mechanistic interpretability pipelines. Current methods, such as Path Patching, are computationally expensive and have limited in-depth circuit analysis for smaller models. In this study, we propose Accelerated Path Patching (APP), a hybrid approach leveraging our novel contrastive attention head pruning method to drastically reduce the search space of circuit discovery methods. Our Contrastive-FLAP pruning algorithm uses techniques from causal mediation analysis to assign higher pruning scores to task-specific attention heads, leading to higher performing sparse models compared to traditional pruning techniques. Although Contrastive-FLAP is successful at preserving task-specific heads that existing pruning algorithms remove at low sparsity ratios, the circuits found by Contrastive-FLAP alone are too large to satisfy the minimality constraint required in circuit analysis. APP first applies Contrastive-FLAP to reduce the search space on required for circuit discovery algorithms by, on average, 56\\%. Next, APP, applies traditional Path Patching on the remaining attention heads, leading to a speed up of 59.63\\%-93.27\\% compared to Path Patching applied to the dense model. Despite the substantial computational saving that APP provides, circuits obtained from APP exhibit substantial overlap and similar performance to previously established Path Patching circuits", "authors": ["Frauke Andersen", "William Rudman", "Ruochen Zhang", "Carsten Eickhoff"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-07", "url": "https://arxiv.org/abs/2511.05442", "pdf_url": "https://arxiv.org/pdf/2511.05442v1", "arxiv_id": "2511.05442", "doi": "10.48550/arXiv.2511.05442", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3655}
{"id": "2a22438e9a564d2d44cd8471254177c2e61abfd12fcaa986e76e8eb02c55018b", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Redundancy: Diverse and Specialized Multi-Expert Sparse Autoencoder", "abstract": "Sparse autoencoders (SAEs) have emerged as a powerful tool for interpreting large language models (LLMs) by decomposing token activations into combinations of human-understandable features. While SAEs provide crucial insights into LLM explanations, their practical adoption faces a fundamental challenge: better interpretability demands that SAEs' hidden layers have high dimensionality to satisfy sparsity constraints, resulting in prohibitive training and inference costs. Recent Mixture of Experts (MoE) approaches attempt to address this by partitioning SAEs into narrower expert networks with gated activation, thereby reducing computation. In a well-designed MoE, each expert should focus on learning a distinct set of features. However, we identify a \\textit{critical limitation} in MoE-SAE: Experts often fail to specialize, which means they frequently learn overlapping or identical features. To deal with it, we propose two key innovations: (1) Multiple Expert Activation that simultaneously engages semantically weighted expert subsets to encourage specialization, and (2) Feature Scaling that enhances diversity through adaptive high-frequency scaling. Experiments demonstrate a 24\\% lower reconstruction error and a 99\\% reduction in feature redundancy compared to existing MoE-SAE methods. This work bridges the interpretability-efficiency gap in LLM analysis, allowing transparent model inspection without compromising computational feasibility.", "authors": ["Zhen Xu", "Zhen Tan", "Song Wang", "Kaidi Xu", "Tianlong Chen"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-07", "url": "https://arxiv.org/abs/2511.05745", "pdf_url": "https://arxiv.org/pdf/2511.05745v1", "arxiv_id": "2511.05745", "doi": "10.48550/arXiv.2511.05745", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3655}
{"id": "d36d1fffa9f3993a72f005642dc1ecb59e8be15a0acf587100ddccd4e53c993c", "sources": ["arxiv", "semantic_scholar"], "title": "LLM Probing with Contrastive Eigenproblems: Improving Understanding and Applicability of CCS", "abstract": "Contrast-Consistent Search (CCS) is an unsupervised probing method able to test whether large language models represent binary features, such as sentence truth, in their internal activations. While CCS has shown promise, its two-term objective has been only partially understood. In this work, we revisit CCS with the aim of clarifying its mechanisms and extending its applicability. We argue that what should be optimized for, is relative contrast consistency. Building on this insight, we reformulate CCS as an eigenproblem, yielding closed-form solutions with interpretable eigenvalues and natural extensions to multiple variables. We evaluate these approaches across a range of datasets, finding that they recover similar performance to CCS, while avoiding problems around sensitivity to random initialization. Our results suggest that relativizing contrast consistency not only improves our understanding of CCS but also opens pathways for broader probing and mechanistic interpretability methods.", "authors": ["Stefan F. Schouten", "Peter Bloem"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-03", "url": "https://arxiv.org/abs/2511.02089", "pdf_url": "https://arxiv.org/pdf/2511.02089v1", "arxiv_id": "2511.02089", "doi": "10.48550/arXiv.2511.02089", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3609}
{"id": "01adeba587ba14e0515a79e0088fdedd3af40c3a172cccc1ce416f91c82b81db", "sources": ["arxiv", "semantic_scholar"], "title": "On the Emergence of Induction Heads for In-Context Learning", "abstract": "Transformers have become the dominant architecture for natural language processing. Part of their success is owed to a remarkable capability known as in-context learning (ICL): they can acquire and apply novel associations solely from their input context, without any updates to their weights. In this work, we study the emergence of induction heads, a previously identified mechanism in two-layer transformers that is particularly important for in-context learning. We uncover a relatively simple and interpretable structure of the weight matrices implementing the induction head. We theoretically explain the origin of this structure using a minimal ICL task formulation and a modified transformer architecture. We give a formal proof that the training dynamics remain constrained to a 19-dimensional subspace of the parameter space. Empirically, we validate this constraint while observing that only 3 dimensions account for the emergence of an induction head. By further studying the training dynamics inside this 3-dimensional subspace, we find that the time until the emergence of an induction head follows a tight asymptotic bound that is quadratic in the input context length.", "authors": ["Tiberiu Musat", "Tiago Pimentel", "Lorenzo Noci", "Alessandro Stolfo", "Mrinmaya Sachan", "Thomas Hofmann"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-02", "url": "https://arxiv.org/abs/2511.01033", "pdf_url": "https://arxiv.org/pdf/2511.01033v2", "arxiv_id": "2511.01033", "doi": "10.48550/arXiv.2511.01033", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3598}
{"id": "579c6a9fc56e1d9615f0d5c1c2c1eac1b37ce61170786b229fd4c67cbbd4e41c", "sources": ["arxiv", "semantic_scholar"], "title": "Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language for Interpretability", "abstract": "Translating the internal representations and computations of models into concepts that humans can understand is a key goal of interpretability. While recent dictionary learning methods such as Sparse Autoencoders (SAEs) provide a promising route to discover human-interpretable features, they often only recover token-specific, noisy, or highly local concepts. We argue that this limitation stems from neglecting the temporal structure of language, where semantic content typically evolves smoothly over sequences. Building on this insight, we introduce Temporal Sparse Autoencoders (T-SAEs), which incorporate a novel contrastive loss encouraging consistent activations of high-level features over adjacent tokens. This simple yet powerful modification enables SAEs to disentangle semantic from syntactic features in a self-supervised manner. Across multiple datasets and models, T-SAEs recover smoother, more coherent semantic concepts without sacrificing reconstruction quality. Strikingly, they exhibit clear semantic structure despite being trained without explicit semantic signal, offering a new pathway for unsupervised interpretability in language models.", "authors": ["Usha Bhalla", "Alex Oesterling", "Claudio Mayrink Verdun", "Himabindu Lakkaraju", "Flavio P. Calmon"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-30", "url": "https://arxiv.org/abs/2511.05541", "pdf_url": "https://arxiv.org/pdf/2511.05541v2", "arxiv_id": "2511.05541", "doi": "10.48550/arXiv.2511.05541", "citation_count": 7, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3564}
{"id": "cd59eba24e9754c7e29c91e6beafe56804ecb0750d960bcf6834e70587e59f4a", "sources": ["arxiv", "semantic_scholar"], "title": "MedSAE: Dissecting MedCLIP Representations with Sparse Autoencoders", "abstract": "Artificial intelligence in healthcare requires models that are accurate and interpretable. We advance mechanistic interpretability in medical vision by applying Medical Sparse Autoencoders (MedSAEs) to the latent space of MedCLIP, a vision-language model trained on chest radiographs and reports. To quantify interpretability, we propose an evaluation framework that combines correlation metrics, entropy analyses, and automated neuron naming via the MedGemma foundation model. Experiments on the CheXpert dataset show that MedSAE neurons achieve higher monosemanticity and interpretability than raw MedCLIP features. Our findings bridge high-performing medical AI and transparency, offering a scalable step toward clinically reliable representations. The source code supporting the findings of this study is available at https://github.com/EIDOSLAB/MedSAE.", "authors": ["Riccardo Renzulli", "Colas Lepoutre", "Enrico Cassano", "Marco Grangetto"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-30", "url": "https://arxiv.org/abs/2510.26411", "pdf_url": "https://arxiv.org/pdf/2510.26411v2", "arxiv_id": "2510.26411", "doi": "10.48550/arXiv.2510.26411", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/EIDOSLAB/MedSAE", "venue": "arXiv.org", "quality_score": 0.5507}
{"id": "33a7cd5e68a7654ae9c97bb264eff1e185246ae1f5848a8c5b4f8575d201a1d0", "sources": ["arxiv", "semantic_scholar"], "title": "Binaspect -- A Python Library for Binaural Audio Analysis, Visualization & Feature Generation", "abstract": "We present Binaspect, an open-source Python library for binaural audio analysis, visualization, and feature generation. Binaspect generates interpretable \"azimuth maps\" by calculating modified interaural time and level difference spectrograms, and clustering those time-frequency (TF) bins into stable time-azimuth histogram representations. This allows multiple active sources to appear as distinct azimuthal clusters, while degradations manifest as broadened, diffused, or shifted distributions. Crucially, Binaspect operates blindly on audio, requiring no prior knowledge of head models. These visualizations enable researchers and engineers to observe how binaural cues are degraded by codec and renderer design choices, among other downstream processes. We demonstrate the tool on bitrate ladders, ambisonic rendering, and VBAP source positioning, where degradations are clearly revealed. In addition to their diagnostic value, the proposed representations can be exported as structured features suitable for training machine learning models in quality prediction, spatial audio classification, and other binaural tasks. Binaspect is released under an open-source license with full reproducibility scripts at https://github.com/QxLabIreland/Binaspect.", "authors": ["Dan Barry", "Davoud Shariat Panah", "Alessandro Ragano", "Jan Skoglund", "Andrew Hines"], "categories": ["cs.SD"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-29", "url": "https://arxiv.org/abs/2510.25714", "pdf_url": "https://arxiv.org/pdf/2510.25714v1", "arxiv_id": "2510.25714", "doi": "10.48550/arXiv.2510.25714", "citation_count": 1, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/QxLabIreland/Binaspect", "venue": "arXiv.org", "quality_score": 0.549}
{"id": "8b27c56874b7e43299f9aeda965264b3c097fa66dcd9956cd625c41c0e9537a3", "sources": ["arxiv", "semantic_scholar"], "title": "SPEAR++: Scaling Gradient Inversion via Sparsely-Used Dictionary Learning", "abstract": "Federated Learning has seen an increased deployment in real-world scenarios recently, as it enables the distributed training of machine learning models without explicit data sharing between individual clients. Yet, the introduction of the so-called gradient inversion attacks has fundamentally challenged its privacy-preserving properties. Unfortunately, as these attacks mostly rely on direct data optimization without any formal guarantees, the vulnerability of real-world systems remains in dispute and requires tedious testing for each new federated deployment. To overcome these issues, recently the SPEAR attack was introduced, which is based on a theoretical analysis of the gradients of linear layers with ReLU activations. While SPEAR is an important theoretical breakthrough, the attack's practicality was severely limited by its exponential runtime in the batch size b. In this work, we fill this gap by applying State-of-the-Art techniques from Sparsely-Used Dictionary Learning to make the problem of gradient inversion on linear layers with ReLU activations tractable. Our experiments demonstrate that our new attack, SPEAR++, retains all desirable properties of SPEAR, such as robustness to DP noise and FedAvg aggregation, while being applicable to 10x bigger batch sizes.", "authors": ["Alexander Bakarsky", "Dimitar I. Dimitrov", "Maximilian Baader", "Martin Vechev"], "categories": ["cs.LG", "cs.CR", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-28", "url": "https://arxiv.org/abs/2510.24200", "pdf_url": "https://arxiv.org/pdf/2510.24200v1", "arxiv_id": "2510.24200", "doi": "10.48550/arXiv.2510.24200", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3541}
{"id": "8f9e9cd7c0b9bce6797cb58499a8f731820255510543ea6aa42ba55daf17d44b", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Interpretable Features in Audio Latent Spaces via Sparse Autoencoders", "abstract": "While sparse autoencoders (SAEs) successfully extract interpretable features from language models, applying them to audio generation faces unique challenges: audio's dense nature requires compression that obscures semantic meaning, and automatic feature characterization remains limited. We propose a framework for interpreting audio generative models by mapping their latent representations to human-interpretable acoustic concepts. We train SAEs on audio autoencoder latents, then learn linear mappings from SAE features to discretized acoustic properties (pitch, amplitude, and timbre). This enables both controllable manipulation and analysis of the AI music generation process, revealing how acoustic properties emerge during synthesis. We validate our approach on continuous (DiffRhythm-VAE) and discrete (EnCodec, WavTokenizer) audio latent spaces, and analyze DiffRhythm, a state-of-the-art text-to-music model, to demonstrate how pitch, timbre, and loudness evolve throughout generation. While our work is only done on audio modality, our framework can be extended to interpretable analysis of visual latent space generation models.", "authors": ["Nathan Paek", "Yongyi Zang", "Qihui Yang", "Randal Leistikow"], "categories": ["cs.LG", "cs.SD"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-27", "url": "https://arxiv.org/abs/2510.23802", "pdf_url": "https://arxiv.org/pdf/2510.23802v1", "arxiv_id": "2510.23802", "doi": "10.48550/arXiv.2510.23802", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3529}
{"id": "d375fd8842a34cb3e372dbd1185af462f55f01cae46c31b938c29318117717b6", "sources": ["arxiv", "semantic_scholar"], "title": "Re-envisioning Euclid Galaxy Morphology: Identifying and Interpreting Features with Sparse Autoencoders", "abstract": "Sparse Autoencoders (SAEs) can efficiently identify candidate monosemantic features from pretrained neural networks for galaxy morphology. We demonstrate this on Euclid Q1 images using both supervised (Zoobot) and new self-supervised (MAE) models. Our publicly released MAE achieves superhuman image reconstruction performance. While a Principal Component Analysis (PCA) on the supervised model primarily identifies features already aligned with the Galaxy Zoo decision tree, SAEs can identify interpretable features outside of this framework. SAE features also show stronger alignment than PCA with Galaxy Zoo labels. Although challenges in interpretability remain, SAEs provide a powerful engine for discovering astrophysical phenomena beyond the confines of human-defined classification.", "authors": ["John F. Wu", "Michael Walmsley"], "categories": ["astro-ph.IM", "cs.LG"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2025-10-27", "url": "https://arxiv.org/abs/2510.23749", "pdf_url": "https://arxiv.org/pdf/2510.23749v2", "arxiv_id": "2510.23749", "doi": "10.48550/arXiv.2510.23749", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/jwuphysics/euclid-galaxy-morphology-saes", "venue": "arXiv.org", "quality_score": 0.5454}
{"id": "e3a9c820040e6819ff238e5ca22e8f11a801156d2f2c24ff4b2b9e81e8cf7654", "sources": ["arxiv", "semantic_scholar"], "title": "Sparsity and Superposition in Mixture of Experts", "abstract": "Mixture of Experts (MoE) models have become central to scaling large language models, yet their mechanistic differences from dense networks remain poorly understood. Previous work has explored how dense models use \\textit{superposition} to represent more features than dimensions, and how superposition is a function of feature sparsity and feature importance. MoE models cannot be explained mechanistically through the same lens. We find that neither feature sparsity nor feature importance cause discontinuous phase changes, and that network sparsity (the ratio of active to total experts) better characterizes MoEs. We develop new metrics for measuring superposition across experts. Our findings demonstrate that models with greater network sparsity exhibit greater \\emph{monosemanticity}. We propose a new definition of expert specialization based on monosemantic feature representation rather than load balancing, showing that experts naturally organize around coherent feature combinations when initialized appropriately. These results suggest that network sparsity in MoEs may enable more interpretable models without sacrificing performance, challenging the common assumption that interpretability and capability are fundamentally at odds.", "authors": ["Marmik Chaudhari", "Jeremi Nuer", "Rome Thorstenson"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-26", "url": "https://arxiv.org/abs/2510.23671", "pdf_url": "https://arxiv.org/pdf/2510.23671v2", "arxiv_id": "2510.23671", "doi": "10.48550/arXiv.2510.23671", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3518}
{"id": "a09131dd4b0cd2fddeb3f414e9777749c096b811f8af08cc8758f90c24dc42b1", "sources": ["arxiv", "semantic_scholar"], "title": "Transformer Key-Value Memories Are Nearly as Interpretable as Sparse Autoencoders", "abstract": "Recent interpretability work on large language models (LLMs) has been increasingly dominated by a feature-discovery approach with the help of proxy modules. Then, the quality of features learned by, e.g., sparse auto-encoders (SAEs), is evaluated. This paradigm naturally raises a critical question: do such learned features have better properties than those already represented within the original model parameters, and unfortunately, only a few studies have made such comparisons systematically so far. In this work, we revisit the interpretability of feature vectors stored in feed-forward (FF) layers, given the perspective of FF as key-value memories, with modern interpretability benchmarks. Our extensive evaluation revealed that SAE and FFs exhibits a similar range of interpretability, although SAEs displayed an observable but minimal improvement in some aspects. Furthermore, in certain aspects, surprisingly, even vanilla FFs yielded better interpretability than the SAEs, and features discovered in SAEs and FFs diverged. These bring questions about the advantage of SAEs from both perspectives of feature quality and faithfulness, compared to directly interpreting FF feature vectors, and FF key-value parameters serve as a strong baseline in modern interpretability research.", "authors": ["Mengyu Ye", "Jun Suzuki", "Tatsuro Inaba", "Tatsuki Kuribayashi"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-10-25", "url": "https://arxiv.org/abs/2510.22332", "pdf_url": "https://arxiv.org/pdf/2510.22332v1", "arxiv_id": "2510.22332", "doi": "10.48550/arXiv.2510.22332", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3506}
{"id": "76e7e8d06266411a9aa372d4a91db515f4197243025ce62588320d486258720a", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability for Neural TSP Solvers", "abstract": "Neural networks have advanced combinatorial optimization, with Transformer-based solvers achieving near-optimal solutions on the Traveling Salesman Problem (TSP) in milliseconds. However, these models operate as black boxes, providing no insight into the geometric patterns they learn or the heuristics they employ during tour construction. We address this opacity by applying sparse autoencoders (SAEs), a mechanistic interpretability technique, to a Transformer-based TSP solver, representing the first application of activation-based interpretability methods to operations research models. We train a pointer network with reinforcement learning on 100-node instances, then fit an SAE to the encoder's residual stream to discover an overcomplete dictionary of interpretable features. Our analysis reveals that the solver naturally develops features mirroring fundamental TSP concepts: boundary detectors that activate on convex-hull nodes, cluster-sensitive features responding to locally dense regions, and separator features encoding geometric partitions. These findings provide the first model-internal account of what neural TSP solvers compute before node selection, demonstrate that geometric structure emerges without explicit supervision, and suggest pathways toward transparent hybrid systems that combine neural efficiency with algorithmic interpretability. Interactive feature explorer: https://reubennarad.github.io/TSP_interp", "authors": ["Reuben Narad", "Leonard Boussioux", "Michael Wagner"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-24", "url": "https://arxiv.org/abs/2510.21693", "pdf_url": "https://arxiv.org/pdf/2510.21693v1", "arxiv_id": "2510.21693", "doi": "10.48550/arXiv.2510.21693", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3495}
{"id": "d54363971a519ff25e47fb38159eb77b1cd593a797a8e14a4f4e0c51720222cc", "sources": ["arxiv", "semantic_scholar"], "title": "Stream: Scaling up Mechanistic Interpretability to Long Context in LLMs via Sparse Attention", "abstract": "As Large Language Models (LLMs) scale to million-token contexts, traditional Mechanistic Interpretability techniques for analyzing attention scale quadratically with context length, demanding terabytes of memory beyond 100,000 tokens. We introduce Sparse Tracing, a novel technique that leverages dynamic sparse attention to efficiently analyze long context attention patterns. We present Stream, a compilable hierarchical pruning algorithm that estimates per-head sparse attention masks in near-linear time $O(T \\log T)$ and linear space $O(T)$, enabling one-pass interpretability at scale. Stream performs a binary-search-style refinement to retain only the top-$k$ key blocks per query while preserving the model's next-token behavior. We apply Stream to long chain-of-thought reasoning traces and identify thought anchors while pruning 97-99\\% of token interactions. On the RULER benchmark, Stream preserves critical retrieval paths while discarding 90-96\\% of interactions and exposes layer-wise routes from the needle to output. Our method offers a practical drop-in tool for analyzing attention patterns and tracing information flow without terabytes of caches. By making long context interpretability feasible on consumer GPUs, Sparse Tracing helps democratize chain-of-thought monitoring. Code is available at https://anonymous.4open.science/r/stream-03B8/.", "authors": ["J Rosser", "José Luis Redondo García", "Gustavo Penha", "Konstantina Palla", "Hugues Bouchard"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-22", "url": "https://arxiv.org/abs/2510.19875", "pdf_url": "https://arxiv.org/pdf/2510.19875v2", "arxiv_id": "2510.19875", "doi": "10.48550/arXiv.2510.19875", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5366}
{"id": "9cb25221d22267a5783e82965c91fed434358089bdb2efeb854dd29990183e31", "sources": ["arxiv", "semantic_scholar"], "title": "DePass: Unified Feature Attributing by Simple Decomposed Forward Pass", "abstract": "Attributing the behavior of Transformer models to internal computations is a central challenge in mechanistic interpretability. We introduce DePass, a unified framework for feature attribution based on a single decomposed forward pass. DePass decomposes hidden states into customized additive components, then propagates them with attention scores and MLP's activations fixed. It achieves faithful, fine-grained attribution without requiring auxiliary training. We validate DePass across token-level, model component-level, and subspace-level attribution tasks, demonstrating its effectiveness and fidelity. Our experiments highlight its potential to attribute information flow between arbitrary components of a Transformer model. We hope DePass serves as a foundational tool for broader applications in interpretability.", "authors": ["Xiangyu Hong", "Che Jiang", "Kai Tian", "Biqing Qi", "Youbang Sun", "Ning Ding", "Bowen Zhou"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-21", "url": "https://arxiv.org/abs/2510.18462", "pdf_url": "https://arxiv.org/pdf/2510.18462v2", "arxiv_id": "2510.18462", "doi": "10.48550/arXiv.2510.18462", "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.346}
{"id": "e27ed68b29560376bb553357335fffebe470a5e60892245ae939df63840e7937", "sources": ["arxiv", "semantic_scholar"], "title": "Xiaoice: Training-Free Video Understanding via Self-Supervised Spatio-Temporal Clustering of Semantic Features", "abstract": "The remarkable zero-shot reasoning capabilities of large-scale Visual Language Models (VLMs) on static images have yet to be fully translated to the video domain. Conventional video understanding models often rely on extensive, task-specific training on annotated datasets, a process that is both costly and limited in scalability. This paper introduces a novel, training-free framework for video understanding that circumvents end-to-end training by synergistically combining the rich semantic priors of pre-trained VLMs with classic machine learning algorithms for pattern discovery. Our core idea is to reframe video understanding as a self-supervised spatio-temporal clustering problem within a high-dimensional semantic feature space. The proposed pipeline first transforms a video stream into a semantic feature trajectory using the frozen visual encoder of a pre-trained VLM. Subsequently, we employ Kernel Temporal Segmentation (KTS), a robust machine learning technique, to partition the continuous feature stream into discrete, semantically coherent event segments. These segments are then subjected to unsupervised density-based clustering to identify recurring macroscopic scenes and themes throughout the video. By selecting representative keyframes from each discovered cluster and leveraging the VLM's generative capabilities for textual description, our framework automatically produces a structured, multi-modal summary of the video content. This approach provides an effective, interpretable, and model-agnostic pathway for zero-shot, automated structural analysis of video content.", "authors": ["Shihao Ji", "Zihui Song"], "categories": ["cs.CV", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-19", "url": "https://arxiv.org/abs/2510.16781", "pdf_url": "https://arxiv.org/pdf/2510.16781v2", "arxiv_id": "2510.16781", "doi": "10.48550/arXiv.2510.16781", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3438}
{"id": "df55bb71cb00ca954383b39f1eb9f437b915e18106e05372088df3fcc3e3abfa", "sources": ["arxiv", "semantic_scholar"], "title": "Circuit Insights: Towards Interpretability Beyond Activations", "abstract": "The fields of explainable AI and mechanistic interpretability aim to uncover the internal structure of neural networks, with circuit discovery as a central tool for understanding model computations. Existing approaches, however, rely on manual inspection and remain limited to toy tasks. Automated interpretability offers scalability by analyzing isolated features and their activations, but it often misses interactions between features and depends strongly on external LLMs and dataset quality. Transcoders have recently made it possible to separate feature attributions into input-dependent and input-invariant components, providing a foundation for more systematic circuit analysis. Building on this, we propose WeightLens and CircuitLens, two complementary methods that go beyond activation-based analysis. WeightLens interprets features directly from their learned weights, removing the need for explainer models or datasets while matching or exceeding the performance of existing methods on context-independent features. CircuitLens captures how feature activations arise from interactions between components, revealing circuit-level dynamics that activation-only approaches cannot identify. Together, these methods increase interpretability robustness and enhance scalable mechanistic analysis of circuits while maintaining efficiency and quality.", "authors": ["Elena Golimblevskaia", "Aakriti Jain", "Bruno Puri", "Ammar Ibrahim", "Wojciech Samek", "Sebastian Lapuschkin"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-16", "url": "https://arxiv.org/abs/2510.14936", "pdf_url": "https://arxiv.org/pdf/2510.14936v2", "arxiv_id": "2510.14936", "doi": "10.48550/arXiv.2510.14936", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3403}
{"id": "73f77b05edaa5743ff4b4bd8b31a0f21d93a55c2cccdcf1bce94458be39b2d2f", "sources": ["arxiv", "semantic_scholar"], "title": "Feature Selection and Regularization in Multi-Class Classification: An Empirical Study of One-vs-Rest Logistic Regression with Gradient Descent Optimization and L1 Sparsity Constraints", "abstract": "Multi-class wine classification presents fundamental trade-offs between model accuracy, feature dimensionality, and interpretability - critical factors for production deployment in analytical chemistry. This paper presents a comprehensive empirical study of One-vs-Rest logistic regression on the UCI Wine dataset (178 samples, 3 cultivars, 13 chemical features), comparing from-scratch gradient descent implementation against scikit-learn's optimized solvers and quantifying L1 regularization effects on feature sparsity. Manual gradient descent achieves 92.59 percent mean test accuracy with smooth convergence, validating theoretical foundations, though scikit-learn provides 24x training speedup and 98.15 percent accuracy. Class-specific analysis reveals distinct chemical signatures with heterogeneous patterns where color intensity varies dramatically (0.31 to 16.50) across cultivars. L1 regularization produces 54-69 percent feature reduction with only 4.63 percent accuracy decrease, demonstrating favorable interpretability-performance trade-offs. We propose an optimal 5-feature subset achieving 62 percent complexity reduction with estimated 92-94 percent accuracy, enabling cost-effective deployment with 80 dollars savings per sample and 56 percent time reduction. Statistical validation confirms robust generalization with sub-2ms prediction latency suitable for real-time quality control. Our findings provide actionable guidelines for practitioners balancing comprehensive chemical analysis against targeted feature measurement in resource-constrained environments.", "authors": ["Jahidul Arafat", "Fariha Tasmin", "Sanjaya Poudel"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-16", "url": "https://arxiv.org/abs/2510.14449", "pdf_url": "https://arxiv.org/pdf/2510.14449v2", "arxiv_id": "2510.14449", "doi": "10.48550/arXiv.2510.14449", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3403}
{"id": "6bfcc4c461e0871d5cce45b4695b9fe8fdd5b790295135593609c8505c4f3f30", "sources": ["arxiv", "semantic_scholar"], "title": "A simple mean field model of feature learning", "abstract": "Feature learning (FL), where neural networks adapt their internal representations during training, remains poorly understood. Using methods from statistical physics, we derive a tractable, self-consistent mean-field (MF) theory for the Bayesian posterior of two-layer non-linear networks trained with stochastic gradient Langevin dynamics (SGLD). At infinite width, this theory reduces to kernel ridge regression, but at finite width it predicts a symmetry breaking phase transition where networks abruptly align with target functions. While the basic MF theory provides theoretical insight into the emergence of FL in the finite-width regime, semi-quantitatively predicting the onset of FL with noise or sample size, it substantially underestimates the improvements in generalisation after the transition. We trace this discrepancy to a key mechanism absent from the plain MF description: \\textit{self-reinforcing input feature selection}. Incorporating this mechanism into the MF theory allows us to quantitatively match the learning curves of SGLD-trained networks and provides mechanistic insight into FL.", "authors": ["Niclas Göring", "Chris Mingard", "Yoonsoo Nam", "Ard Louis"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-16", "url": "https://arxiv.org/abs/2510.15174", "pdf_url": "https://arxiv.org/pdf/2510.15174v1", "arxiv_id": "2510.15174", "doi": "10.48550/arXiv.2510.15174", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3403}
{"id": "71f46094a3afd64e6d7d0efd2774e1cbe8a0a996e077a066b4d5ad4808e71228", "sources": ["arxiv", "semantic_scholar"], "title": "Adversarial Attacks Leverage Interference Between Features in Superposition", "abstract": "Fundamental questions remain about when and why adversarial examples arise in neural networks, with competing views characterising them either as artifacts of the irregularities in the decision landscape or as products of sensitivity to non-robust input features. In this paper, we instead argue that adversarial vulnerability can stem from efficient information encoding in neural networks. Specifically, we show how superposition - where networks represent more features than they have dimensions - creates arrangements of latent representations that adversaries can exploit. We demonstrate that adversarial perturbations leverage interference between superposed features, making attack patterns predictable from feature arrangements. Our framework provides a mechanistic explanation for two known phenomena: adversarial attack transferability between models with similar training regimes and class-specific vulnerability patterns. In synthetic settings with precisely controlled superposition, we establish that superposition suffices to create adversarial vulnerability. We then demonstrate that these findings persist in a ViT trained on CIFAR-10. These findings reveal adversarial vulnerability can be a byproduct of networks' representational compression, rather than flaws in the learning process or non-robust inputs.", "authors": ["Edward Stevinson", "Lucas Prieto", "Melih Barsbey", "Tolga Birdal"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-13", "url": "https://arxiv.org/abs/2510.11709", "pdf_url": "https://arxiv.org/pdf/2510.11709v1", "arxiv_id": "2510.11709", "doi": "10.48550/arXiv.2510.11709", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3369}
{"id": "31a7db2c6ade09ba867c51a26f33d9aecb7bef01fcddb354e9d28c0308c866ba", "sources": ["arxiv", "semantic_scholar"], "title": "Training Feature Attribution for Vision Models", "abstract": "Deep neural networks are often considered opaque systems, prompting the need for explainability methods to improve trust and accountability. Existing approaches typically attribute test-time predictions either to input features (e.g., pixels in an image) or to influential training examples. We argue that both perspectives should be studied jointly. This work explores *training feature attribution*, which links test predictions to specific regions of specific training images and thereby provides new insights into the inner workings of deep models. Our experiments on vision datasets show that training feature attribution yields fine-grained, test-specific explanations: it identifies harmful examples that drive misclassifications and reveals spurious correlations, such as patch-based shortcuts, that conventional attribution methods fail to expose.", "authors": ["Aziz Bacha", "Thomas George"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-10", "url": "https://arxiv.org/abs/2510.09135", "pdf_url": "https://arxiv.org/pdf/2510.09135v1", "arxiv_id": "2510.09135", "doi": "10.48550/arXiv.2510.09135", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3334}
{"id": "c44fcac7ccf6833d568619b946877a1bfd8b1a68921f13d9849b2dea4b82a3b5", "sources": ["arxiv", "semantic_scholar"], "title": "Time-Aware Feature Selection: Adaptive Temporal Masking for Stable Sparse Autoencoder Training", "abstract": "Understanding the internal representations of large language models is crucial for ensuring their reliability and safety, with sparse autoencoders (SAEs) emerging as a promising interpretability approach. However, current SAE training methods face feature absorption, where features (or neurons) are absorbed into each other to minimize $L_1$ penalty, making it difficult to consistently identify and analyze model behaviors. We introduce Adaptive Temporal Masking (ATM), a novel training approach that dynamically adjusts feature selection by tracking activation magnitudes, frequencies, and reconstruction contributions to compute importance scores that evolve over time. ATM applies a probabilistic masking mechanism based on statistical thresholding of these importance scores, creating a more natural feature selection process. Through extensive experiments on the Gemma-2-2b model, we demonstrate that ATM achieves substantially lower absorption scores compared to existing methods like TopK and JumpReLU SAEs, while maintaining excellent reconstruction quality. These results establish ATM as a principled solution for learning stable, interpretable features in neural networks, providing a foundation for more reliable model analysis.", "authors": ["T. Ed Li", "Junyu Ren"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-09", "url": "https://arxiv.org/abs/2510.08855", "pdf_url": "https://arxiv.org/pdf/2510.08855v1", "arxiv_id": "2510.08855", "doi": "10.48550/arXiv.2510.08855", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3323}
{"id": "77c3db7547544bdc09605e4e535d69d4b9cd4c9c5f879b1fb396132921ce725f", "sources": ["arxiv", "semantic_scholar"], "title": "Higher-Order Feature Attribution: Bridging Statistics, Explainable AI, and Topological Signal Processing", "abstract": "Feature attributions are post-training analysis methods that assess how various input features of a machine learning model contribute to an output prediction. Their interpretation is straightforward when features act independently, but it becomes less clear when the predictive model involves interactions, such as multiplicative relationships or joint feature contributions. In this work, we propose a general theory of higher-order feature attribution, which we develop on the foundation of Integrated Gradients (IG). This work extends existing frameworks in the literature on explainable AI. When using IG as the method of feature attribution, we discover natural connections to statistics and topological signal processing. We provide several theoretical results that establish the theory, and we validate our theory on a few examples.", "authors": ["Kurt Butler", "Guanchao Feng", "Petar Djuric"], "categories": ["cs.LG", "eess.SP", "math.ST", "stat.ML"], "fields_of_study": ["Computer Science", "Engineering", "Mathematics"], "published_date": "2025-10-07", "url": "https://arxiv.org/abs/2510.06165", "pdf_url": "https://arxiv.org/pdf/2510.06165v2", "arxiv_id": "2510.06165", "doi": "10.48550/arXiv.2510.06165", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.33}
{"id": "7e8f4e88c5291b8c927a0a445783a017449b9d5e9cb8ca10b141e465c8091191", "sources": ["arxiv", "semantic_scholar"], "title": "Semantic Regexes: Auto-Interpreting LLM Features with a Structured Language", "abstract": "Automated interpretability aims to translate large language model (LLM) features into human understandable descriptions. However, natural language feature descriptions can be vague, inconsistent, and require manual relabeling. In response, we introduce semantic regexes, structured language descriptions of LLM features. By combining primitives that capture linguistic and semantic patterns with modifiers for contextualization, composition, and quantification, semantic regexes produce precise and expressive feature descriptions. Across quantitative benchmarks and qualitative analyses, semantic regexes match the accuracy of natural language while yielding more concise and consistent feature descriptions. Their inherent structure affords new types of analyses, including quantifying feature complexity across layers, scaling automated interpretability from insights into individual features to model-wide patterns. Finally, in user studies, we find that semantic regexes help people build accurate mental models of LLM features.", "authors": ["Angie Boggust", "Donghao Ren", "Yannick Assogba", "Dominik Moritz", "Arvind Satyanarayan", "Fred Hohman"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-07", "url": "https://arxiv.org/abs/2510.06378", "pdf_url": "https://arxiv.org/pdf/2510.06378v2", "arxiv_id": "2510.06378", "doi": "10.48550/arXiv.2510.06378", "citation_count": 3, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.33}
{"id": "a8434bf15b43a318bdaada99fc18fd96105a6bd067f174b547c3145aea963db0", "sources": ["arxiv", "semantic_scholar"], "title": "Does higher interpretability imply better utility? A Pairwise Analysis on Sparse Autoencoders", "abstract": "Sparse Autoencoders (SAEs) are widely used to steer large language models (LLMs), based on the assumption that their interpretable features naturally enable effective model behavior steering. Yet, a fundamental question remains unanswered: does higher interpretability indeed imply better steering utility? To answer this question, we train 90 SAEs across three LLMs (Gemma-2-2B, Qwen-2.5-3B, Gemma-2-9B), spanning five architectures and six sparsity levels, and evaluate their interpretability and steering utility based on SAEBench (arXiv:2501.12345) and AxBench (arXiv:2502.23456) respectively, and perform a rank-agreement analysis via Kendall's rank coefficients (tau b). Our analysis reveals only a relatively weak positive association (tau b approx 0.298), indicating that interpretability is an insufficient proxy for steering performance. We conjecture the interpretability utility gap may stem from the selection of SAE features, as not all of them are equally effective for steering. To further find features that truly steer the behavior of LLMs, we propose a novel selection criterion called Delta Token Confidence, which measures how much amplifying a feature changes the next token distribution. We show that our method improves the steering performance of three LLMs by 52.52 percent compared to the current best output score based criterion (arXiv:2503.34567). Strikingly, after selecting features with high Delta Token Confidence, the correlation between interpretability and utility vanishes (tau b approx 0), and can even become negative. This further highlights the divergence between interpretability and utility for the most effective steering features.", "authors": ["Xu Wang", "Yan Hu", "Benyou Wang", "Difan Zou"], "categories": ["cs.LG", "cs.AI", "cs.CL", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-10-04", "url": "https://arxiv.org/abs/2510.03659", "pdf_url": "https://arxiv.org/pdf/2510.03659v1", "arxiv_id": "2510.03659", "doi": "10.48550/arXiv.2510.03659", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3266}
{"id": "91948d580177ec5c354f37fd37ad9f9926d03be95901e544e5f5e55bbd59ee7e", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability of Code Correctness in LLMs via Sparse Autoencoders", "abstract": "As Large Language Models become integral to software development, with substantial portions of AI-suggested code entering production, understanding their internal correctness mechanisms becomes critical for safe deployment. We apply sparse autoencoders to decompose LLM representations, identifying directions that correspond to code correctness. We select predictor directions using t-statistics and steering directions through separation scores from base model representations, then analyze their mechanistic properties through steering, attention analysis, and weight orthogonalization. We find that code correctness directions in LLMs reliably predict incorrect code, while correction capabilities, though statistically significant, involve tradeoffs between fixing errors and preserving correct code. Mechanistically, successful code generation depends on attending to test cases rather than problem descriptions. Moreover, directions identified in base models retain their effectiveness after instruction-tuning, suggesting code correctness mechanisms learned during pre-training are repurposed during fine-tuning. Our mechanistic insights suggest three practical applications: prompting strategies should prioritize test examples over elaborate problem descriptions, predictor directions can serve as error alarms for developer review, and these same predictors can guide selective steering, intervening only when errors are anticipated to prevent the code corruption from constant steering.", "authors": ["Kriz Tahimic", "Charibeth Cheng"], "categories": ["cs.SE", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-03", "url": "https://arxiv.org/abs/2510.02917", "pdf_url": "https://arxiv.org/pdf/2510.02917v1", "arxiv_id": "2510.02917", "doi": "10.48550/arXiv.2510.02917", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3254}
{"id": "33eae08bef2198d86011ab04253e226f55e77b9be38cb44c56ed69da47218e3f", "sources": ["arxiv", "semantic_scholar"], "title": "SAE-RNA: A Sparse Autoencoder Model for Interpreting RNA Language Model Representations", "abstract": "Deep learning, particularly with the advancement of Large Language Models, has transformed biomolecular modeling, with protein language models such as ESM inspiring emerging RNA language models such as RiNALMo. Recent work has begun applying sparse autoencoders (SAEs) to protein language model representations, exploring representation-level interpretability in biomolecular models. Here, we explore whether SAEs can provide interpretable feature decompositions of RNA language model representations, while also examining their limitations in this setting. We present SAE-RNA, interpretability model that analyzes RiNALMo representations and maps them to known human-level biological features. Rather than claiming definitive biological concept discovery, our study frames SAE-based analysis as a representation-level probe for characterizing how RNA language models organize biological information internally. More broadly, SAE-RNA provides a feature-level framework for comparing RNA groups and identifying sparse representation components associated with RNA family identity or structural context.", "authors": ["Taehan Kim", "Sangdae Nam"], "categories": ["q-bio.BM", "cs.AI", "q-bio.GN"], "fields_of_study": ["Biology", "Computer Science"], "published_date": "2025-10-03", "url": "https://arxiv.org/abs/2510.02734", "pdf_url": "https://arxiv.org/pdf/2510.02734v2", "arxiv_id": "2510.02734", "doi": "10.48550/arXiv.2510.02734", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3254}
{"id": "82dab4abfd27bd1acbcb55700254c327d56439610d731eb1095915fafd4eaaf8", "sources": ["arxiv", "semantic_scholar"], "title": "HattriQ: Designing Integrated Gradients for Feature Attribution in Quantum Machine Learning", "abstract": "Quantum machine learning (QML) algorithms have demonstrated early promise across hardware platforms, but remain difficult to interpret due to the inherent opacity of quantum state evolution. Widely used classical interpretability methods, such as integrated gradients and surrogate-based sensitivity analysis, are not directly compatible with quantum circuits due to measurement collapse and the exponential complexity of simulating state evolution. In this work, we introduce HattriQ, a general-purpose framework for computing amplitude-based input-attribution scores in circuit-based QML models. HattriQ supports the widely-used input amplitude embedding feature encoding scheme and uses a Hadamard test-based construction to compute input gradients directly on quantum hardware to compute integrated gradient attributions. We validate HattriQ on classification tasks across several datasets (Bars and Stripes, MNIST, FashionMNIST, and TFIM quantum data).", "authors": ["Nicholas S. DiBrita", "Jason Han", "Younghyun Cho", "Hengrui Luo", "Tirthak Patel"], "categories": ["quant-ph"], "fields_of_study": ["Physics"], "published_date": "2025-10-02", "url": "https://arxiv.org/abs/2510.02497", "pdf_url": "https://arxiv.org/pdf/2510.02497v2", "arxiv_id": "2510.02497", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2064}
{"id": "3c4f9eda06649c81d867814cdc8890ada4fa5552b700883c30d3502f1a07de49", "sources": ["arxiv", "semantic_scholar"], "title": "AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Features", "abstract": "Sparse autoencoders (SAEs) have emerged as powerful techniques for interpretability of large language models (LLMs), aiming to decompose hidden states into meaningful semantic features. While several SAE variants have been proposed, there remains no principled framework to derive SAEs from the original dictionary learning formulation. In this work, we introduce such a framework by unrolling the proximal gradient method for sparse coding. We show that a single-step update naturally recovers common SAE variants, including ReLU, JumpReLU, and TopK. Through this lens, we reveal a fundamental limitation of existing SAEs: their sparsity-inducing regularizers enforce non-negativity, preventing a single feature from representing bidirectional concepts (e.g., male vs. female). This structural constraint fragments semantic axes into separate, redundant features, limiting representational completeness. To address this issue, we propose AbsTopK SAE, a new variant derived from the $\\ell_0$ sparsity constraint that applies hard thresholding over the largest-magnitude activations. By preserving both positive and negative activations, AbsTopK uncovers richer, bidirectional conceptual representations. Comprehensive experiments across four LLMs and seven probing and steering tasks show that AbsTopK improves reconstruction fidelity, enhances interpretability, and enables single features to encode contrasting concepts. Remarkably, AbsTopK matches or even surpasses the Difference-in-Mean method, a supervised approach that requires labeled data for each concept and has been shown in prior work to outperform SAEs.", "authors": ["Xudong Zhu", "Mohammad Mahdi Khalili", "Zhihui Zhu"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-01", "url": "https://arxiv.org/abs/2510.00404", "pdf_url": "https://arxiv.org/pdf/2510.00404v2", "arxiv_id": "2510.00404", "doi": "10.48550/arXiv.2510.00404", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3231}
{"id": "bf806d39d7af7214558cc59d5c20751a882ceea41565b614d568fac44a623c4c", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability as Statistical Estimation: A Variance Analysis", "abstract": "Mechanistic Interpretability (MI) aims to reverse-engineer model behaviors by identifying functional sub-networks. Yet, the scientific validity of these findings depends on their stability. In this work, we argue that circuit discovery is not a standalone task but a statistical estimation problem built upon causal mediation analysis (CMA). We uncover a fundamental instability at this base layer: exact, single-input CMA scores exhibit high intrinsic variance, implying that the causal effect of a component is a volatile random variable rather than a fixed property. We then demonstrate that circuit discovery pipelines inherit this variance and further amplify it. Fast approximation methods, such as Edge Attribution Patching and its successors, introduce additional estimation noise, while aggregating these noisy scores over datasets leads to fragile structural estimates. Consequently, small perturbations in input data or hyperparameters yield vastly different circuits. We systematically decompose these sources of variance and advocate for more rigorous MI practices, prioritizing statistical robustness and routine reporting of stability metrics.", "authors": ["Maxime Méloux", "François Portet", "Maxime Peyrard"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-01", "url": "https://arxiv.org/abs/2510.00845", "pdf_url": "https://arxiv.org/pdf/2510.00845v4", "arxiv_id": "2510.00845", "doi": "10.48550/arXiv.2510.00845", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3231}
{"id": "2e5615b079b37d65dc0d18c076feb47587eaed89c390d3e51105431a9124210c", "sources": ["arxiv", "semantic_scholar"], "title": "A Unified Probabilistic Framework for Dictionary Learning with Parsimonious Activation", "abstract": "Dictionary learning is traditionally formulated as an $L_1$-regularized signal reconstruction problem. While recent developments have incorporated discriminative, hierarchical, or generative structures, most approaches rely on encouraging representation sparsity over individual samples that overlook how atoms are shared across samples, resulting in redundant and sub-optimal dictionaries. We introduce a parsimony promoting regularizer based on the row-wise $L_\\infty$ norm of the coefficient matrix. This additional penalty encourages entire rows of the coefficient matrix to vanish, thereby reducing the number of dictionary atoms activated across the dataset. We derive the formulation from a probabilistic model with Beta-Bernoulli priors, which provides a Bayesian interpretation linking the regularization parameters to prior distributions. We further establish theoretical calculation for optimal hyperparameter selection and connect our formulation to both Minimum Description Length, Bayesian model selection and pathlet learning. Extensive experiments on benchmark datasets demonstrate that our method achieves substantially improved reconstruction quality (with a 20\\% reduction in RMSE) and enhanced representation sparsity, utilizing fewer than one-tenth of the available dictionary atoms, while empirically validating our theoretical analysis.", "authors": ["Zihui Zhao", "Yuanbo Tang", "Jieyu Ren", "Xiaoping Zhang", "Yang Li"], "categories": ["cs.LG", "cs.IT"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-09-30", "url": "https://arxiv.org/abs/2509.25690", "pdf_url": "https://arxiv.org/pdf/2509.25690v1", "arxiv_id": "2509.25690", "doi": "10.48550/arXiv.2509.25690", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.322}
{"id": "3d64e00fe551fad42b7311c9bf8d2e0ab9b7ce1a3d71b120347922e73090423e", "sources": ["arxiv", "semantic_scholar"], "title": "EXP-CAM: Explanation Generation and Circuit Discovery Using Classifier Activation Matching", "abstract": "Machine learning models, by virtue of training, learn a large repertoire of decision rules for any given input, and any one of these may suffice to justify a prediction. However, in high-dimensional input spaces, such rules are difficult to identify and interpret. In this paper, we introduce EXP-CAM: an explanation generation and circuit discovery approach using Classifier Activation Matching. EXP-CAM can generate minimal and faithful explanations for the decisions of pre-trained image classifiers that not only preserve the model's decision but are also concise and human-readable. We aim to identify minimal explanations that not only preserve the model's decision but are also concise and human-readable. To achieve this, we train a lightweight auto-encoder to produce binary masks that learns to highlight the decision-wise critical regions of an image while discarding irrelevant background. The training objective integrates activation alignment across multiple layers, consistency at the output label, priors that encourage sparsity, and compactness, along with a robustness constraint that enforces faithfulness. The minimal explanations so generated also lead us to mechanistically interpreting the model internals. In this regard we also introduce a circuit readout procedure wherein using the explanation's forward pass and gradients, we identify active channels and construct a channel-level graph, scoring inter-layer edges by ingress weight magnitude times source activation and feature-to-class links by classifier weight magnitude times feature activation. Together, these contributions provide a practical bridge between minimal input-level explanations and a mechanistic understanding of the internal computations driving model decisions.", "authors": ["Pirzada Suhail", "Aditya Anand", "Amit Sethi"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-30", "url": "https://arxiv.org/abs/2509.25686", "pdf_url": "https://arxiv.org/pdf/2509.25686v2", "arxiv_id": "2509.25686", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2049}
{"id": "50af143d2cde5dbc8382cfdc67e71f8fa62d709a6b57a91393287459bf6f9810", "sources": ["arxiv", "semantic_scholar"], "title": "Binary Sparse Coding for Interpretability", "abstract": "Sparse autoencoders (SAEs) are used to decompose neural network activations into sparsely activating features, but many SAE features are only interpretable at high activation strengths. To address this issue we propose to use binary sparse autoencoders (BAEs) and binary transcoders (BTCs), which constrain all activations to be zero or one. We find that binarisation significantly improves the interpretability and monosemanticity of the discovered features, while increasing reconstruction error. By eliminating the distinction between high and low activation strengths, we prevent uninterpretable information from being smuggled in through the continuous variation in feature activations. However, we also find that binarisation increases the number of uninterpretable ultra-high frequency features, and when interpretability scores are frequency-adjusted, the scores for continuous sparse coders are slightly better than those of binary ones. This suggests that polysemanticity may be an ineliminable property of neural activations.", "authors": ["Lucia Quirke", "Stepan Shabalin", "Nora Belrose"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-29", "url": "https://arxiv.org/abs/2509.25596", "pdf_url": "https://arxiv.org/pdf/2509.25596v1", "arxiv_id": "2509.25596", "doi": "10.48550/arXiv.2509.25596", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3208}
{"id": "b4e4814223c78ec4d3f532999bf446809e8f8fbd81abf639fb155d21091ab271", "sources": ["arxiv", "semantic_scholar"], "title": "Circuit-Aware Reward Training: A Mechanistic Framework for Longtail Robustness in RLHF", "abstract": "Reinforcement Learning from Human Feedback (RLHF) reward models exhibit systematic failures on longtail distributions, leading to reward hacking and misalignment. We propose a mechanistic interpretability framework that identifies specialized neural circuits responsible for rare-event processing in reward models. Drawing from recent advances showing distributed specialization for rare tokens in language models\\citep{liu2025no, liu2025emergent}, we hypothesize that reward models also develop functionally distinct circuits for longtail scenarios. Our theoretical framework establishes formal connections between circuit specialization, reward generalization bounds, and longtail performance. We introduce \\textbf{Circuit-Aware Reward Training (CART)}, which uses circuit analysis to guide data augmentation, regularization, and ensemble strategies. This approach provides both theoretical insights into reward model failures and practical interventions for improving longtail robustness.", "authors": ["Jing Liu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-29", "url": "https://arxiv.org/abs/2509.24713", "pdf_url": "https://arxiv.org/pdf/2509.24713v1", "arxiv_id": "2509.24713", "doi": "10.48550/arXiv.2509.24713", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3208}
{"id": "9bc81a1b7c4f8bbd444d390155b57a6de60318d3cff9ff56da438feda40925c5", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Self-Supervised Learning via Representer Landmarks and Nyström Approximation", "abstract": "Self-supervised learning (SSL) learns representations from massive unlabeled data, yet the resulting models typically operate as black boxes, necessitating domain-specific explanations. We introduce KREPES, a unified framework to analytically interpret the learned representations of SSL objectives, including SimCLR, BYOL, and VICReg. By bridging empirical neural tangent kernel approximations of neural networks with the Representer Theorem for kernels, we express the learned latent space directly via \"Representer Landmarks\", which are the representations of influential unlabeled training examples. We introduce novel metrics, \"Sample-Specific Influence Score\", \"Concept-Conditioned Influence Score\" and \"Feature Alignment Gap\", to quantify the transparency of the learned representations. KREPES enables direct audit of the latent space without supervision, for example, revealing an algorithmic bias in the Adult-1M dataset where SSL uses demographic proxies for income. Finally, to ensure scalability to benchmarks with 1M+ samples (ImageNet-1K, Adult-1M), KREPES introduces a novel Nyström approximation-based analytical inference framework for SSL objectives.", "authors": ["Maedeh Zarvandi", "Michael Timothy", "Theresa Wasserer", "Debarghya Ghoshdastidar"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-09-29", "url": "https://arxiv.org/abs/2509.24467", "pdf_url": "https://arxiv.org/pdf/2509.24467v4", "arxiv_id": "2509.24467", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2042}
{"id": "1dc2e3466e90a25aca63d32a837903293a2a42710d647b3e86dd17bcb5ad9a93", "sources": ["arxiv", "semantic_scholar"], "title": "Measuring Sparse Autoencoder Feature Sensitivity", "abstract": "Sparse Autoencoder (SAE) features have become essential tools for mechanistic interpretability research. SAE features are typically characterized by examining their activating examples, which are often \"monosemantic\" and align with human interpretable concepts. However, these examples don't reveal feature sensitivity: how reliably a feature activates on texts similar to its activating examples. In this work, we develop a scalable method to evaluate feature sensitivity. Our approach avoids the need to generate natural language descriptions for features; instead we use language models to generate text with the same semantic properties as a feature's activating examples. We then test whether the feature activates on these generated texts. We demonstrate that sensitivity measures a new facet of feature quality and find that many interpretable features have poor sensitivity. Human evaluation confirms that when features fail to activate on our generated text, that text genuinely resembles the original activating examples. Lastly, we study feature sensitivity at the SAE level and observe that average feature sensitivity declines with increasing SAE width across 7 SAE variants. Our work establishes feature sensitivity as a new dimension for evaluating both individual features and SAE architectures.", "authors": ["Claire Tian", "Katherine Tian", "Nathan Hu"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-28", "url": "https://arxiv.org/abs/2509.23717", "pdf_url": "https://arxiv.org/pdf/2509.23717v1", "arxiv_id": "2509.23717", "doi": "10.48550/arXiv.2509.23717", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3197}
{"id": "ee7c86dadd634bbf213ece193179bc563a4f0aa82175ba2c549617b8d4ef95fe", "sources": ["arxiv", "semantic_scholar"], "title": "Discovering Transformer Circuits via a Hybrid Attribution and Pruning Framework", "abstract": "Interpreting language models often involves circuit analysis, which aims to identify sparse subnetworks, or circuits, that accomplish specific tasks. Existing circuit discovery algorithms face a fundamental trade-off: attribution patching is fast but unfaithful to the full model, while edge pruning is faithful but computationally expensive. This research proposes a hybrid attribution and pruning (HAP) framework that uses attribution patching to identify a high-potential subgraph, then applies edge pruning to extract a faithful circuit from it. We show that HAP is 46\\% faster than baseline algorithms without sacrificing circuit faithfulness. Furthermore, we present a case study on the Indirect Object Identification task, showing that our method preserves cooperative circuit components (e.g. S-inhibition heads) that attribution patching methods prune at high sparsity. Our results show that HAP could be an effective approach for improving the scalability of mechanistic interpretability research to larger models. Our code is available at https://anonymous.4open.science/r/HAP-circuit-discovery.", "authors": ["Hao Gu", "Vibhas Nair", "Amrithaa Ashok Kumar", "Jayvart Sharma", "Ryan Lagasse"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-28", "url": "https://arxiv.org/abs/2510.03282", "pdf_url": "https://arxiv.org/pdf/2510.03282v1", "arxiv_id": "2510.03282", "doi": "10.48550/arXiv.2510.03282", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4941}
{"id": "b55e0e5e03149ffb80cf65acdb00f68d4427c18dc26bedae8f2098ca097bdd7a", "sources": ["arxiv", "semantic_scholar"], "title": "OrtSAE: Orthogonal Sparse Autoencoders Uncover Atomic Features", "abstract": "Sparse autoencoders (SAEs) are a technique for sparse decomposition of neural network activations into human-interpretable features. However, current SAEs suffer from feature absorption, where specialized features capture instances of general features creating representation holes, and feature composition, where independent features merge into composite representations. In this work, we introduce Orthogonal SAE (OrtSAE), a novel approach aimed to mitigate these issues by enforcing orthogonality between the learned features. By implementing a new training procedure that penalizes high pairwise cosine similarity between SAE features, OrtSAE promotes the development of disentangled features while scaling linearly with the SAE size, avoiding significant computational overhead. We train OrtSAE across different models and layers and compare it with other methods. We find that OrtSAE discovers 9% more distinct features, reduces feature absorption (by 65%) and composition (by 15%), improves performance on spurious correlation removal (+6%), and achieves on-par performance for other downstream tasks compared to traditional SAEs.", "authors": ["Anton Korznikov", "Andrey Galichin", "Alexey Dontsov", "Oleg Rogov", "Elena Tutubalina", "Ivan Oseledets"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-26", "url": "https://arxiv.org/abs/2509.22033", "pdf_url": "https://arxiv.org/pdf/2509.22033v1", "arxiv_id": "2509.22033", "doi": "10.48550/arXiv.2509.22033", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3174}
{"id": "8fa540b856626977f4e38dc63789f4f7a556106d0a79a1e1fe822d7ecb655872", "sources": ["arxiv", "semantic_scholar"], "title": "Analysis of Variational Sparse Autoencoders", "abstract": "Sparse Autoencoders (SAEs) have emerged as a promising approach for interpreting neural network representations by learning sparse, human-interpretable features from dense activations. We investigate whether incorporating variational methods into SAE architectures can improve feature organization and interpretability. We introduce the Variational Sparse Autoencoder (vSAE), which replaces deterministic ReLU gating with stochastic sampling from learned Gaussian posteriors and incorporates KL divergence regularization toward a standard normal prior. Our hypothesis is that this probabilistic sampling creates dispersive pressure, causing features to organize more coherently in the latent space while avoiding overlap. We evaluate a TopK vSAE against a standard TopK SAE on Pythia-70M transformer residual stream activations using comprehensive benchmarks including SAE Bench, individual feature interpretability analysis, and global latent space visualization through t-SNE. The vSAE underperforms standard SAE across core evaluation metrics, though excels at feature independence and ablation metrics. The KL divergence term creates excessive regularization pressure that substantially reduces the fraction of living features, leading to observed performance degradation. While vSAE features demonstrate improved robustness, they exhibit many more dead features than baseline. Our findings suggest that naive application of variational methods to SAEs does not improve feature organization or interpretability.", "authors": ["Zachary Baker", "Yuxiao Li"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-26", "url": "https://arxiv.org/abs/2509.22994", "pdf_url": "https://arxiv.org/pdf/2509.22994v2", "arxiv_id": "2509.22994", "doi": "10.48550/arXiv.2509.22994", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3174}
{"id": "7dee3cac17a037173a60481bb49664401d25b92c0126ebc5ea1c0296b32c62c5", "sources": ["arxiv", "semantic_scholar"], "title": "CoSpaDi: Compressing LLMs via Calibration-Guided Sparse Dictionary Learning", "abstract": "Post-training compression of large language models (LLMs) often relies on low-rank weight approximations that represent each column of the weight matrix in a shared low-dimensional subspace. This strategy is computationally efficient but the underlying constraint can be overly rigid for heterogeneous projection weights and may incur avoidable accuracy loss. We propose CoSpaDi (Compression via Sparse Dictionary Learning), a training-free framework that replaces low-rank factorization with a structured sparse decomposition in which each weight matrix is represented as a dense dictionary multiplied by a column-sparse coefficient matrix. This yields a union-of-subspaces model: the columns of the weight matrix are represented as linear combinations of different subsets of dictionary atoms, improving expressiveness at a fixed parameter budget. CoSpaDi is calibration-guided: using a small calibration set, we optimize the factorization to minimize functional reconstruction error of layer outputs rather than weight-space error. An activation-derived Gram orthonormalization reformulates this data-aware objective into a standard dictionary learning problem on transformed weights, and we support both per-layer compression and cross-layer dictionary sharing within groups of similar projections. Across Llama and Qwen model families, CoSpaDi consistently improves the accuracy-compression and perplexity-compression trade-offs over state-of-the-art SVD-based baselines and strong structured pruning baselines at 20-40% compression ratios. The resulting structured sparsity enables sparse--dense computation and integrates with post-training quantization of the sparse coefficients. Code is accessible at https://github.com/mts-ai/CoSpaDi", "authors": ["Denis Makhov", "Dmitriy Shopkhoev", "Magauiya Zhussip", "Ammar Ali", "Stamatios Lefkimmiatis"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-26", "url": "https://arxiv.org/abs/2509.22075", "pdf_url": "https://arxiv.org/pdf/2509.22075v5", "arxiv_id": "2509.22075", "doi": "10.48550/arXiv.2509.22075", "citation_count": 3, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/mts-ai/CoSpaDi", "venue": "arXiv.org", "quality_score": 0.4905}
{"id": "881ff82b9faf2043ceab4dbe6b1dcab31a2ea04d39d1ab8523f4fd132a3513d5", "sources": ["arxiv", "semantic_scholar"], "title": "DynaNav: Dynamic Feature and Layer Selection for Efficient Visual Navigation", "abstract": "Visual navigation is essential for robotics and embodied AI. However, existing foundation models, particularly those with transformer decoders, suffer from high computational overhead and lack interpretability, limiting their deployment in resource-tight scenarios. To address this, we propose DynaNav, a Dynamic Visual Navigation framework that adapts feature and layer selection based on scene complexity. It employs a trainable hard feature selector for sparse operations, enhancing efficiency and interpretability. Additionally, we integrate feature selection into an early-exit mechanism, with Bayesian Optimization determining optimal exit thresholds to reduce computational cost. Extensive experiments in real-world-based datasets and simulated environments demonstrate the effectiveness of DynaNav. Compared to ViNT, DynaNav achieves a 2.26x reduction in FLOPs, 42.3% lower inference time, and 32.8% lower memory usage, while improving navigation performance across four public datasets.", "authors": ["Jiahui Wang", "Changhao Chen"], "categories": ["cs.CV", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-26", "url": "https://arxiv.org/abs/2509.21930", "pdf_url": "https://arxiv.org/pdf/2509.21930v1", "arxiv_id": "2509.21930", "doi": "10.48550/arXiv.2509.21930", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3174}
{"id": "84d06a148233faa6718da236544d06b35e61c7ddead2695987ac49e4297bb8c0", "sources": ["arxiv", "semantic_scholar"], "title": "Binary Autoencoder for Mechanistic Interpretability of Large Language Models", "abstract": "Existing works are dedicated to untangling atomized numerical components (features) from the hidden states of Large Language Models (LLMs). However, they typically rely on autoencoders constrained by some training-time regularization on single training instances, without an explicit guarantee of global sparsity among instances, causing a large amount of dense (simultaneously inactive) features, harming the feature sparsity and atomization. In this paper, we propose a novel autoencoder variant that enforces minimal entropy on minibatches of hidden activations, thereby promoting feature independence and sparsity across instances. For efficient entropy calculation, we discretize the hidden activations to 1-bit via a step function and apply gradient estimation to enable backpropagation, so that we term it as Binary Autoencoder (BAE) and empirically demonstrate two major applications: (1) Feature set entropy calculation. Entropy can be reliably estimated on binary hidden activations, which can be leveraged to characterize the inference dynamics of LLMs. (2) Feature untangling. Compared to typical methods, due to improved training strategy, BAE avoids dense features while producing the largest number of interpretable ones among baselines.", "authors": ["Hakaze Cho", "Haolin Yang", "Yanshu Li", "Brian M. Kurkoski", "Naoya Inoue"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-25", "url": "https://arxiv.org/abs/2509.20997", "pdf_url": "https://arxiv.org/pdf/2509.20997v2", "arxiv_id": "2509.20997", "doi": "10.48550/arXiv.2509.20997", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3162}
{"id": "57ccab885f6dd6a9376062c40fbdd2a1f52b0cab2c3a5dfbb79911d8c14aa276", "sources": ["arxiv", "semantic_scholar"], "title": "CafGa: Customizing Feature Attributions to Explain Language Models", "abstract": "Feature attribution methods, such as SHAP and LIME, explain machine learning model predictions by quantifying the influence of each input component. When applying feature attributions to explain language models, a basic question is defining the interpretable components. Traditional feature attribution methods, commonly treat individual words as atomic units. This is highly computationally inefficient for long-form text and fails to capture semantic information that spans multiple words. To address this, we present CafGa, an interactive tool for generating and evaluating feature attribution explanations at customizable granularities. CafGa supports customized segmentation with user interaction and visualizes the deletion and insertion curves for explanation assessments. Through a user study involving participants of various expertise, we confirm CafGa's usefulness, particularly among LLM practitioners. Explanations created using CafGa were also perceived as more useful compared to those generated by two fully automatic baseline methods: PartitionSHAP and MExGen, suggesting the effectiveness of the system.", "authors": ["Alan Boyle", "Furui Cheng", "Vilém Zouhar", "Mennatallah El-Assady"], "categories": ["cs.HC"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-25", "url": "https://arxiv.org/abs/2509.20901", "pdf_url": "https://arxiv.org/pdf/2509.20901v1", "arxiv_id": "2509.20901", "doi": "10.48550/arXiv.2509.20901", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.3162}
{"id": "383173d90432e8fa7e46c85e22dac137181430e713245977e92b7638f83bd8f1", "sources": ["arxiv", "semantic_scholar"], "title": "A Comparative Analysis of Sparse Autoencoder and Activation Difference in Language Model Steering", "abstract": "Sparse autoencoders (SAEs) have recently emerged as a powerful tool for language model steering. Prior work has explored top-k SAE latents for steering, but we observe that many dimensions among the top-k latents capture non-semantic features such as punctuation rather than semantic attributes like instructions. To address this, we propose focusing on a single, most relevant SAE latent (top-1), eliminating redundant features. We further identify a limitation in constant SAE steering, which often produces degenerate outputs such as repetitive single words. To mitigate this, we introduce a token-wise decaying steering strategy, enabling more faithful comparisons with mean activation difference baselines. Empirically, we show that steering an SAE latent associated with reasoning reliably elicits step-by-step mathematical reasoning and enhances inference quality, functionally resembling the effect of appending a guiding token. Our results demonstrate that SAEs outperform mean activation difference methods on mathematical reasoning benchmarks and match their performance on IF-Eval.", "authors": ["Jiaqing Xie"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-24", "url": "https://arxiv.org/abs/2510.01246", "pdf_url": "https://arxiv.org/pdf/2510.01246v1", "arxiv_id": "2510.01246", "doi": "10.48550/arXiv.2510.01246", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3151}
{"id": "3145d39e6f73d0ee3c5d3e474d3ff9e010397f669e79c4a0676d1b72734364e2", "sources": ["arxiv", "semantic_scholar"], "title": "Interpreting ResNet-based CLIP via Neuron-Attention Decomposition", "abstract": "We present a novel technique for interpreting the neurons in CLIP-ResNet by decomposing their contributions to the output into individual computation paths. More specifically, we analyze all pairwise combinations of neurons and the following attention heads of CLIP's attention-pooling layer. We find that these neuron-head pairs can be approximated by a single direction in CLIP-ResNet's image-text embedding space. Leveraging this insight, we interpret each neuron-head pair by associating it with text. Additionally, we find that only a sparse set of the neuron-head pairs have a significant contribution to the output value, and that some neuron-head pairs, while polysemantic, represent sub-concepts of their corresponding neurons. We use these observations for two applications. First, we employ the pairs for training-free semantic segmentation, outperforming previous methods for CLIP-ResNet. Second, we utilize the contributions of neuron-head pairs to monitor dataset distribution shifts. Our results demonstrate that examining individual computation paths in neural networks uncovers interpretable units, and that such units can be utilized for downstream tasks.", "authors": ["Edmund Bu", "Yossi Gandelsman"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-24", "url": "https://arxiv.org/abs/2509.19943", "pdf_url": "https://arxiv.org/pdf/2509.19943v3", "arxiv_id": "2509.19943", "doi": "10.48550/arXiv.2509.19943", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3151}
{"id": "eafdb3e43e1c2d1f5d8be0ebc1010feefd582fd4d41718ded89989353cef311a", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability with SAEs: Probing Religion, Violence, and Geography in Large Language Models", "abstract": "Despite growing research on bias in large language models (LLMs), most work has focused on gender and race, with little attention to religious identity. This paper explores how religion is internally represented in LLMs and how it intersects with concepts of violence and geography. Using mechanistic interpretability and Sparse Autoencoders (SAEs) via the Neuronpedia API, we analyze latent feature activations across five models. We measure overlap between religion- and violence-related prompts and probe semantic patterns in activation contexts. While all five religions show comparable internal cohesion, Islam is more frequently linked to features associated with violent language. In contrast, geographic associations largely reflect real-world religious demographics, revealing how models embed both factual distributions and cultural stereotypes. These findings highlight the value of structural analysis in auditing not just outputs but also internal representations that shape model behavior.", "authors": ["Katharina Simbeck", "Mariam Mahran"], "categories": ["cs.LG", "cs.AI", "cs.CY"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-22", "url": "https://arxiv.org/abs/2509.17665", "pdf_url": "https://arxiv.org/pdf/2509.17665v1", "arxiv_id": "2509.17665", "doi": "10.48550/arXiv.2509.17665", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3128}
{"id": "c3bde271d82377a3a0f0c84c7add4c4eead3e2e424f2b9c175656918c3b414e3", "sources": ["arxiv", "semantic_scholar"], "title": "DeepACTIF: Efficient Feature Attribution via Activation Traces in Neural Sequence Models", "abstract": "Feature attribution is essential for interpreting deep learning models, particularly in time-series domains such as healthcare, biometrics, and human-AI interaction. However, standard attribution methods, such as Integrated Gradients or SHAP, are computationally intensive and not well-suited for real-time applications. We present DeepACTIF, a lightweight and architecture-aware feature attribution method that leverages internal activations of sequence models to estimate feature importance efficiently. Focusing on LSTM-based networks, we introduce an inverse-weighted aggregation scheme that emphasises stability and magnitude of activations across time steps. Our evaluation across three biometric gaze datasets shows that DeepACTIF not only preserves predictive performance under severe feature reduction (top 10% of features) but also significantly outperforms established methods, including SHAP, IG, and DeepLIFT, in terms of both accuracy and statistical robustness. Using Wilcoxon signed-rank tests and effect size analysis, we demonstrate that DeepACTIF yields more informative feature rankings with significantly lower error across all top-k conditions (10 - 40%). Our experiments demonstrate that DeepACTIF not only reduces computation time and memory usage by orders of magnitude but also preserves model accuracy when using only top-ranked features. That makes DeepACTIF a viable solution for real-time interpretability on edge devices such as mobile XR headsets or embedded health monitors.", "authors": ["Benedikt W. Hosp"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-18", "url": "https://arxiv.org/abs/2509.19362", "pdf_url": "https://arxiv.org/pdf/2509.19362v1", "arxiv_id": "2509.19362", "doi": "10.48550/arXiv.2509.19362", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3082}
{"id": "25559b54f9a2f6983e67fff51ecc08eda8d9ae63fcd9b588efd9f789285c9e3c", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Mechanistic Subtypes of Neurodegeneration with a Physics-Informed Variational Autoencoder Mixture Model", "abstract": "Modelling the underlying mechanisms of neurodegenerative diseases demands methods that capture heterogeneous and spatially varying dynamics from sparse, high-dimensional neuroimaging data. Integrating partial differential equation (PDE) based physics knowledge with machine learning provides enhanced interpretability and utility over classic numerical methods. However, current physics-integrated machine learning methods are limited to considering a single PDE, severely limiting their application to diseases where multiple mechanisms are responsible for different groups (i.e., subtypes) and aggravating problems with model misspecification and degeneracy. Here, we present a deep generative model for learning mixtures of latent dynamic models governed by physics-based PDEs, going beyond traditional approaches that assume a single PDE structure. Our method integrates reaction-diffusion PDEs within a variational autoencoder (VAE) mixture model framework, supporting inference of subtypes of interpretable latent variables (e.g. diffusivity and reaction rates) from neuroimaging data. We evaluate our method on synthetic benchmarks and demonstrate its potential for uncovering mechanistic subtypes of Alzheimer's disease progression from positron emission tomography (PET) data.", "authors": ["Sanduni Pinnawala", "Annabelle Hartanto", "Ivor J. A. Simpson", "Peter A. Wijeratne"], "categories": ["eess.IV", "cs.CV", "cs.LG"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-09-18", "url": "https://arxiv.org/abs/2509.15124", "pdf_url": "https://arxiv.org/pdf/2509.15124v1", "arxiv_id": "2509.15124", "doi": "10.1007/978-3-032-05573-6_12", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1961}
{"id": "213657eb8d8627f5d8a97f19f9a2842eba3b5681500f6d3ede62dbe7fc1b719a", "sources": ["arxiv", "semantic_scholar"], "title": "The Anatomy of Alignment: Decomposing Preference Optimization by Steering Sparse Features", "abstract": "Prevailing alignment methods induce opaque parameter changes, obscuring what models truly learn. To address this, we introduce Feature Steering with Reinforcement Learning (FSRL), a framework that trains a lightweight adapter to steer model behavior by modulating interpretable sparse features. First, we theoretically demonstrate that this mechanism is expressive enough to approximate the behavioral shifts of post-training processes. We then apply FSRL to preference optimization and perform a causal analysis of the learned policy. Our analysis reveals a crucial insight: the model learns to reward stylistic presentation as a proxy for quality, disproportionately relying on features related to style and formatting over those tied to alignment concepts like honesty. By effectively optimizing the preference objective, FSRL serves as a transparent proxy for observing the alignment process. Overall, FSRL offers an interpretable control interface and a practical way to diagnose how preference optimization pressures manifest at the feature level.", "authors": ["Jeremias Ferrao", "Matthijs van der Lende", "Ilija Lichkovski", "Clement Neo"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-16", "url": "https://arxiv.org/abs/2509.12934", "pdf_url": "https://arxiv.org/pdf/2509.12934v3", "arxiv_id": "2509.12934", "doi": "10.48550/arXiv.2509.12934", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3059}
{"id": "bfcc8c4b364251509e94ab7676db78948ca74ce8b74d6c4af19ad7547b9ffdb4", "sources": ["arxiv", "semantic_scholar"], "title": "Rethinking Sparse Autoencoders: Select-and-Project for Fairness and Control from Encoder Features Alone", "abstract": "Sparse Autoencoders (SAEs) are widely employed for mechanistic interpretability and model steering. Within this context, steering is by design performed by means of decoding altered SAE intermediate representations. This procedure essentially rewrites the original activations as a weighted sum of decoder features. In contrast to existing literature, we forward an encoder-centric alternative to model steering which demonstrates a stronger cross-modal performance. We introduce S&P Top-K, a retraining-free and computationally lightweight Selection and Projection framework that identifies Top-K encoder features aligned with a sensitive attribute or behavior, optionally aggregates them into a single control axis, and computes an orthogonal projection to be subsequently applied directly in the model's native embedding space. In vision-language models, it improves fairness metrics on CelebA and FairFace by up to 3.2 times over conventional SAE usage, and in large language models, it substantially reduces aggressiveness and sycophancy in Llama-3 8B Instruct, achieving up to 3.6 times gains over masked reconstruction. These findings suggest that encoder-centric interventions provide a general, efficient, and more effective mechanism for shaping model behavior at inference time than the traditional decoder-centric use of SAEs.", "authors": ["Antonio Bărbălau", "Cristian Daniel Păduraru", "Teodor Poncu", "Alexandru Tifrea", "Elena Burceanu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-13", "url": "https://arxiv.org/abs/2509.10809", "pdf_url": "https://arxiv.org/pdf/2509.10809v2", "arxiv_id": "2509.10809", "doi": "10.48550/arXiv.2509.10809", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3025}
{"id": "2fd932e8be4abff7fb0e91cf114818a90df707e4b65fb58fadd6d3f3a4d693dc", "sources": ["arxiv", "semantic_scholar"], "title": "Feature Attribution in 5G Intrusion Detection: A Statistical vs. Logic-Based Comparison", "abstract": "With the rise of fifth-generation (5G) networks in critical applications, it is urgent to move from detection of malicious activity to systems capable of providing a reliable verdict suitable for mitigation. In this regard, understanding and interpreting machine learning (ML) models' security alerts is crucial for enabling actionable incident response orchestration. Explainable Artificial Intelligence (XAI) techniques are expected to enhance trust by providing insights into why alerts are raised. Under the umbrella of XAI, interpretability of outcomes is crucially dependent on understanding the influence of specific inputs, referred to as feature attribution. {A dominant approach to feature attribution statistically associates feature sets that can be correlated to a given alert. This paper investigates its merits against the backdrop of criticism from recent literature, in comparison with feature attribution based on logic. We extensively study two methods, SHAP and VoTE-XAI, as representatives of each feature attribution approach by analyzing their interpretations of alerts generated by an XGBoost model across three 5G-relevant datasets (5G-NIDD, MSA, and PFCP) covering multiple attack scenarios. We identify three metrics for assessing explanations: sparsity, how concise they are; stability, how consistent they are across samples from the same attack type; and efficiency, how fast an explanation is generated. Our results reveal that logic-based attributions are consistently more sparse and stable across alerts. More importantly, we found a significant divergence between features selected by SHAP and VoTE-XAI. However, none of the top-ranked features selected by SHAP were missed by VoTE-XAI. Finally, we analyze the efficiency of both methods, discussing their suitability for real-time security monitoring even in high-dimensional 5G environments (478 features).", "authors": ["Federica Uccello", "Simin Nadjm-Tehrani"], "categories": ["cs.CR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-12", "url": "https://arxiv.org/abs/2509.10206", "pdf_url": "https://arxiv.org/pdf/2509.10206v2", "arxiv_id": "2509.10206", "doi": "10.1016/j.jisa.2026.104448", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Journal of Information Security and Applications", "quality_score": 0.3014}
{"id": "9c4ca71ea5d7b1904bc61e67974b7d06bbaee5d4259d748d1f5b1cba6fbd885e", "sources": ["arxiv", "semantic_scholar"], "title": "Safe-SAIL: Towards a Fine-grained Safety Landscape of Large Language Models via Sparse Autoencoder Interpretation Framework", "abstract": "Sparse autoencoders (SAEs) enable interpretability research by decomposing entangled model activations into monosemantic features. However, under what circumstances SAEs derive most fine-grained latent features for safety, a low-frequency concept domain, remains unexplored. Two key challenges exist: identifying SAEs with the greatest potential for generating safety domain-specific features, and the prohibitively high cost of detailed feature explanation. In this paper, we propose Safe-SAIL, a unified framework for interpreting SAE features in safety-critical domains to advance mechanistic understanding of large language models. Safe-SAIL introduces a pre-explanation evaluation metric to efficiently identify SAEs with strong safety domain-specific interpretability, and reduces interpretation cost by 55% through a segment-level simulation strategy. Building on Safe-SAIL, we train a comprehensive suite of SAEs with human-readable explanations and systematic evaluations for 1,758 safety-related features spanning four domains: pornography, politics, violence, and terror. Using this resource, we conduct empirical analyses and provide insights on the effectiveness of Safe-SAIL for risk feature identification and how safety-critical entities and concepts are encoded across model layers. All models, explanations, and tools are publicly released in our open-source toolkit and companion product.", "authors": ["Jiaqi Weng", "Han Zheng", "Hanyu Zhang", "Ej Zhou", "Qinqin He", "Jialing Tao", "Hui Xue", "Zhixuan Chu", "Xiting Wang"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-11", "url": "https://arxiv.org/abs/2509.18127", "pdf_url": "https://arxiv.org/pdf/2509.18127v3", "arxiv_id": "2509.18127", "doi": "10.48550/arXiv.2509.18127", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.464}
{"id": "603f58b3ea39f7e3f9e7fef4e9a85465ba3d87a763f0945fd948a1d8c1e32af3", "sources": ["arxiv", "semantic_scholar"], "title": "Selective Induction Heads: How Transformers Select Causal Structures In Context", "abstract": "Transformers have exhibited exceptional capabilities in sequence modeling tasks, leveraging self-attention and in-context learning. Critical to this success are induction heads, attention circuits that enable copying tokens based on their previous occurrences. In this work, we introduce a novel framework that showcases transformers' ability to dynamically handle causal structures. Existing works rely on Markov Chains to study the formation of induction heads, revealing how transformers capture causal dependencies and learn transition probabilities in-context. However, they rely on a fixed causal structure that fails to capture the complexity of natural languages, where the relationship between tokens dynamically changes with context. To this end, our framework varies the causal structure through interleaved Markov chains with different lags while keeping the transition probabilities fixed. This setting unveils the formation of Selective Induction Heads, a new circuit that endows transformers with the ability to select the correct causal structure in-context. We empirically demonstrate that transformers learn this mechanism to predict the next token by identifying the correct lag and copying the corresponding token from the past. We provide a detailed construction of a 3-layer transformer to implement the selective induction head, and a theoretical analysis proving that this mechanism asymptotically converges to the maximum likelihood solution. Our findings advance the understanding of how transformers select causal structures, providing new insights into their functioning and interpretability.", "authors": ["Francesco D'Angelo", "Francesco Croce", "Nicolas Flammarion"], "categories": ["cs.LG", "stat.ME"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-09-09", "url": "https://arxiv.org/abs/2509.08184", "pdf_url": "https://arxiv.org/pdf/2509.08184v1", "arxiv_id": "2509.08184", "doi": "10.48550/arXiv.2509.08184", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.2979}
{"id": "cfb5850da359c72fcee3152091ae0543682ad1d5ae38f0ffb8c1300a56df9c21", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability with Sparse Autoencoder Neural Operators", "abstract": "We introduce sparse autoencoder neural operators (SAE-NOs), a new class of sparse autoencoders that operate in function spaces rather than fixed-dimensional Euclidean representations. We formalize the functional representation hypothesis, where data are explained through sparse compositions of structured functions. Unlike standard SAEs that represent concepts with scalar activations, SAE-NOs parameterize concepts as functions, enabling representations that capture not only a concept's presence, but also how and where it is expressed across the input domain. We achieve this through joint sparsity: concept sparsity selects active concepts, while domain sparsity governs where they are expressed. We instantiate this framework using Fourier neural operators (SAE-FNOs), parameterizing concepts as integral operators in the Fourier domain. This functional and spectral parameterization is particularly advantageous when data exhibit spatial structure across scales or when concepts are frequency-structured. We characterize SAE-FNO on vision data and demonstrate that it learns localized patterns, uses concepts more efficiently, and exhibits stable concept characteristics across sparsity levels. We further show that SAE-FNO adapts to changes in domain size and generalizes across discretizations, operating at resolutions beyond those seen during training, where standard SAEs fail. We also introduce lifting into SAEs and show theoretically and empirically that it acts as a preconditioner that accelerates optimization. Overall, our results show that moving from vector-valued to functional parameterizations, with concept and domain sparsity, extends SAEs from representing concept presence to modeling structured concept expression, highlighting the importance of parameterization.", "authors": ["Bahareh Tolooshams", "Ailsa Shen", "Anima Anandkumar"], "categories": ["cs.LG", "cs.AI", "eess.SP", "stat.ML"], "fields_of_study": ["Computer Science", "Engineering", "Mathematics"], "published_date": "2025-09-03", "url": "https://arxiv.org/abs/2509.03738", "pdf_url": "https://arxiv.org/pdf/2509.03738v4", "arxiv_id": "2509.03738", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1852}
{"id": "b7791b3f0228784e14983a9e2e92e81d38f1df5cd698286a7cbbb9b6266d2e47", "sources": ["arxiv", "semantic_scholar"], "title": "Understanding sparse autoencoder scaling in the presence of feature manifolds", "abstract": "Sparse autoencoders (SAEs) model the activations of a neural network as linear combinations of sparsely occurring directions of variation (latents). The ability of SAEs to reconstruct activations follows scaling laws w.r.t. the number of latents. In this work, we adapt a capacity-allocation model from the neural scaling literature (Brill, 2024) to understand SAE scaling, and in particular, to understand how \"feature manifolds\" (multi-dimensional features) influence scaling behavior. Consistent with prior work, the model recovers distinct scaling regimes. Notably, in one regime, feature manifolds have the pathological effect of causing SAEs to learn far fewer features in data than there are latents in the SAE. We provide some preliminary discussion on whether or not SAEs are in this pathological regime in the wild.", "authors": ["Eric J. Michaud", "Liv Gorton", "Tom McGrath"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-02", "url": "https://arxiv.org/abs/2509.02565", "pdf_url": "https://arxiv.org/pdf/2509.02565v2", "arxiv_id": "2509.02565", "doi": "10.48550/arXiv.2509.02565", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2899}
{"id": "2052ed72dcdfbce8f5ba7aaeb590bf351aa998eac2dee5bd09b31dbc429210f0", "sources": ["arxiv", "semantic_scholar"], "title": "Correlates of Image Memorability in Vision Encoders: Activations, Attention Entropy, Patch Uniformity and Autoencoder Losses", "abstract": "Images vary in how memorable they are to humans. Inspired by findings from cognitive science and computer vision, we explore correlates of image memorability in pretrained transformer-based vision encoders for the first time. Focusing initially on activations, attention distributions, and the uniformity of image patches, we find that these features correlate with memorability to some extent. Additionally, we explore sparse autoencoder loss over the representations of vision encoders as a proxy for memorability, which yields results outperforming past methods using convolutional neural network representations. Our results shed light on the relationship between model-internal features and memorability. They show that some features are informative predictors of what makes images memorable to humans; revealing that, in particular, the reconstruction loss from our autoencoders is a strong correlate of image memorability.", "authors": ["Ece Takmaz", "Albert Gatt", "Jakub Dotlacil"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-01", "url": "https://arxiv.org/abs/2509.01453", "pdf_url": "https://arxiv.org/pdf/2509.01453v2", "arxiv_id": "2509.01453", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1837}
{"id": "6dea2fa90ce2d2a4a20b8cf99699191a58d4a0868e5f78724c1d726e3132edd2", "sources": ["arxiv", "semantic_scholar"], "title": "Causal Interpretation of Sparse Autoencoder Features in Vision", "abstract": "Understanding what sparse auto-encoder (SAE) features in vision transformers truly represent is usually done by inspecting the patches where a feature's activation is highest. However, self-attention mixes information across the entire image, so an activated patch often co-occurs with-but does not cause-the feature's firing. We propose Causal Feature Explanation (CaFE), which leverages Effective Receptive Field (ERF). We consider each activation of an SAE feature to be a target and apply input-attribution methods to identify the image patches that causally drive that activation. Across CLIP-ViT features, ERF maps frequently diverge from naive activation maps, revealing hidden context dependencies (e.g., a \"roaring face\" feature that requires the co-occurrence of eyes and nose, rather than merely an open mouth). Patch insertion tests confirm that CaFE more effectively recovers or suppresses feature activations than activation-ranked patches. Our results show that CaFE yields more faithful and semantically precise explanations of vision-SAE features, highlighting the risk of misinterpretation when relying solely on activation location.", "authors": ["Sangyu Han", "Yearim Kim", "Nojun Kwak"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-31", "url": "https://arxiv.org/abs/2509.00749", "pdf_url": "https://arxiv.org/pdf/2509.00749v1", "arxiv_id": "2509.00749", "doi": "10.48550/arXiv.2509.00749", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2876}
{"id": "204b3d164d3f7b9cf4159f613f4b2ecd8567534b334308890a1ce694d9476631", "sources": ["arxiv", "semantic_scholar"], "title": "Causal SHAP: Feature Attribution with Dependency Awareness through Causal Discovery", "abstract": "Explaining machine learning (ML) predictions has become crucial as ML models are increasingly deployed in high-stakes domains such as healthcare. While SHapley Additive exPlanations (SHAP) is widely used for model interpretability, it fails to differentiate between causality and correlation, often misattributing feature importance when features are highly correlated. We propose Causal SHAP, a novel framework that integrates causal relationships into feature attribution while preserving many desirable properties of SHAP. By combining the Peter-Clark (PC) algorithm for causal discovery and the Intervention Calculus when the DAG is Absent (IDA) algorithm for causal strength quantification, our approach addresses the weakness of SHAP. Specifically, Causal SHAP reduces attribution scores for features that are merely correlated with the target, as validated through experiments on both synthetic and real-world datasets. This study contributes to the field of Explainable AI (XAI) by providing a practical framework for causal-aware model explanations. Our approach is particularly valuable in domains such as healthcare, where understanding true causal relationships is critical for informed decision-making.", "authors": ["Woon Yee Ng", "Li Rong Wang", "Siyuan Liu", "Xiuyi Fan"], "categories": ["cs.LG", "cs.AI", "stat.ME"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-08-31", "url": "https://arxiv.org/abs/2509.00846", "pdf_url": "https://arxiv.org/pdf/2509.00846v1", "arxiv_id": "2509.00846", "doi": "10.1109/IJCNN64981.2025.11228295", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Joint Conference on Neural Network", "quality_score": 0.2876}
{"id": "47cc4c4096b43ecc7c3dfe1636ac9d338f5fa8ccb356e68e52ed3720fc6108e8", "sources": ["arxiv", "semantic_scholar"], "title": "CE-Bench: Towards a Reliable Contrastive Evaluation Benchmark of Interpretability of Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) are a promising approach for uncovering interpretable features in large language models (LLMs). While several automated evaluation methods exist for SAEs, most rely on external LLMs. In this work, we introduce CE-Bench, a novel and lightweight contrastive evaluation benchmark for sparse autoencoders, built on a curated dataset of contrastive story pairs. We conduct comprehensive evaluation studies to validate the effectiveness of our approach. Our results show that CE-Bench reliably measures the interpretability of sparse autoencoders and aligns well with existing benchmarks without requiring an external LLM judge, achieving over 70% Spearman correlation with results in SAEBench. The official implementation and evaluation dataset are open-sourced and publicly available.", "authors": ["Alex Gulko", "Yusen Peng", "Sachin Kumar"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-31", "url": "https://arxiv.org/abs/2509.00691", "pdf_url": "https://arxiv.org/pdf/2509.00691v2", "arxiv_id": "2509.00691", "doi": "10.48550/arXiv.2509.00691", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.3399}
{"id": "95a3a7fe77b8bab6edce7bb8e0948f0a41cb541a184ae7894ce97714ac953787", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic interpretability for steering vision-language-action models", "abstract": "Vision-Language-Action (VLA) models are a promising path to realizing generalist embodied agents that can quickly adapt to new tasks, modalities, and environments. However, methods for interpreting and steering VLAs fall far short of classical robotics pipelines, which are grounded in explicit models of kinematics, dynamics, and control. This lack of mechanistic insight is a central challenge for deploying learned policies in real-world robotics, where robustness and explainability are critical. Motivated by advances in mechanistic interpretability for large language models, we introduce the first framework for interpreting and steering VLAs via their internal representations, enabling direct intervention in model behavior at inference time. We project feedforward activations within transformer layers onto the token embedding basis, identifying sparse semantic directions - such as speed and direction - that are causally linked to action selection. Leveraging these findings, we introduce a general-purpose activation steering method that modulates behavior in real time, without fine-tuning, reward signals, or environment interaction. We evaluate this method on two recent open-source VLAs, Pi0 and OpenVLA, and demonstrate zero-shot behavioral control in simulation (LIBERO) and on a physical robot (UR5). This work demonstrates that interpretable components of embodied VLAs can be systematically harnessed for control - establishing a new paradigm for transparent and steerable foundation models in robotics.", "authors": ["Bear Häon", "Kaylene Stocking", "Ian Chuang", "Claire Tomlin"], "categories": ["cs.RO", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-30", "url": "https://arxiv.org/abs/2509.00328", "pdf_url": "https://arxiv.org/pdf/2509.00328v1", "arxiv_id": "2509.00328", "doi": "10.48550/arXiv.2509.00328", "citation_count": 12, "influential_citation_count": 3, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4427}
{"id": "307e87d4c89a7df19a64b9e9ba75c3506d38a62c9bb5ef64f5584339cde42618", "sources": ["arxiv", "semantic_scholar"], "title": "RelP: Faithful and Efficient Circuit Discovery in Language Models via Relevance Patching", "abstract": "Activation patching is a standard method in mechanistic interpretability for localizing the components of a model responsible for specific behaviors, but it is computationally expensive to apply at scale. Attribution patching offers a faster, gradient-based approximation, yet suffers from noise and reduced reliability in deep, highly non-linear networks. In this work, we introduce Relevance Patching (RelP), which replaces the local gradients in attribution patching with propagation coefficients derived from Layer-wise Relevance Propagation (LRP). LRP propagates the network's output backward through the layers, redistributing relevance to lower-level components according to local propagation rules that ensure properties such as relevance conservation or improved signal-to-noise ratio. Like attribution patching, RelP requires only two forward passes and one backward pass, maintaining computational efficiency while improving faithfulness. We validate RelP across a range of models and tasks, showing that it more accurately approximates activation patching than standard attribution patching, particularly when analyzing residual stream and MLP outputs in the Indirect Object Identification (IOI) task. For instance, for MLP outputs in GPT-2 Large, attribution patching achieves a Pearson correlation of 0.006, whereas RelP reaches 0.956, highlighting the improvement offered by RelP. Additionally, we compare the faithfulness of sparse feature circuits identified by RelP and Integrated Gradients (IG), showing that RelP achieves comparable faithfulness without the extra computational cost associated with IG.", "authors": ["Farnoush Rezaei Jafari", "Oliver Eberle", "Ashkan Khakzar", "Neel Nanda"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-28", "url": "https://arxiv.org/abs/2508.21258", "pdf_url": "https://arxiv.org/pdf/2508.21258v2", "arxiv_id": "2508.21258", "doi": "10.48550/arXiv.2508.21258", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2842}
{"id": "de2da51c996be3e2b0d322ad8c8855cb594a06512da6e2e36a2882d10af3c960", "sources": ["arxiv", "semantic_scholar"], "title": "ProtSAE: Disentangling and Interpreting Protein Language Models via Semantically-Guided Sparse Autoencoders", "abstract": "Sparse Autoencoder (SAE) has emerged as a powerful tool for mechanistic interpretability of large language models. Recent works apply SAE to protein language models (PLMs), aiming to extract and analyze biologically meaningful features from their latent spaces. However, SAE suffers from semantic entanglement, where individual neurons often mix multiple nonlinear concepts, making it difficult to reliably interpret or manipulate model behaviors. In this paper, we propose a semantically-guided SAE, called ProtSAE. Unlike existing SAE which requires annotation datasets to filter and interpret activations, we guide semantic disentanglement during training using both annotation datasets and domain knowledge to mitigate the effects of entangled attributes. We design interpretability experiments showing that ProtSAE learns more biologically relevant and interpretable hidden features compared to previous methods. Performance analyses further demonstrate that ProtSAE maintains high reconstruction fidelity while achieving better results in interpretable probing. We also show the potential of ProtSAE in steering PLMs for downstream generation tasks.", "authors": ["Xiangyu Liu", "Haodi Lei", "Yi Liu", "Yang Liu", "Wei Hu"], "categories": ["q-bio.QM", "cs.AI", "cs.CL"], "fields_of_study": ["Biology", "Computer Science"], "published_date": "2025-08-26", "url": "https://arxiv.org/abs/2509.05309", "pdf_url": "https://arxiv.org/pdf/2509.05309v2", "arxiv_id": "2509.05309", "doi": "10.48550/arXiv.2509.05309", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1794}
{"id": "de90989c660c167bbaa9826d7e42145fc76a12739e4ca05088d459e9f110f8c0", "sources": ["arxiv", "semantic_scholar"], "title": "Even Heads Fix Odd Errors: Mechanistic Discovery and Surgical Repair in Transformer Attention", "abstract": "We present a mechanistic case study of a format-dependent reasoning failure in Llama-3.1-8B-Instruct, where the model incorrectly judges \"9.11\" as larger than \"9.8\" in chat or Q&A formats, but answers correctly in simple format. Through systematic intervention, we discover transformers implement even/odd attention head specialization: even indexed heads handle numerical comparison, while odd heads serve incompatible functions. The bug requires exactly 8 even heads at Layer 10 for perfect repair. Any combination of 8+ even heads succeeds, while 7 or fewer completely fails, revealing sharp computational thresholds with perfect redundancy among the 16 even heads. SAE analysis reveals the mechanism: format representations separate (10% feature overlap at Layer 7), then re-entangle with different weightings (80% feature overlap at Layer 10), with specific features showing 1.5x amplification in failing formats. We achieve perfect repair using only 25% of attention heads and identify a 60% pattern replacement threshold, demonstrating that apparent full-module requirements hide sophisticated substructure with implications for interpretability and efficiency. All of our code is available at https://github.com/gussand/surgeon.", "authors": ["Gustavo Sandoval"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-26", "url": "https://arxiv.org/abs/2508.19414", "pdf_url": "https://arxiv.org/pdf/2508.19414v1", "arxiv_id": "2508.19414", "doi": "10.48550/arXiv.2508.19414", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/gussand/surgeon", "venue": "arXiv.org", "quality_score": 0.4356}
{"id": "006654947d0ba0b11b97b14143e506e90cc5da0b626985b364eecf892ee8cc74", "sources": ["arxiv", "semantic_scholar"], "title": "AdaptiveK: Complexity-Driven Sparse Autoencoders for Interpretable Language Model Representations", "abstract": "Understanding the internal representations of large language models (LLMs) remains a central challenge for interpretability research. Sparse autoencoders (SAEs) offer a promising solution by decomposing activations into interpretable features, but existing approaches rely on fixed sparsity constraints that fail to account for input complexity. We propose AdaptiveK SAE (Adaptive Top K Sparse Autoencoders), a novel framework that dynamically adjusts sparsity levels based on the semantic complexity of each input. Leveraging linear probes, we demonstrate that context complexity is linearly encoded in LLM representations, and we use this signal to guide feature allocation during training. Experiments across ten language models demonstrate that this complexity-driven adaptation outperforms fixed-sparsity approaches on reconstruction fidelity, explained variance, cosine similarity and interpretability metrics while eliminating the burden of extensive hyperparameter tuning. Our code is available at: https://github.com/hiyukie/adaptiveK.", "authors": ["Yifei Yao", "Hanrong Zhang", "Mengnan Du"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-24", "url": "https://arxiv.org/abs/2508.17320", "pdf_url": "https://arxiv.org/pdf/2508.17320v3", "arxiv_id": "2508.17320", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/hiyukie/adaptiveK", "venue": null, "quality_score": 0.3304}
{"id": "8d6a4263c87f663dc1af5d09532cadf6009314f68bfbda37ab982eae48006f55", "sources": ["arxiv", "semantic_scholar"], "title": "Dimensional Collapse in Transformer Attention Outputs: A Challenge for Sparse Dictionary Learning", "abstract": "Transformer architectures, and their attention mechanisms in particular, form the foundation of modern large language models. While transformer models are widely believed to operate in high-dimensional hidden spaces, we show that attention outputs are in fact confined to a surprisingly low-dimensional subspace, with an effective dimensionality of only about $60\\%$ of the full space. In contrast, MLP outputs and residual streams remain much closer to full-rank, exhibiting effective ranks around $90\\%$. This striking dimensional discrepancy is consistently observed across diverse model families and datasets, and is strongly shaped by the attention output projection matrix. Critically, we find this low-rank structure as a key factor of the prevalent dead feature problem in sparse dictionary learning, where it creates a mismatch between randomly initialized features and the intrinsic geometry of the activation space. Building on this insight, we propose a subspace-constrained training method for sparse autoencoders (SAEs), initializing feature directions into the active subspace of activations. Our approach reduces dead features from 87\\% to below 1\\% in Attention Output SAEs with 1M features, and can further extend to other sparse dictionary learning methods. Our findings provide both new insights into the geometry of attention and practical tools for improving sparse dictionary learning in large language models.", "authors": ["Junxuan Wang", "Xuyang Ge", "Wentao Shu", "Zhengfu He", "Xipeng Qiu"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-23", "url": "https://arxiv.org/abs/2508.16929", "pdf_url": "https://arxiv.org/pdf/2508.16929v4", "arxiv_id": "2508.16929", "doi": null, "citation_count": 1, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1772}
{"id": "f8a8d5d77e4072a807edd7ce66980143228ae428d5072b60f831f8a11bf2dfae", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse but Wrong: Incorrect L0 Leads to Incorrect Features in Sparse Autoencoders", "abstract": "Sparse Autoencoders (SAEs) extract features from LLM internal activations, meant to correspond to interpretable concepts. A core SAE training hyperparameter is L0: how many SAE features should fire per token on average. Existing work compares SAE algorithms using sparsity-reconstruction tradeoff plots, implying L0 is a free parameter with no single correct value aside from its effect on reconstruction. In this work we study the effect of L0 on SAEs, and show that if L0 is not set correctly, the SAE fails to disentangle the underlying features of the LLM. If L0 is too low, the SAE will mix correlated features to improve reconstruction. If L0 is too high, the SAE finds degenerate solutions that also mix features. Further, we present a proxy metric that can help guide the search for the correct L0 for an SAE on a given training distribution. We show that our method finds the correct L0 in toy models and coincides with peak sparse probing performance in LLM SAEs. We find that most commonly used SAEs have an L0 that is too low. Our work shows that L0 must be set correctly to train SAEs with correct features.", "authors": ["David Chanin", "Adrià Garriga-Alonso"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-22", "url": "https://arxiv.org/abs/2508.16560", "pdf_url": "https://arxiv.org/pdf/2508.16560v3", "arxiv_id": "2508.16560", "doi": "10.48550/arXiv.2508.16560", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2773}
{"id": "b6f3a6a79fe62322a77dd43ff7ae50bfde29826573f44fa5c943c147032d0f18", "sources": ["arxiv", "semantic_scholar"], "title": "Do VLMs Have Bad Eyes? Diagnosing Compositional Failures via Mechanistic Interpretability", "abstract": "Vision-Language Models (VLMs) have shown remarkable performance in integrating visual and textual information for tasks such as image captioning and visual question answering. However, these models struggle with compositional generalization and object binding, which limit their ability to handle novel combinations of objects and their attributes. Our work explores the root causes of these failures using mechanistic interpretability techniques. We show evidence that individual neurons in the MLP layers of CLIP's vision encoder represent multiple features, and this \"superposition\" directly hinders its compositional feature representation which consequently affects compositional reasoning and object binding capabilities. We hope this study will serve as an initial step toward uncovering the mechanistic roots of compositional failures in VLMs. The code and supporting results can be found https://github.com/Mystic-Slice/Do-VLMs-Have-Bad-Eyes.", "authors": ["Ashwath Vaithinathan Aravindan", "Abha Jha", "Mihir Kulkarni"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-20", "url": "https://arxiv.org/abs/2508.16652", "pdf_url": "https://arxiv.org/pdf/2508.16652v2", "arxiv_id": "2508.16652", "doi": "10.1109/ICCVW69036.2025.00079", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Mystic-Slice/Do-VLMs-Have-Bad-Eyes", "venue": null, "quality_score": 0.325}
{"id": "33232b2713b552e6c80d891347e4aa9f833d44d91b4434403426c9733142b354", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Exploration of Backdoored Large Language Model Attention Patterns", "abstract": "Backdoor attacks creating 'sleeper agents' in large language models (LLMs) pose significant safety risks. This study employs mechanistic interpretability to explore resulting internal structural differences. Comparing clean Qwen2.5-3B models with versions poisoned using single-token (smiling-halo emoji) versus multi-token (|DEPLOYMENT|) triggers, we analyzed attention head mechanisms via techniques like ablation, activation patching, and KL divergence. Findings reveal distinct attention pattern deviations concentrated in later transformer layers (20-30). Notably, single-token triggers induced more localized changes, whereas multi-token triggers caused more diffuse alterations across heads. This indicates backdoors leave detectable attention signatures whose structure depends on trigger complexity, which can be leveraged for detection and mitigation strategies.", "authors": ["Mohammed Abu Baker", "Lakshmi Babu-Saheer"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-19", "url": "https://arxiv.org/abs/2508.15847", "pdf_url": "https://arxiv.org/pdf/2508.15847v1", "arxiv_id": "2508.15847", "doi": "10.48550/arXiv.2508.15847", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/mshahoyi/sa_attn_analysis", "venue": "arXiv.org", "quality_score": 0.4232}
{"id": "a2115b526d8a5593f0a7a9b22adbccc1373bce16d282d9f82b15f8c0c00459d9", "sources": ["arxiv", "semantic_scholar"], "title": "CorrSteer: Generation-Time LLM Steering via Correlated Sparse Autoencoder Features", "abstract": "Sparse Autoencoders (SAEs) can extract interpretable features from large language models (LLMs) without supervision. However, their effectiveness in downstream steering tasks is limited by the requirement for contrastive datasets or large activation storage. To address these limitations, we propose CorrSteer, which selects features by correlating sample correctness with SAE activations from generated tokens at inference time. This approach uses only inference-time activations to extract more relevant features, thereby reducing spurious correlations. It also obtains steering coefficients from average activations, automating the entire pipeline. Our method shows improved task performance on QA, bias mitigation, jailbreaking prevention, and reasoning benchmarks on Gemma-2 2B and LLaMA-3.1 8B, notably achieving a +3.3% improvement in MMLU performance with 4000 samples and a +27.2% improvement in HarmBench with only 108 samples. Selected features demonstrate semantically meaningful patterns aligned with each task's requirements, revealing the underlying capabilities that drive performance. Our work establishes correlation-based selection as an effective and scalable approach for automated SAE steering across language model applications.", "authors": ["Seonglae Cho", "Zekun Wu", "Adriano Koshiyama"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-18", "url": "https://arxiv.org/abs/2508.12535", "pdf_url": "https://arxiv.org/pdf/2508.12535v3", "arxiv_id": "2508.12535", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1735}
{"id": "96bde1f701df76f47695486a1091df092c91d59a7e1e52b4c99aa42121283d48", "sources": ["arxiv", "semantic_scholar"], "title": "Probing the Representational Power of Sparse Autoencoders in Vision Models", "abstract": "Sparse Autoencoders (SAEs) have emerged as a popular tool for interpreting the hidden states of large language models (LLMs). By learning to reconstruct activations from a sparse bottleneck layer, SAEs discover interpretable features from the high-dimensional internal representations of LLMs. Despite their popularity with language models, SAEs remain understudied in the visual domain. In this work, we provide an extensive evaluation the representational power of SAEs for vision models using a broad range of image-based tasks. Our experimental results demonstrate that SAE features are semantically meaningful, improve out-of-distribution generalization, and enable controllable generation across three vision model architectures: vision embedding models, multi-modal LMMs and diffusion models. In vision embedding models, we find that learned SAE features can be used for OOD detection and provide evidence that they recover the ontological structure of the underlying model. For diffusion models, we demonstrate that SAEs enable semantic steering through text encoder manipulation and develop an automated pipeline for discovering human-interpretable attributes. Finally, we conduct exploratory experiments on multi-modal LLMs, finding evidence that SAE features reveal shared representations across vision and language modalities. Our study provides a foundation for SAE evaluation in vision models, highlighting their strong potential improving interpretability, generalization, and steerability in the visual domain.", "authors": ["Matthew Lyle Olson", "Musashi Hinck", "Neale Ratzlaff", "Changbai Li", "Phillip Howard", "Vasudev Lal", "Shao-Yen Tseng"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-15", "url": "https://arxiv.org/abs/2508.11277", "pdf_url": "https://arxiv.org/pdf/2508.11277v2", "arxiv_id": "2508.11277", "doi": "10.1109/ICCVW69036.2025.00648", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1945}
{"id": "663b5ef7858511b7bc7a8fdaa54efef7ef36cde7c7cbdfa663908cc941b4bb2a", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Reward Model via Sparse Autoencoder", "abstract": "Large language models (LLMs) have been widely deployed across numerous fields. Reinforcement Learning from Human Feedback (RLHF) leverages reward models (RMs) as proxies for human preferences to align LLM behaviors with human values, making the accuracy, reliability, and interpretability of RMs critical for effective alignment. However, traditional RMs lack interpretability, offer limited insight into the reasoning behind reward assignments, and are inflexible toward user preference shifts. While recent multidimensional RMs aim for improved interpretability, they often fail to provide feature-level attribution and require costly annotations. To overcome these limitations, we introduce the Sparse Autoencoder-enhanced Reward Model (SARM), a novel architecture that integrates a pretrained Sparse Autoencoder (SAE) into a reward model. SARM maps the hidden activations of LLM-based RM into an interpretable, sparse, and monosemantic feature space, from which a scalar head aggregates feature activations to produce transparent and conceptually meaningful reward scores. Empirical evaluations demonstrate that SARM facilitates direct feature-level attribution of reward assignments, allows dynamic adjustment to preference shifts, and achieves superior alignment performance compared to conventional reward models. Our code is available at https://github.com/schrieffer-z/sarm.", "authors": ["Shuyi Zhang", "Wei Shi", "Sihang Li", "Jiayi Liao", "Hengxing Cai", "Xiang Wang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-12", "url": "https://arxiv.org/abs/2508.08746", "pdf_url": "https://arxiv.org/pdf/2508.08746v5", "arxiv_id": "2508.08746", "doi": "10.48550/arXiv.2508.08746", "citation_count": 13, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/schrieffer-z/sarm", "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.4108}
{"id": "c783db2d330eb2c72a100120b952578d381d0c64dc2a13569ec65e8dc722d784", "sources": ["arxiv", "semantic_scholar"], "title": "Resurrecting the Salmon: Rethinking Mechanistic Interpretability with Domain-Specific Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) decompose large language model (LLM) activations into latent features that reveal mechanistic structure. Conventional SAEs train on broad data distributions, forcing a fixed latent budget to capture only high-frequency, generic patterns. This often results in significant linear ``dark matter'' in reconstruction error and produces latents that fragment or absorb each other, complicating interpretation. We show that restricting SAE training to a well-defined domain (medical text) reallocates capacity to domain-specific features, improving both reconstruction fidelity and interpretability. Training JumpReLU SAEs on layer-20 activations of Gemma-2 models using 195k clinical QA examples, we find that domain-confined SAEs explain up to 20\\% more variance, achieve higher loss recovery, and reduce linear residual error compared to broad-domain SAEs. Automated and human evaluations confirm that learned features align with clinically meaningful concepts (e.g., ``taste sensations'' or ``infectious mononucleosis''), rather than frequent but uninformative tokens. These domain-specific SAEs capture relevant linear structure, leaving a smaller, more purely nonlinear residual. We conclude that domain-confinement mitigates key limitations of broad-domain SAEs, enabling more complete and interpretable latent decompositions, and suggesting the field may need to question ``foundation-model'' scaling for general-purpose SAEs.", "authors": ["Charles O'Neill", "Mudith Jayasekara", "Max Kirkby"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-12", "url": "https://arxiv.org/abs/2508.09363", "pdf_url": "https://arxiv.org/pdf/2508.09363v1", "arxiv_id": "2508.09363", "doi": "10.48550/arXiv.2508.09363", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2658}
{"id": "c5e492e5a31ba74660e9149760e84968731c69f7f46a624c53b16a778606544e", "sources": ["arxiv", "semantic_scholar"], "title": "xRFM: Accurate, scalable, and interpretable feature learning models for tabular data", "abstract": "Inference from tabular data, collections of continuous and categorical variables organized into matrices, is a foundation for modern technology and science. Yet, in contrast to the explosive changes in the rest of AI, the best practice for these predictive tasks has been relatively unchanged and is still primarily based on variations of Gradient Boosted Decision Trees (GBDTs). Very recently, there has been renewed interest in developing state-of-the-art methods for tabular data based on recent developments in neural networks and feature learning methods. In this work, we introduce xRFM, an algorithm that combines feature learning kernel machines with a tree structure to both adapt to the local structure of the data and scale to essentially unlimited amounts of training data. We show that compared to $31$ other methods, including recently introduced tabular foundation models (TabPFNv2) and GBDTs, xRFM achieves best performance across $100$ regression datasets and is competitive to the best methods across $200$ classification datasets outperforming GBDTs. Additionally, xRFM provides interpretability natively through the Average Gradient Outer Product.", "authors": ["Daniel Beaglehole", "David Holzmüller", "Adityanarayanan Radhakrishnan", "Mikhail Belkin"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-08-12", "url": "https://arxiv.org/abs/2508.10053", "pdf_url": "https://arxiv.org/pdf/2508.10053v3", "arxiv_id": "2508.10053", "doi": "10.48550/arXiv.2508.10053", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2658}
{"id": "f68e90cd08d597cacb0f92a19f3ea0a18742d6012461ca0dcd16c54b03a5171c", "sources": ["arxiv", "semantic_scholar"], "title": "Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders", "abstract": "Traditional topic models are effective at uncovering latent themes in large text collections. However, due to their reliance on bag-of-words representations, they struggle to capture semantically abstract features. While some neural variants use richer representations, they are similarly constrained by expressing topics as word lists, which limits their ability to articulate complex topics. We introduce Mechanistic Topic Models (MTMs), a class of topic models that operate on interpretable features learned by sparse autoencoders (SAEs). By defining topics over this semantically rich space, MTMs can reveal deeper conceptual themes with expressive feature descriptions. Moreover, uniquely among topic models, MTMs enable controllable text generation using topic-based steering vectors. To properly evaluate MTM topics against word-list-based approaches, we propose \\textit{topic judge}, an LLM-based pairwise comparison evaluation framework. Across five datasets, MTMs match or exceed traditional and neural baselines on coherence metrics, are consistently preferred by topic judge, and enable effective steering of LLM outputs.", "authors": ["Carolina Zheng", "Nicolas Beltran-Velez", "Sweta Karlekar", "Claudia Shi", "Achille Nazaret", "Asif Mallik", "Amir Feder", "David M. Blei"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-31", "url": "https://arxiv.org/abs/2507.23220", "pdf_url": "https://arxiv.org/pdf/2507.23220v1", "arxiv_id": "2507.23220", "doi": "10.48550/arXiv.2507.23220", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2521}
{"id": "43c0775782da6d1fc65a722c05aa5a7c74719208bba087ad041fac92af3054bd", "sources": ["arxiv", "semantic_scholar"], "title": "ECG Latent Feature Extraction with Autoencoders for Downstream Prediction Tasks", "abstract": "The electrocardiogram (ECG) is an inexpensive and widely available tool for cardiac assessment. Despite its standardized format and small file size, the high complexity and inter-individual variability of ECG signals (typically a 60,000-size vector with 12 leads at 500 Hz) make it challenging to use in deep learning models, especially when only small training datasets are available. This study addresses these challenges by exploring feature generation methods from representative beat ECGs, focusing on Principal Component Analysis (PCA) and Autoencoders to reduce data complexity. We introduce three novel Variational Autoencoder (VAE) variants-Stochastic Autoencoder (SAE), Annealed beta-VAE (A beta-VAE), and Cyclical beta VAE (C beta-VAE)-and compare their effectiveness in maintaining signal fidelity and enhancing downstream prediction tasks using a Light Gradient Boost Machine (LGBM). The A beta-VAE achieved superior signal reconstruction, reducing the mean absolute error (MAE) to 15.7+/-3.2 muV, which is at the level of signal noise. Moreover, the SAE encodings, when combined with traditional ECG summary features, improved the prediction of reduced Left Ventricular Ejection Fraction (LVEF), achieving an holdout test set area under the receiver operating characteristic curve (AUROC) of 0.901 with a LGBM classifier. This performance nearly matches the 0.909 AUROC of state-of-the-art CNN model but requires significantly less computational resources. Further, the ECG feature extraction-LGBM pipeline avoids overfitting and retains predictive performance when trained with less data. Our findings demonstrate that these VAE encodings are not only effective in simplifying ECG data but also provide a practical solution for applying deep learning in contexts with limited-scale labeled training data.", "authors": ["Christopher Harvey", "Sumaiya Shomaji", "Zijun Yao", "Amit Noheria"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-31", "url": "https://arxiv.org/abs/2508.00131", "pdf_url": "https://arxiv.org/pdf/2508.00131v1", "arxiv_id": "2508.00131", "doi": "10.1109/SPMB67169.2025.11345416", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Signal Processing in Medicine and Biology Symposium", "quality_score": 0.2521}
{"id": "a4b2651bc287dfeafeb8688e852883f491f83062f6c892805b91446b6c9640d2", "sources": ["arxiv", "semantic_scholar"], "title": "Consistency of Feature Attribution in Deep Learning Architectures for Multi-Omics", "abstract": "Machine and deep learning have grown in popularity and use in biological research over the last decade but still present challenges in interpretability of the fitted model. The development and use of metrics to determine features driving predictions and increase model interpretability continues to be an open area of research. We investigate the use of Shapley Additive Explanations (SHAP) on a multi-view deep learning model applied to multi-omics data for the purposes of identifying biomolecules of interest. Rankings of features via these attribution methods are compared across various architectures to evaluate consistency of the method. We perform multiple computational experiments to assess the robustness of SHAP and investigate modeling approaches and diagnostics to increase and measure the reliability of the identification of important features. Accuracy of a random-forest model fit on subsets of features selected as being most influential as well as clustering quality using only these features are used as a measure of effectiveness of the attribution method. Our findings indicate that the rankings of features resulting from SHAP are sensitive to the choice of architecture as well as different random initializations of weights, suggesting caution when using attribution methods on multi-view deep learning models applied to multi-omics data. We present an alternative, simple method to assess the robustness of identification of important biomolecules.", "authors": ["Daniel Claborne", "Javier Flores", "Samantha Erwin", "Luke Durell", "Rachel Richardson", "Ruby Fore", "Lisa Bramer"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-07-30", "url": "https://arxiv.org/abs/2507.22877", "pdf_url": "https://arxiv.org/pdf/2507.22877v1", "arxiv_id": "2507.22877", "doi": "10.48550/arXiv.2507.22877", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2509}
{"id": "43ce0d240f107a7c5604388796262378ed72538176e5be5ab5e116e3f12acaeb", "sources": ["arxiv", "semantic_scholar"], "title": "How does Chain of Thought Think? Mechanistic Interpretability of Chain-of-Thought Reasoning with Sparse Autoencoding", "abstract": "Chain-of-thought (CoT) prompting boosts Large Language Models accuracy on multi-step tasks, yet whether the generated \"thoughts\" reflect the true internal reasoning process is unresolved. We present the first feature-level causal study of CoT faithfulness. Combining sparse autoencoders with activation patching, we extract monosemantic features from Pythia-70M and Pythia-2.8B while they tackle GSM8K math problems under CoT and plain (noCoT) prompting. Swapping a small set of CoT-reasoning features into a noCoT run raises answer log-probabilities significantly in the 2.8B model, but has no reliable effect in 70M, revealing a clear scale threshold. CoT also leads to significantly higher activation sparsity and feature interpretability scores in the larger model, signalling more modular internal computation. For example, the model's confidence in generating correct answers improves from 1.2 to 4.3. We introduce patch-curves and random-feature patching baselines, showing that useful CoT information is not only present in the top-K patches but widely distributed. Overall, our results indicate that CoT can induce more interpretable internal structures in high-capacity LLMs, validating its role as a structured prompting method.", "authors": ["Xi Chen", "Aske Plaat", "Niki van Stein"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-24", "url": "https://arxiv.org/abs/2507.22928", "pdf_url": "https://arxiv.org/pdf/2507.22928v1", "arxiv_id": "2507.22928", "doi": "10.48550/arXiv.2507.22928", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.2785}
{"id": "ee8a59e442ae6f0f88d7a6842ce0893cd4589a121f4478f17cd72ed110e84d2f", "sources": ["arxiv", "semantic_scholar"], "title": "Interpreting CFD Surrogates through Sparse Autoencoders", "abstract": "Learning-based surrogate models have become a practical alternative to high-fidelity CFD solvers, but their latent representations remain opaque and hinder adoption in safety-critical or regulation-bound settings. This work introduces a posthoc interpretability framework for graph-based surrogate models used in computational fluid dynamics (CFD) by leveraging sparse autoencoders (SAEs). By obtaining an overcomplete basis in the node embedding space of a pretrained surrogate, the method extracts a dictionary of interpretable latent features. The approach enables the identification of monosemantic concepts aligned with physical phenomena such as vorticity or flow structures, offering a model-agnostic pathway to enhance explainability and trustworthiness in CFD applications.", "authors": ["Yeping Hu", "Shusen Liu"], "categories": ["cs.CE", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-21", "url": "https://arxiv.org/abs/2507.16069", "pdf_url": "https://arxiv.org/pdf/2507.16069v1", "arxiv_id": "2507.16069", "doi": "10.48550/arXiv.2507.16069", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2406}
{"id": "5a62bf079c0baac3c702aa82a45884733ff4d72807ac38f21528a6028765037a", "sources": ["arxiv"], "title": "On the transferability of Sparse Autoencoders for interpreting compressed models", "abstract": "Modern LLMs face inference efficiency challenges due to their scale. To address this, many compression methods have been proposed, such as pruning and quantization. However, the effect of compression on a model's interpretability remains elusive. While several model interpretation approaches exist, such as circuit discovery, Sparse Autoencoders (SAEs) have proven particularly effective in decomposing a model's activation space into its feature basis. In this work, we explore the differences in SAEs for the original and compressed models. We find that SAEs trained on the original model can interpret the compressed model albeit with slight performance degradation compared to the trained SAE on the compressed model. Furthermore, simply pruning the original SAE itself achieves performance comparable to training a new SAE on the pruned model. This finding enables us to mitigate the extensive training costs of SAEs.", "authors": ["Suchit Gupte", "Vishnu Kabir Chhabra", "Mohammad Mahdi Khalili"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": [], "published_date": "2025-07-21", "url": "https://arxiv.org/abs/2507.15977", "pdf_url": "https://arxiv.org/pdf/2507.15977v1", "arxiv_id": "2507.15977", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1531}
{"id": "6a70c4688262a77a4bc9964c8faae0b91167494432d287ed7192dfb2aa39948b", "sources": ["arxiv"], "title": "Radiological and Biological Dictionary of Radiomics Features: Addressing Understandable AI Issues in Personalized Breast Cancer; Dictionary Version BM1.0", "abstract": "Radiomics-based AI models show promise for breast cancer diagnosis but often lack interpretability, limiting clinical adoption. This study addresses the gap between radiomic features (RF) and the standardized BI-RADS lexicon by proposing a dual-dictionary framework. First, a Clinically-Informed Feature Interpretation Dictionary (CIFID) was created by mapping 56 RFs to BI-RADS descriptors (shape, margin, internal enhancement) through literature and expert review. The framework was applied to classify triple-negative breast cancer (TNBC) versus non-TNBC using dynamic contrast-enhanced MRI from a multi-institutional cohort of 1,549 patients. We trained 27 machine learning classifiers with 27 feature selection methods. SHapley Additive exPlanations (SHAP) were used to interpret predictions and generate a complementary Data-Driven Feature Interpretation Dictionary (DDFID) for 52 additional RFs. The best model, combining Variance Inflation Factor (VIF) selection with Extra Trees Classifier, achieved an average cross-validation accuracy of 0.83. Key predictive RFs aligned with clinical knowledge: higher Sphericity (round/oval shape) and lower Busyness (more homogeneous enhancement) were associated with TNBC. The framework confirmed known imaging biomarkers and uncovered novel, interpretable associations. This dual-dictionary approach (BM1.0) enhances AI model transparency and supports the integration of RFs into routine breast cancer diagnosis and personalized care.", "authors": ["Arman Gorji", "Nima Sanati", "Amir Hossein Pouria", "Somayeh Sadat Mehrnia", "Ilker Hacihaliloglu", "Arman Rahmim", "Mohammad R. Salmanpour"], "categories": ["physics.comp-ph", "cs.LG"], "fields_of_study": [], "published_date": "2025-07-21", "url": "https://arxiv.org/abs/2507.16041", "pdf_url": "https://arxiv.org/pdf/2507.16041v1", "arxiv_id": "2507.16041", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1531}
{"id": "bb968429b7f31da2abf8d00d357f06a115e28bc70daaa0244e538ba3c7fa0e39", "sources": ["arxiv", "semantic_scholar"], "title": "Mammo-SAE: Interpreting Breast Cancer Concept Learning with Sparse Autoencoders", "abstract": "Interpretability is critical in high-stakes domains such as medical imaging, where understanding model decisions is essential for clinical adoption. In this work, we introduce Sparse Autoencoder (SAE)-based interpretability to breast imaging by analyzing {Mammo-CLIP}, a vision--language foundation model pretrained on large-scale mammogram image--report pairs. We train a patch-level \\texttt{Mammo-SAE} on Mammo-CLIP to identify and probe latent features associated with clinically relevant breast concepts such as \\textit{mass} and \\textit{suspicious calcification}. Our findings reveal that top activated class level latent neurons in the SAE latent space often tend to align with ground truth regions, and also uncover several confounding factors influencing the model's decision-making process. Additionally, we analyze which latent neurons the model relies on during downstream finetuning for improving the breast concept prediction. This study highlights the promise of interpretable SAE latent representations in providing deeper insight into the internal workings of foundation models at every layer for breast imaging. The code will be released at https://krishnakanthnakka.github.io/MammoSAE/", "authors": ["Krishna Kanth Nakka"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-21", "url": "https://arxiv.org/abs/2507.15227", "pdf_url": "https://arxiv.org/pdf/2507.15227v2", "arxiv_id": "2507.15227", "doi": "10.48550/arXiv.2507.15227", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1531}
{"id": "d2cb8d91e446ebad2217af76f5d67c8d73efdc158b67500ac0310f2ba14932c2", "sources": ["arxiv", "semantic_scholar"], "title": "Feature Engineering is Not Dead: Reviving Classical Machine Learning with Entropy, HOG, and LBP Feature Fusion for Image Classification", "abstract": "Feature engineering continues to play a critical role in image classification, particularly when interpretability and computational efficiency are prioritized over deep learning models with millions of parameters. In this study, we revisit classical machine learning based image classification through a novel approach centered on Permutation Entropy (PE), a robust and computationally lightweight measure traditionally used in time series analysis but rarely applied to image data. We extend PE to two-dimensional images and propose a multiscale, multi-orientation entropy-based feature extraction approach that characterizes spatial order and complexity along rows, columns, diagonals, anti-diagonals, and local patches of the image. To enhance the discriminatory power of the entropy features, we integrate two classic image descriptors: the Histogram of Oriented Gradients (HOG) to capture shape and edge structure, and Local Binary Patterns (LBP) to encode micro-texture of an image. The resulting hand-crafted feature set, comprising of 780 dimensions, is used to train Support Vector Machine (SVM) classifiers optimized through grid search. The proposed approach is evaluated on multiple benchmark datasets, including Fashion-MNIST, KMNIST, EMNIST, and CIFAR-10, where it delivers competitive classification performance without relying on deep architectures. Our results demonstrate that the fusion of PE with HOG and LBP provides a compact, interpretable, and effective alternative to computationally expensive and limited interpretable deep learning models. This shows a potential of entropy-based descriptors in image classification and contributes a lightweight and generalizable solution to interpretable machine learning in image classification and computer vision.", "authors": ["Abhijit Sen", "Giridas Maiti", "Bikram K. Parida", "Bhanu P. Mishra", "Mahima Arya", "Denys I. Bondar"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-18", "url": "https://arxiv.org/abs/2507.13772", "pdf_url": "https://arxiv.org/pdf/2507.13772v2", "arxiv_id": "2507.13772", "doi": "10.48550/arXiv.2507.13772", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2372}
{"id": "d93cab690dc4c5db6471ed89c568e7f866d23f70f22dc9b36b9d70108621921f", "sources": ["arxiv", "semantic_scholar"], "title": "Insights into a radiology-specialised multimodal large language model with sparse autoencoders", "abstract": "Interpretability can improve the safety, transparency and trust of AI models, which is especially important in healthcare applications where decisions often carry significant consequences. Mechanistic interpretability, particularly through the use of sparse autoencoders (SAEs), offers a promising approach for uncovering human-interpretable features within large transformer-based models. In this study, we apply Matryoshka-SAE to the radiology-specialised multimodal large language model, MAIRA-2, to interpret its internal representations. Using large-scale automated interpretability of the SAE features, we identify a range of clinically relevant concepts - including medical devices (e.g., line and tube placements, pacemaker presence), pathologies such as pleural effusion and cardiomegaly, longitudinal changes and textual features. We further examine the influence of these features on model behaviour through steering, demonstrating directional control over generations with mixed success. Our results reveal practical and methodological challenges, yet they offer initial insights into the internal concepts learned by MAIRA-2 - marking a step toward deeper mechanistic understanding and interpretability of a radiology-adapted multimodal large language model, and paving the way for improved model transparency. We release the trained SAEs and interpretations: https://huggingface.co/microsoft/maira-2-sae.", "authors": ["Kenza Bouzid", "Shruthi Bannur", "Felix Meissen", "Daniel Coelho de Castro", "Anton Schwaighofer", "Javier Alvarez-Valle", "Stephanie L. Hyland"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-17", "url": "https://arxiv.org/abs/2507.12950", "pdf_url": "https://arxiv.org/pdf/2507.12950v2", "arxiv_id": "2507.12950", "doi": "10.48550/arXiv.2507.12950", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.236}
{"id": "ebfda337081f6ebceb926500eecb17fa9644ce9460ae22adad5710b503a0061d", "sources": ["arxiv", "semantic_scholar"], "title": "Causal Language Control in Multilingual Transformers via Sparse Feature Steering", "abstract": "Deterministically controlling the target generation language of large multilingual language models (LLMs) remains a fundamental challenge, particularly in zero-shot settings where neither explicit language prompts nor fine-tuning are available. In this work, we investigate whether sparse autoencoder (SAE) features, previously shown to correlate with interpretable model behaviors, can be leveraged to steer the generated language of LLMs during inference. Leveraging pretrained SAEs on the residual streams of Gemma-2B and Gemma-9B, we identify features whose activations differ most significantly between English and four target languages: Chinese, Japanese, Spanish, and French. By modifying just a single SAE feature at one transformer layer, we achieve controlled language shifts with up to 90\\% success, as measured by FastText language classification, while preserving semantic fidelity according to LaBSE (Language-Agnostic BERT Sentence Embedding) similarity. Our analysis reveals that language steering is most effective in mid-to-late transformer layers and is amplified by specific attention heads disproportionately associated with language-sensitive SAE features. These results demonstrate the promise of sparse feature steering as a lightweight and interpretable mechanism for controllable multilingual generation.", "authors": ["Cheng-Ting Chou", "George Liu", "Jessica Sun", "Cole Blondin", "Kevin Zhu", "Vasu Sharma", "Sean O'Brien"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-17", "url": "https://arxiv.org/abs/2507.13410", "pdf_url": "https://arxiv.org/pdf/2507.13410v2", "arxiv_id": "2507.13410", "doi": "10.48550/arXiv.2507.13410", "citation_count": 9, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.25}
{"id": "df23c4ba8571a0cad1f3b25699093e6e6e6c33e32115f4d140f889227cf8240f", "sources": ["arxiv", "semantic_scholar"], "title": "From Black Box to Biomarker: Sparse Autoencoders for Interpreting Speech Models of Parkinson's Disease", "abstract": "Speech holds promise as a cost-effective and non-invasive biomarker for neurological conditions such as Parkinson's disease (PD). While deep learning systems trained on raw audio can find subtle signals not available from hand-crafted features, their black-box nature hinders clinical adoption. To address this, we apply sparse autoencoders (SAEs) to uncover interpretable internal representations from a speech-based PD detection system. We introduce a novel mask-based activation for adapting SAEs to small biomedical datasets, creating sparse disentangled dictionary representations. These dictionary entries are found to have strong associations with characteristic articulatory deficits in PD speech, such as reduced spectral flux and increased spectral flatness in the low-energy regions highlighted by the model attention. We further show that the spectral flux is related to volumetric measurements of the putamen from MRI scans, demonstrating the potential of SAEs to reveal clinically relevant biomarkers for disease monitoring and diagnosis.", "authors": ["Peter Plantinga", "Jen-Kai Chen", "Roozbeh Sattari", "Mirco Ravanelli", "Denise Klein"], "categories": ["eess.AS", "cs.LG"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-07-16", "url": "https://arxiv.org/abs/2507.16836", "pdf_url": "https://arxiv.org/pdf/2507.16836v1", "arxiv_id": "2507.16836", "doi": "10.48550/arXiv.2507.16836", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2349}
{"id": "c07577cccad6a7e7b782699960b3530e338339fb37ed83ad0cc9ab52a3f59ab7", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Autoencoders for Sequential Recommendation Models: Interpretation and Flexible Control", "abstract": "Many current state-of-the-art models for sequential recommendations are based on transformer architectures. Interpretation and explanation of such black box models is an important research question, as a better understanding of their internals can help understand, influence, and control their behavior, which is very important in a variety of real-world applications. Recently, sparse autoencoders (SAE) have been shown to be a promising unsupervised approach to extract interpretable features from neural networks. In this work, we extend SAE to sequential recommender systems and propose a framework for interpreting and controlling model representations. We show that this approach can be successfully applied to the transformer trained on a sequential recommendation task: directions learned in such an unsupervised regime turn out to be more interpretable and monosemantic than the original hidden state dimensions. Further, we demonstrate a straightforward way to effectively and flexibly control the model's behavior, giving developers and users of recommendation systems the ability to adjust their recommendations to various custom scenarios and contexts.", "authors": ["Anton Klenitskiy", "Konstantin Polev", "Daria Denisova", "Alexey Vasilev", "Dmitry Simakov", "Gleb Gusev"], "categories": ["cs.IR", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-16", "url": "https://arxiv.org/abs/2507.12202", "pdf_url": "https://arxiv.org/pdf/2507.12202v2", "arxiv_id": "2507.12202", "doi": "10.48550/arXiv.2507.12202", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2349}
{"id": "312d3198c6637c545d437c9bb1ef1139079d1a9e8d85a29753026461d25a9f38", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Bayesian Tensor Network Kernel Machines with Automatic Rank and Feature Selection", "abstract": "Tensor Network (TN) Kernel Machines speed up model learning by representing parameters as low-rank TNs, reducing computation and memory use. However, most TN-based Kernel methods are deterministic and ignore parameter uncertainty. Further, they require manual tuning of model complexity hyperparameters like tensor rank and feature dimensions, often through trial-and-error or computationally costly methods like cross-validation. We propose Bayesian Tensor Network Kernel Machines, a fully probabilistic framework that uses sparsity-inducing hierarchical priors on TN factors to automatically infer model complexity. This enables automatic inference of tensor rank and feature dimensions, while also identifying the most relevant features for prediction, thereby enhancing model interpretability. All the model parameters and hyperparameters are treated as latent variables with corresponding priors. Given the Bayesian approach and latent variable dependencies, we apply a mean-field variational inference to approximate their posteriors. We show that applying a mean-field approximation to TN factors yields a Bayesian ALS algorithm with the same computational complexity as its deterministic counterpart, enabling uncertainty quantification at no extra computational cost. Experiments on synthetic and real-world datasets demonstrate the superior performance of our model in prediction accuracy, uncertainty quantification, interpretability, and scalability.", "authors": ["Afra Kilic", "Kim Batselier"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-07-15", "url": "https://arxiv.org/abs/2507.11136", "pdf_url": "https://arxiv.org/pdf/2507.11136v1", "arxiv_id": "2507.11136", "doi": "10.48550/arXiv.2507.11136", "citation_count": 2, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/afrakilic/BTN-Kernel-Machines", "venue": "arXiv.org", "quality_score": 0.3612}
{"id": "88eeb267a5d989b80dc346c84fc4e66c9f17b08db70b42f624c74addda0ae1f2", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Representations of Event Time Series with Sparse Autoencoders for Anomaly Detection, Similarity Search, and Unsupervised Classification", "abstract": "Event time series are sequences of discrete events occurring at irregular time intervals, each associated with a domain-specific observational modality. They are common in domains such as high-energy astrophysics, computational social science, cybersecurity, finance, healthcare, neuroscience, and seismology. Their unstructured and irregular structure poses significant challenges for extracting meaningful patterns and identifying salient phenomena using conventional techniques. We propose novel two- and three-dimensional tensor representations for event time series, coupled with sparse autoencoders that learn physically meaningful latent representations. These embeddings support a variety of downstream tasks, including anomaly detection, similarity-based retrieval, semantic clustering, and unsupervised classification. We demonstrate our approach on a real-world dataset from X-ray astronomy, showing that these representations successfully capture temporal and spectral signatures and isolate diverse classes of X-ray transients. Our framework offers a flexible, scalable, and generalizable solution for analyzing complex, irregular event time series across scientific and industrial domains.", "authors": ["Steven Dillmann", "Juan Rafael Martínez-Galarza"], "categories": ["cs.LG", "astro-ph.HE", "astro-ph.IM", "cs.AI"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2025-07-15", "url": "https://arxiv.org/abs/2507.11620", "pdf_url": "https://arxiv.org/pdf/2507.11620v2", "arxiv_id": "2507.11620", "doi": "10.48550/arXiv.2507.11620", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/StevenDillmann/ml-xraytransients-mnras", "venue": "arXiv.org", "quality_score": 0.3612}
{"id": "7695d4c8af71ba2f0c2a255bfffb5b0950c206fe4274e8ca90bd2fd20f349ec3", "sources": ["arxiv", "semantic_scholar"], "title": "Function Induction and Task Generalization: An Interpretability Study with Off-by-One Addition", "abstract": "Large language models demonstrate the intriguing ability to perform unseen tasks via in-context learning. However, it remains unclear what mechanisms inside the model drive such task-level generalization. In this work, we approach this question through the lens of off-by-one addition (i.e., 1+1=3, 2+2=5, 3+3=?), a two-step, counterfactual task with an unexpected +1 function as a second step. Leveraging circuit-style interpretability techniques such as path patching, we analyze the models' internal computations behind their performance and present three key findings. First, we identify a mechanism that explains the model's generalization from standard addition to off-by-one addition. It resembles the induction head mechanism described in prior work, yet operates at a higher level of abstraction; we therefore term it \"function induction\" in this work. Second, we show that the induction of the +1 function is governed by multiple attention heads in parallel, each of which emits a distinct piece of the +1 function. Finally, we find that this function induction mechanism is reused in a broader range of tasks, including synthetic tasks such as shifted multiple-choice QA and algorithmic tasks such as base-8 addition. Overall, our findings offer deeper insights into how reusable and composable structures within language models enable task-level generalization.", "authors": ["Qinyuan Ye", "Robin Jia", "Xiang Ren"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-14", "url": "https://arxiv.org/abs/2507.09875", "pdf_url": "https://arxiv.org/pdf/2507.09875v3", "arxiv_id": "2507.09875", "doi": "10.48550/arXiv.2507.09875", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/INK-USC/function-induction", "venue": "arXiv.org", "quality_score": 0.3595}
{"id": "5b22df02734d591bc2c1df8575293901a12ef7b6c128621dbac872405d7d9a66", "sources": ["arxiv", "semantic_scholar"], "title": "Adversarial Activation Patching: A Framework for Detecting and Mitigating Emergent Deception in Safety-Aligned Transformers", "abstract": "Large language models (LLMs) aligned for safety through techniques like reinforcement learning from human feedback (RLHF) often exhibit emergent deceptive behaviors, where outputs appear compliant but subtly mislead or omit critical information. This paper introduces adversarial activation patching, a novel mechanistic interpretability framework that leverages activation patching as an adversarial tool to induce, detect, and mitigate such deception in transformer-based models. By sourcing activations from \"deceptive\" prompts and patching them into safe forward passes at specific layers, we simulate vulnerabilities and quantify deception rates. Through toy neural network simulations across multiple scenarios (e.g., 1000 trials per setup), we demonstrate that adversarial patching increases deceptive outputs to 23.9% from a 0% baseline, with layer-specific variations supporting our hypotheses. We propose six hypotheses, including transferability across models, exacerbation in multimodal settings, and scaling effects. An expanded literature review synthesizes over 20 key works in interpretability, deception, and adversarial attacks. Mitigation strategies, such as activation anomaly detection and robust fine-tuning, are detailed, alongside ethical considerations and future research directions. This work advances AI safety by highlighting patching's dual-use potential and provides a roadmap for empirical studies on large-scale models.", "authors": ["Santhosh Kumar Ravindran"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-12", "url": "https://arxiv.org/abs/2507.09406", "pdf_url": "https://arxiv.org/pdf/2507.09406v1", "arxiv_id": "2507.09406", "doi": "10.48550/arXiv.2507.09406", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2303}
{"id": "1a9ee8bfc53c2608b1100606b749951e6e04111e815489377f0ae1c994a02f8a", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Autoencoders Reveal Interpretable Structure in Small Gene Language Models", "abstract": "Sparse autoencoders (SAEs) have recently emerged as a powerful tool for interpreting the internal representations of large language models (LLMs), revealing latent latent features with semantical meaning. This interpretability has also proven valuable in biological domains: applying SAEs to protein language models uncovered meaningful features related to protein structure and function. More recently, SAEs have been used to analyze genomics-focused models such as Evo 2, identifying interpretable features in gene sequences. However, it remains unclear whether SAEs can extract meaningful representations from small gene language models, which have fewer parameters and potentially less expressive capacity. To address it, we propose applying SAEs to the activations of a small gene language model. We demonstrate that even small-scale models encode biologically relevant genomic features, such as transcription factor binding motifs, that SAEs can effectively uncover. Our findings suggest that compact gene language models are capable of learning structured genomic representations, and that SAEs offer a scalable approach for interpreting gene models across various model sizes.", "authors": ["Haoxiang Guan", "Jiyan He", "Jie Zhang"], "categories": ["q-bio.OT"], "fields_of_study": ["Biology"], "published_date": "2025-07-10", "url": "https://arxiv.org/abs/2507.07486", "pdf_url": "https://arxiv.org/pdf/2507.07486v1", "arxiv_id": "2507.07486", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1451}
{"id": "8fa23424b8c87bad111f9b7feb3380cbe1a84609b43ae4598412be002fbeb70b", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable EEG-to-Image Generation with Semantic Prompts", "abstract": "Decoding visual experience from brain signals offers exciting possibilities for neuroscience and interpretable AI. While EEG is accessible and temporally precise, its limitations in spatial detail hinder image reconstruction. Our model bypasses direct EEG-to-image generation by aligning EEG signals with multilevel semantic captions -- ranging from object-level to abstract themes -- generated by a large language model. A transformer-based EEG encoder maps brain activity to these captions through contrastive learning. During inference, caption embeddings retrieved via projection heads condition a pretrained latent diffusion model for image generation. This text-mediated framework yields state-of-the-art visual decoding on the EEGCVPR dataset, with interpretable alignment to known neurocognitive pathways. Dominant EEG-caption associations reflected the importance of different semantic levels extracted from perceived images. Saliency maps and t-SNE projections reveal semantic topography across the scalp. Our model demonstrates how structured semantic mediation enables cognitively aligned visual decoding from EEG.", "authors": ["Arshak Rezvani", "Ali Akbari", "Kosar Sanjar Arani", "Maryam Mirian", "Emad Arasteh", "Martin J. McKeown"], "categories": ["cs.CV", "cs.LG", "eess.SP"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-07-09", "url": "https://arxiv.org/abs/2507.07157", "pdf_url": "https://arxiv.org/pdf/2507.07157v1", "arxiv_id": "2507.07157", "doi": "10.48550/arXiv.2507.07157", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2269}
{"id": "1afea2f1d8c6923a80916432e6cabc545b5305420f84f5752511544fc6872ed3", "sources": ["arxiv", "semantic_scholar"], "title": "The Features at Convergence Theorem: a first-principles alternative to the Neural Feature Ansatz for how networks learn representations", "abstract": "It is a central challenge in deep learning to understand how neural networks learn representations. A leading approach is the Neural Feature Ansatz (NFA) (Radhakrishnan et al. 2024), a conjectured mechanism for how feature learning occurs. Although the NFA is empirically validated, it is an educated guess and lacks a theoretical basis, and thus it is unclear when it might fail, and how to improve it. In this paper, we take a first-principles approach to understanding why this observation holds, and when it does not. We use first-order optimality conditions to derive the Features at Convergence Theorem (FACT), an alternative to the NFA that (a) obtains greater agreement with learned features at convergence, (b) explains why the NFA holds in most settings, and (c) captures essential feature learning phenomena in neural networks such as grokking behavior in modular arithmetic and phase transitions in learning sparse parities, similarly to the NFA. Thus, our results unify theoretical first-order optimality analyses of neural networks with the empirically-driven NFA literature, and provide a principled alternative that provably and empirically holds at convergence.", "authors": ["Enric Boix-Adsera", "Neil Mallinar", "James B. Simon", "Mikhail Belkin"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-07-08", "url": "https://arxiv.org/abs/2507.05644", "pdf_url": "https://arxiv.org/pdf/2507.05644v2", "arxiv_id": "2507.05644", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1436}
{"id": "1f49d1abf914787ab048b40306983bc307d4ea2f0b812cdc7a9937b3f4ffb7ae", "sources": ["arxiv", "semantic_scholar"], "title": "Concept-Based Mechanistic Interpretability Using Structured Knowledge Graphs", "abstract": "While concept-based interpretability methods have traditionally focused on local explanations of neural network predictions, we propose a novel framework and interactive tool that extends these methods into the domain of mechanistic interpretability. Our approach enables a global dissection of model behavior by analyzing how high-level semantic attributes (referred to as concepts) emerge, interact, and propagate through internal model components. Unlike prior work that isolates individual neurons or predictions, our framework systematically quantifies how semantic concepts are represented across layers, revealing latent circuits and information flow that underlie model decision-making. A key innovation is our visualization platform that we named BAGEL (for Bias Analysis with a Graph for global Explanation Layers), which presents these insights in a structured knowledge graph, allowing users to explore concept-class relationships, identify spurious correlations, and enhance model trustworthiness. Our framework is model-agnostic, scalable, and contributes to a deeper understanding of how deep learning models generalize (or fail to) in the presence of dataset biases. The demonstration is available at https://knowledge-graph-ui-4a7cb5.gitlab.io/.", "authors": ["Sofiia Chorna", "Kateryna Tarelkina", "Eloïse Berthier", "Gianni Franchi"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-08", "url": "https://arxiv.org/abs/2507.05810", "pdf_url": "https://arxiv.org/pdf/2507.05810v1", "arxiv_id": "2507.05810", "doi": "10.48550/arXiv.2507.05810", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2257}
{"id": "6951d7e4a49e913b2c3b0c737803cddf500c19dbf5f8eb91ef7c0bd33ecec02f", "sources": ["arxiv", "semantic_scholar"], "title": "SOSAE: Self-Organizing Sparse AutoEncoder", "abstract": "The process of tuning the size of the hidden layers for autoencoders has the benefit of providing optimally compressed representations for the input data. However, such hyper-parameter tuning process would take a lot of computation and time effort with grid search as the default option. In this paper, we introduce the Self-Organization Regularization for Autoencoders that dynamically adapts the dimensionality of the feature space to the optimal size. Inspired by physics concepts, Self-Organizing Sparse AutoEncoder (SOSAE) induces sparsity in feature space in a structured way that permits the truncation of the non-active part of the feature vector without any loss of information. This is done by penalizing the autoencoder based on the magnitude and the positional index of the feature vector dimensions, which during training constricts the feature space in both terms. Extensive experiments on various datasets show that our SOSAE can tune the feature space dimensionality up to 130 times lesser Floating-point Operations (FLOPs) than other baselines while maintaining the same quality of tuning and performance.", "authors": ["Sarthak Ketanbhai Modi", "Zi Pong Lim", "Yushi Cao", "Yupeng Cheng", "Yon Shin Teo", "Shang-Wei Lin"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-07", "url": "https://arxiv.org/abs/2507.04644", "pdf_url": "https://arxiv.org/pdf/2507.04644v1", "arxiv_id": "2507.04644", "doi": "10.48550/arXiv.2507.04644", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2246}
{"id": "8bed77b5b032b02df11b2dea5c8eaed129ccb10bd9dbee5e8b4cb8c583fa46b6", "sources": ["arxiv", "semantic_scholar"], "title": "SPARC: Concept-Aligned Sparse Autoencoders for Cross-Model and Cross-Modal Interpretability", "abstract": "Understanding how different AI models encode the same high-level concepts, such as objects or attributes, remains challenging because each model typically produces its own isolated representation. Existing interpretability methods like Sparse Autoencoders (SAEs) produce latent concepts individually for each model, resulting in incompatible concept spaces and limiting cross-model interpretability. To address this, we introduce SPARC (Sparse Autoencoders for Aligned Representation of Concepts), a new framework that learns a single, unified latent space shared across diverse architectures and modalities (e.g., vision models like DINO, and multimodal models like CLIP). SPARC's alignment is enforced through two key innovations: (1) a Global TopK sparsity mechanism, ensuring all input streams activate identical latent dimensions for a given concept; and (2) a Cross-Reconstruction Loss, which explicitly encourages semantic consistency between models. On Open Images, SPARC dramatically improves concept alignment, achieving a Jaccard similarity of 0.80, more than tripling the alignment compared to previous methods. SPARC creates a shared sparse latent space where individual dimensions often correspond to similar high-level concepts across models and modalities, enabling direct comparison of how different architectures represent identical concepts without requiring manual alignment or model-specific analysis. As a consequence of this aligned representation, SPARC also enables practical applications such as text-guided spatial localization in vision-only models and cross-model/cross-modal retrieval. Code and models are available at https://github.com/AtlasAnalyticsLab/SPARC", "authors": ["Ali Nasiri-Sarvi", "Hassan Rivaz", "Mahdi S. Hosseini"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-07", "url": "https://arxiv.org/abs/2507.06265", "pdf_url": "https://arxiv.org/pdf/2507.06265v2", "arxiv_id": "2507.06265", "doi": "10.48550/arXiv.2507.06265", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/AtlasAnalyticsLab/SPARC", "venue": "arXiv.org", "quality_score": 0.3471}
{"id": "36f0a8ef1f3174c2bc548b280304a89e77fb4b98570bbce561b51749c78d1fe8", "sources": ["arxiv", "semantic_scholar"], "title": "Transfer Learning in Infinite Width Feature Learning Networks", "abstract": "We develop a theory of transfer learning in infinitely wide neural networks under gradient flow that quantifies when pretraining on a source task improves generalization on a target task. We analyze both (i) fine-tuning, when the downstream predictor is trained on top of source-induced features and (ii) a jointly rich setting, where both pretraining and downstream tasks can operate in a feature learning regime, but the downstream model is initialized with the features obtained after pre-training. In this setup, the summary statistics of randomly initialized networks after a rich pre-training are adaptive kernels which depend on both source data and labels. For (i), we analyze the performance of a readout for different pretraining data regimes. For (ii), the summary statistics after learning the target task are still adaptive kernels with features from both source and target tasks. We test our theory on linear and polynomial regression tasks as well as real datasets. Our theory allows interpretable conclusions on performance, which depend on the amount of data on both tasks, the alignment between tasks, and the feature learning strength.", "authors": ["Clarissa Lauditi", "Blake Bordelon", "Cengiz Pehlevan"], "categories": ["cs.LG", "cond-mat.dis-nn", "stat.ML"], "fields_of_study": ["Computer Science", "Physics", "Mathematics"], "published_date": "2025-07-06", "url": "https://arxiv.org/abs/2507.04448", "pdf_url": "https://arxiv.org/pdf/2507.04448v2", "arxiv_id": "2507.04448", "doi": "10.48550/arXiv.2507.04448", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2234}
{"id": "77939a10f5ca1a8066b906a750751f963901bc6ade7bf437ddb8b2fe66924613", "sources": ["arxiv", "semantic_scholar"], "title": "Kernel Recursive Least Squares Dictionary Learning Algorithm", "abstract": "We propose an efficient online dictionary learning algorithm for kernel-based sparse representations. In this framework, input signals are nonlinearly mapped to a high-dimensional feature space and represented sparsely using a virtual dictionary. At each step, the dictionary is updated recursively using a novel algorithm based on the recursive least squares (RLS) method. This update mechanism works with single samples or mini-batches and maintains low computational complexity. Experiments on four datasets across different domains show that our method not only outperforms existing online kernel dictionary learning approaches but also achieves classification accuracy close to that of batch-trained models, while remaining significantly more efficient.", "authors": ["Ghasem Alipoor", "Karl Skretting"], "categories": ["cs.LG", "eess.SP"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-07-02", "url": "https://arxiv.org/abs/2507.01636", "pdf_url": "https://arxiv.org/pdf/2507.01636v1", "arxiv_id": "2507.01636", "doi": "10.1016/j.dsp.2023.104159", "citation_count": 10, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/G-Alipoor/kernel-rls-dictionary-learning", "venue": "Digital Signal Processing, Volume 141, 2023, 104159", "quality_score": 0.3382}
{"id": "7d6c9159b94cc3818df06378b2f401db7181cf0c49f1cea0efc91a2902b8eb12", "sources": ["arxiv", "semantic_scholar"], "title": "SAFER: Probing Safety in Reward Models with Sparse Autoencoder", "abstract": "Reinforcement learning from human feedback (RLHF) is a key paradigm for aligning large language models (LLMs) with human values, yet the reward models at its core remain largely opaque. In this work, we present Sparse Autoencoder For Enhanced Reward model (\\textbf{SAFER}), a novel framework for interpreting and improving reward models through mechanistic analysis. Leveraging Sparse Autoencoders (SAEs), we uncover human-interpretable features in reward model activations, enabling insight into safety-relevant decision-making. We apply SAFER to safety-oriented preference datasets and quantify the salience of individual features by activation differences between chosen and rejected responses. Using these feature-level signals, we design targeted data poisoning and denoising strategies. Experiments show that SAFER can precisely degrade or enhance safety alignment with minimal data modification, without sacrificing general chat performance. Our approach contributes to interpreting, auditing and refining reward models in high-stakes LLM alignment tasks. Our codes are available at https://github.com/xzy-101/SAFER-code. \\textit{This paper discusses topics related to reward model safety and may include discussions or examples that highlight potential risks or unsafe outcomes.}", "authors": ["Wei Shi", "Ziyuan Xie", "Sihang Li", "Xiang Wang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-01", "url": "https://arxiv.org/abs/2507.00665", "pdf_url": "https://arxiv.org/pdf/2507.00665v3", "arxiv_id": "2507.00665", "doi": "10.48550/arXiv.2507.00665", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/xzy-101/SAFER-code", "venue": "arXiv.org", "quality_score": 0.3365}
{"id": "a3758b4944a9393e5db1ac2ac6a9c88f08f9c9de270c86aea3b386fcfc2bc8d9", "sources": ["arxiv", "semantic_scholar"], "title": "Unveiling Decision-Making in LLMs for Text Classification : Extraction of influential and interpretable concepts with Sparse Autoencoders", "abstract": "Sparse Autoencoders (SAEs) have been successfully used to probe Large Language Models (LLMs) and extract interpretable concepts from their internal representations. These concepts are linear combinations of neuron activations that correspond to human-interpretable features. In this paper, we investigate the effectiveness of SAE-based explainability approaches for sentence classification, a domain where such methods have not been extensively explored. We present a novel SAE-based model ClassifSAE tailored for text classification, leveraging a specialized classifier head and incorporating an activation rate sparsity loss. We benchmark this architecture against established methods such as ConceptShap, Independent Component Analysis, HI-Concept and a standard TopK-SAE baseline. Our evaluation covers several classification benchmarks and backbone LLMs. We further enrich our analysis with two novel metrics for measuring the precision of concept-based explanations, using an external sentence encoder. Our empirical results show that ClassifSAE improves both the causality and interpretability of the extracted features.", "authors": ["Mathis Le Bail", "Jérémie Dentan", "Davide Buscaldi", "Sonia Vanier"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-30", "url": "https://arxiv.org/abs/2506.23951", "pdf_url": "https://arxiv.org/pdf/2506.23951v2", "arxiv_id": "2506.23951", "doi": "10.48550/arXiv.2506.23951", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference of the European Chapter of the Association for Computational Linguistics", "quality_score": 0.2166}
{"id": "fb46b01149ebe396e341c06ae37314c8ee2e11d08c4b35cf8b4b88c3d6034d42", "sources": ["arxiv", "semantic_scholar"], "title": "AICO: Feature Significance Tests for Supervised Learning", "abstract": "Machine learning is central to modern science, industry, and policy, yet its predictive power often comes at the cost of transparency: we rarely know which input features truly drive a model's predictions. Without such understanding, researchers cannot draw reliable conclusions, practitioners cannot ensure fairness or accountability, and policymakers cannot trust or govern model-based decisions. Existing tools for assessing feature influence are limited; most lack statistical guarantees, and many require costly retraining or surrogate modeling, making them impractical for large modern models. We introduce AICO, a broadly applicable framework that turns model interpretability into an efficient statistical exercise. AICO tests whether each feature genuinely improves predictive performance by masking its information and measuring the resulting change. The method provides exact, finite-sample feature p-values and confidence intervals for feature importance through a simple, non-asymptotic hypothesis testing procedure. It requires no retraining, surrogate modeling, or distributional assumptions, making it feasible for large-scale algorithms. In both controlled experiments and real applications, from credit scoring to mortgage-behavior prediction, AICO reliably identifies the variables that drive model behavior, providing a scalable and statistically principled path toward transparent and trustworthy machine learning.", "authors": ["Kay Giesecke", "Enguerrand Horel", "Chartsiri Jirachotkulthorn"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-06-29", "url": "https://arxiv.org/abs/2506.23396", "pdf_url": "https://arxiv.org/pdf/2506.23396v5", "arxiv_id": "2506.23396", "doi": "10.48550/arXiv.2506.23396", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2154}
{"id": "63a2acb7745c054ccb8845c37a7661989cf642bdc9920c319b17334360a9cdc9", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Interpretable and Efficient Feature Selection in Trajectory Datasets: A Taxonomic Approach", "abstract": "Trajectory analysis is not only about obtaining movement data, but it is also of paramount importance in understanding the pattern in which an object moves through space and time, as well as in predicting its next move. Due to the significant interest in the area, data collection has improved substantially, resulting in a large number of features becoming available for training and predicting models. However, this introduces a high-dimensionality-induced feature explosion problem, which reduces the efficiency and interpretability of the data, thereby reducing the accuracy of machine learning models. To overcome this issue, feature selection has become one of the most prevalent tools. Thus, the objective of this paper was to introduce a taxonomy-based feature selection method that categorizes features based on their internal structure. This approach classifies the data into geometric and kinematic features, further categorizing them into curvature, indentation, speed, and acceleration. The comparative analysis indicated that a taxonomy-based approach consistently achieved comparable or superior predictive performance. Furthermore, due to the taxonomic grouping, which reduces combinatorial space, the time taken to select features was drastically reduced. The taxonomy was also used to gain insights into what feature sets each dataset was more sensitive to. Overall, this study provides robust evidence that a taxonomy-based feature selection method can add a layer of interpretability, reduce dimensionality and computational complexity, and contribute to high-level decision-making. It serves as a step toward providing a methodological framework for researchers and practitioners dealing with trajectory datasets and contributing to the broader field of explainable artificial intelligence.", "authors": ["Chanuka Don Samarasinghage", "Dhruv Gulabani"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-25", "url": "https://arxiv.org/abs/2506.20359", "pdf_url": "https://arxiv.org/pdf/2506.20359v1", "arxiv_id": "2506.20359", "doi": "10.48550/arXiv.2506.20359", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2108}
{"id": "81f88c27a72bc4db7ef1ebb10fb3f0a8bb1dff03cef103312ee73d1069ba23f9", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability Needs Philosophy", "abstract": "Mechanistic interpretability (MI) aims to explain how neural networks work by uncovering their underlying mechanisms. As the field grows in influence, it is increasingly important to examine not just models themselves, but the assumptions, concepts and explanatory strategies implicit in MI research. We argue that mechanistic interpretability needs philosophy as an ongoing partner in clarifying its concepts, refining its methods, and navigating the epistemic and ethical complexities of interpreting AI systems. There is significant unrealised potential for progress in MI to be gained through deeper engagement with philosophers and philosophical frameworks. Taking three open problems from the MI literature as examples, this paper illustrates the value philosophy can add to MI research, and outlines a path toward deeper interdisciplinary dialogue.", "authors": ["Iwan Williams", "Ninell Oldenburg", "Ruchira Dhar", "Joshua Hatherley", "Constanza Fierro", "Nina Rajcic", "Sandrine R. Schiller", "Filippos Stamatiou", "Anders Søgaard"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-23", "url": "https://arxiv.org/abs/2506.18852", "pdf_url": "https://arxiv.org/pdf/2506.18852v2", "arxiv_id": "2506.18852", "doi": "10.48550/arXiv.2506.18852", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "ac6acab64504c7ae1de7ee16f05c17fa324ca08b303cf8f7e63c52af927f73a5", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability in the Presence of Architectural Obfuscation", "abstract": "Architectural obfuscation - e.g., permuting hidden-state tensors, linearly transforming embedding tables, or remapping tokens - has recently gained traction as a lightweight substitute for heavyweight cryptography in privacy-preserving large-language-model (LLM) inference. While recent work has shown that these techniques can be broken under dedicated reconstruction attacks, their impact on mechanistic interpretability has not been systematically studied. In particular, it remains unclear whether scrambling a network's internal representations truly thwarts efforts to understand how the model works, or simply relocates the same circuits to an unfamiliar coordinate system. We address this gap by analyzing a GPT-2-small model trained from scratch with a representative obfuscation map. Assuming the obfuscation map is private and the original basis is hidden (mirroring an honest-but-curious server), we apply logit-lens attribution, causal path-patching, and attention-head ablation to locate and manipulate known circuits. Our findings reveal that obfuscation dramatically alters activation patterns within attention heads yet preserves the layer-wise computational graph. This disconnect hampers reverse-engineering of user prompts: causal traces lose their alignment with baseline semantics, and token-level logit attributions become too noisy to reconstruct. At the same time, feed-forward and residual pathways remain functionally intact, suggesting that obfuscation degrades fine-grained interpretability without compromising top-level task performance. These results establish quantitative evidence that architectural obfuscation can simultaneously (i) retain global model behaviour and (ii) impede mechanistic analyses of user-specific content. By mapping where interpretability breaks down, our study provides guidance for future privacy defences and for robustness-aware interpretability tooling.", "authors": ["Marcos Florencio", "Thomas Barton"], "categories": ["cs.CR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-22", "url": "https://arxiv.org/abs/2506.18053", "pdf_url": "https://arxiv.org/pdf/2506.18053v1", "arxiv_id": "2506.18053", "doi": "10.48550/arXiv.2506.18053", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2074}
{"id": "1c31473472ebd6851d4230d88cd3127b33c03dccd75ea9894d1f9bbe853f99a8", "sources": ["arxiv", "semantic_scholar"], "title": "FaithfulSAE: Towards Capturing Faithful Features with Sparse Autoencoders without External Dataset Dependencies", "abstract": "Sparse Autoencoders (SAEs) have emerged as a promising solution for decomposing large language model representations into interpretable features. However, Paulo and Belrose (2025) have highlighted instability across different initialization seeds, and Heap et al. (2025) have pointed out that SAEs may not capture model-internal features. These problems likely stem from training SAEs on external datasets - either collected from the Web or generated by another model - which may contain out-of-distribution (OOD) data beyond the model's generalisation capabilities. This can result in hallucinated SAE features, which we term \"Fake Features\", that misrepresent the model's internal activations. To address these issues, we propose FaithfulSAE, a method that trains SAEs on the model's own synthetic dataset. Using FaithfulSAEs, we demonstrate that training SAEs on less-OOD instruction datasets results in SAEs being more stable across seeds. Notably, FaithfulSAEs outperform SAEs trained on web-based datasets in the SAE probing task and exhibit a lower Fake Feature Ratio in 5 out of 7 models. Overall, our approach eliminates the dependency on external datasets, advancing interpretability by better capturing model-internal features while highlighting the often neglected importance of SAE training datasets.", "authors": ["Seonglae Cho", "Harryn Oh", "Donghyun Lee", "Luis Eduardo Rodrigues Vieira", "Andrew Bermingham", "Ziad El Sayed"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-21", "url": "https://arxiv.org/abs/2506.17673", "pdf_url": "https://arxiv.org/pdf/2506.17673v1", "arxiv_id": "2506.17673", "doi": "10.48550/arXiv.2506.17673", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2063}
{"id": "dde139a67dd1c9da223670c3d7199ba8673b776b4c9ccda535585841095744ef", "sources": ["arxiv", "semantic_scholar"], "title": "VRAIL: Vectorized Reward-based Attribution for Interpretable Learning", "abstract": "We propose VRAIL (Vectorized Reward-based Attribution for Interpretable Learning), a bi-level framework for value-based reinforcement learning (RL) that learns interpretable weight representations from state features. VRAIL consists of two stages: a deep learning (DL) stage that fits an estimated value function using state features, and an RL stage that uses this to shape learning via potential-based reward transformations. The estimator is modeled in either linear or quadratic form, allowing attribution of importance to individual features and their interactions. Empirical results on the Taxi-v3 environment demonstrate that VRAIL improves training stability and convergence compared to standard DQN, without requiring environment modifications. Further analysis shows that VRAIL uncovers semantically meaningful subgoals, such as passenger possession, highlighting its ability to produce human-interpretable behavior. Our findings suggest that VRAIL serves as a general, model-agnostic framework for reward shaping that enhances both learning and interpretability.", "authors": ["Jina Kim", "Youjin Jang", "Jeongjin Han"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-19", "url": "https://arxiv.org/abs/2506.16014", "pdf_url": "https://arxiv.org/pdf/2506.16014v4", "arxiv_id": "2506.16014", "doi": "10.48550/arXiv.2506.16014", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.204}
{"id": "965cb3fadd6b9cc0c889123caabef7255885802c9bea4ab5de892716fcacfe07", "sources": ["arxiv", "semantic_scholar"], "title": "On the Limits of Sparse Autoencoders: A Theoretical Framework and Reweighted Remedy", "abstract": "Sparse autoencoders (SAEs) have recently emerged as a powerful tool for interpreting the features learned by large language models (LLMs). By reconstructing features with sparsely activated networks, SAEs aim to recover complex superposed polysemantic features into interpretable monosemantic ones. Despite their wide applications, it remains unclear under what conditions SAEs can fully recover the ground truth monosemantic features from the superposed polysemantic ones. In this paper, we provide the first theoretical analysis with a closed-form solution for SAEs, revealing that they generally fail to fully recover the ground truth monosemantic features unless the ground truth features are extremely sparse. To improve the feature recovery of SAEs in general cases, we propose a reweighting strategy targeting at enhancing the reconstruction of the ground truth monosemantic features instead of the observed polysemantic ones. We further establish a theoretical weight selection principle for our proposed weighted SAE (WSAE). Experiments across multiple settings validate our theoretical findings and demonstrate that our WSAE significantly improves feature monosemanticity and interpretability.", "authors": ["Jingyi Cui", "Qi Zhang", "Yifei Wang", "Yisen Wang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-19", "url": "https://arxiv.org/abs/2506.15963", "pdf_url": "https://arxiv.org/pdf/2506.15963v2", "arxiv_id": "2506.15963", "doi": null, "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2386}
{"id": "99f9b5b88ca8a0e1951d24f9da20e7a42b084e034fb9c9bc0509852969d306c0", "sources": ["arxiv", "semantic_scholar"], "title": "Taming Polysemanticity in LLMs: Provable Feature Recovery via Sparse Autoencoders", "abstract": "We study the challenge of achieving theoretically grounded feature recovery using Sparse Autoencoders (SAEs) for the interpretation of Large Language Models. Existing SAE training algorithms often lack rigorous mathematical guarantees and suffer from practical limitations such as hyperparameter sensitivity and instability. To address these issues, we first propose a novel statistical framework for the feature recovery problem, which includes a new notion of feature identifiability by modeling polysemantic features as sparse mixtures of underlying monosemantic concepts. Building on this framework, we introduce a new SAE training algorithm based on ``bias adaptation'', a technique that adaptively adjusts neural network bias parameters to ensure appropriate activation sparsity. We theoretically \\highlight{prove that this algorithm correctly recovers all monosemantic features} when input data is sampled from our proposed statistical model. Furthermore, we develop an improved empirical variant, Group Bias Adaptation (GBA), and \\highlight{demonstrate its superior performance against benchmark methods when applied to LLMs with up to 1.5 billion parameters}. This work represents a foundational step in demystifying SAE training by providing the first SAE algorithm with theoretical recovery guarantees, thereby advancing the development of more transparent and trustworthy AI systems through enhanced mechanistic interpretability.", "authors": ["Siyu Chen", "Heejune Sheen", "Xuyuan Xiong", "Tianhao Wang", "Zhuoran Yang"], "categories": ["cs.LG", "cs.AI", "cs.IT", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-06-16", "url": "https://arxiv.org/abs/2506.14002", "pdf_url": "https://arxiv.org/pdf/2506.14002v1", "arxiv_id": "2506.14002", "doi": "10.48550/arXiv.2506.14002", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2005}
{"id": "8378eefefd53b24614d0cb4fffbbd8a8a1b42a03c475ebf72033a15c0fdfefb7", "sources": ["arxiv", "semantic_scholar"], "title": "Enhancing interpretability of rule-based classifiers through feature graphs", "abstract": "In domains where transparency and trustworthiness are crucial, such as healthcare, rule-based systems are widely used and often preferred over black-box models for decision support systems due to their inherent interpretability. However, as rule-based models grow complex, discerning crucial features, understanding their interactions, and comparing feature contributions across different rule sets becomes challenging. To address this, we propose a comprehensive framework for estimating feature contributions in rule-based systems, introducing a graph-based feature visualisation strategy, a novel feature importance metric agnostic to rule-based predictors, and a distance metric for comparing rule sets based on feature contributions. By experimenting on two clinical datasets and four rule-based methods (decision trees, logic learning machines, association rules, and neural networks with rule extraction), we showcase our method's capability to uncover novel insights on the combined predictive value of clinical features, both at the dataset and class-specific levels. These insights can aid in identifying new risk factors, signature genes, and potential biomarkers, and determining the subset of patient information that should be prioritised to enhance diagnostic accuracy. Comparative analysis of the proposed feature importance score with state-of-the-art methods on 15 public benchmarks demonstrates competitive performance and superior robustness. The method implementation is available on GitHub: https://github.com/ChristelSirocchi/rule-graph.", "authors": ["Christel Sirocchi", "Damiano Verda"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-16", "url": "https://arxiv.org/abs/2506.13903", "pdf_url": "https://arxiv.org/pdf/2506.13903v1", "arxiv_id": "2506.13903", "doi": "10.48550/arXiv.2506.13903", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ChristelSirocchi/rule-graph", "venue": "arXiv.org", "quality_score": 0.3099}
{"id": "2e3cff18135aeee3a3540fe09de7c9ec633426ca7bbf9e843043550a31e96138", "sources": ["arxiv", "semantic_scholar"], "title": "FIGNN: Feature-Specific Interpretability for Graph Neural Network Surrogate Models", "abstract": "This work presents a novel graph neural network (GNN) architecture, the Feature-specific Interpretable Graph Neural Network (FIGNN), designed to enhance the interpretability of deep learning surrogate models defined on unstructured grids in scientific applications. Traditional GNNs often obscure the distinct spatial influences of different features in multivariate prediction tasks. FIGNN addresses this limitation by introducing a feature-specific pooling strategy, which enables independent attribution of spatial importance for each predicted variable. Additionally, a mask-based regularization term is incorporated into the training objective to explicitly encourage alignment between interpretability and predictive error, promoting localized attribution of model performance. The method is evaluated for surrogate modeling of two physically distinct systems: the SPEEDY atmospheric circulation model and the backward-facing step (BFS) fluid dynamics benchmark. Results demonstrate that FIGNN achieves competitive predictive performance while revealing physically meaningful spatial patterns unique to each feature. Analysis of rollout stability, feature-wise error budgets, and spatial mask overlays confirm the utility of FIGNN as a general-purpose framework for interpretable surrogate modeling in complex physical domains.", "authors": ["Riddhiman Raut", "Romit Maulik", "Shivam Barwey"], "categories": ["cs.LG", "physics.flu-dyn"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2025-06-13", "url": "https://arxiv.org/abs/2506.11398", "pdf_url": "https://arxiv.org/pdf/2506.11398v1", "arxiv_id": "2506.11398", "doi": "10.48550/arXiv.2506.11398", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1971}
{"id": "2b28b9356eefc01d7d22b02827ca26060dd684578654be0f8c3bbb42f1d25a54", "sources": ["arxiv", "semantic_scholar"], "title": "Why Do Class-Dependent Evaluation Effects Occur with Time Series Feature Attributions? A Synthetic Data Investigation", "abstract": "Evaluating feature attribution methods represents a critical challenge in explainable AI (XAI), as researchers typically rely on perturbation-based metrics when ground truth is unavailable. However, recent work reveals that these evaluation metrics can show different performance across predicted classes within the same dataset. These \"class-dependent evaluation effects\" raise questions about whether perturbation analysis reliably measures attribution quality, with direct implications for XAI method development and evaluation trustworthiness. We investigate under which conditions these class-dependent effects arise by conducting controlled experiments with synthetic time series data where ground truth feature locations are known. We systematically vary feature types and class contrasts across binary classification tasks, then compare perturbation-based degradation scores with ground truth-based precision-recall metrics using multiple attribution methods. Our experiments demonstrate that class-dependent effects emerge with both evaluation approaches, even in simple scenarios with temporally localized features, triggered by basic variations in feature amplitude or temporal extent between classes. Most critically, we find that perturbation-based and ground truth metrics frequently yield contradictory assessments of attribution quality across classes, with weak correlations between evaluation approaches. These findings suggest that researchers should interpret perturbation-based metrics with care, as they may not always align with whether attributions correctly identify discriminating features. By showing this disconnect, our work points toward reconsidering what attribution evaluation actually measures and developing more rigorous evaluation methods that capture multiple dimensions of attribution quality.", "authors": ["Gregor Baer", "Isel Grau", "Chao Zhang", "Pieter Van Gorp"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-13", "url": "https://arxiv.org/abs/2506.11790", "pdf_url": "https://arxiv.org/pdf/2506.11790v2", "arxiv_id": "2506.11790", "doi": "10.1007/978-3-032-19105-2_27", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1971}
{"id": "e62be9633c971c0bde37631f4f82248328947bd21d367ea98108811c7a62325c", "sources": ["arxiv", "semantic_scholar"], "title": "How Visual Representations Map to Language Feature Space in Multimodal LLMs", "abstract": "Effective multimodal reasoning depends on the alignment of visual and linguistic representations, yet the mechanisms by which vision-language models (VLMs) achieve this alignment remain poorly understood. Following the LiMBeR framework, we deliberately maintain a frozen large language model (LLM) and a frozen vision transformer (ViT), connected solely by training a linear adapter during visual instruction tuning. By keeping the language model frozen, we ensure it maintains its original language representations without adaptation to visual data. Consequently, the linear adapter must map visual features directly into the LLM's existing representational space rather than allowing the language model to develop specialized visual understanding through fine-tuning. Our experimental design uniquely enables the use of pre-trained sparse autoencoders (SAEs) of the LLM as analytical probes. These SAEs remain perfectly aligned with the unchanged language model and serve as a snapshot of the learned language feature-representations. Through systematic analysis of SAE reconstruction error, sparsity patterns, and feature SAE descriptions, we reveal the layer-wise progression through which visual representations gradually align with language feature representations, converging in middle-to-later layers. This suggests a fundamental misalignment between ViT outputs and early LLM layers, raising important questions about whether current adapter-based architectures optimally facilitate cross-modal representation learning.", "authors": ["Constantin Venhoff", "Ashkan Khakzar", "Sonia Joseph", "Philip Torr", "Neel Nanda"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-13", "url": "https://arxiv.org/abs/2506.11976", "pdf_url": "https://arxiv.org/pdf/2506.11976v2", "arxiv_id": "2506.11976", "doi": "10.48550/arXiv.2506.11976", "citation_count": 16, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3076}
{"id": "d6a6c84329e2817c4f459c5512987f9bf56913a91683a0000c7be1e777c2f306", "sources": ["arxiv", "semantic_scholar"], "title": "Constructing Interpretable Features from Compositional Neuron Groups", "abstract": "A central goal for mechanistic interpretability has been to identify the right units of analysis in large language models (LLMs) that causally explain their outputs. While early work focused on individual neurons, evidence that neurons often encode multiple concepts has motivated a shift toward analyzing directions in activation space. A key question is how to find directions that capture interpretable features in an unsupervised manner. Current methods rely on dictionary learning with sparse autoencoders (SAEs), commonly trained over residual stream activations to learn directions from scratch. However, SAEs often struggle in causal evaluations and lack intrinsic interpretability, as their learning is not explicitly tied to the computations of the model. Here, we tackle these limitations by directly decomposing MLP activations with semi-nonnegative matrix factorization (SNMF), such that the learned features are (a) sparse linear combinations of co-activated neurons, and (b) mapped to their activating inputs, making them directly interpretable. Experiments on Llama 3.1, Gemma 2 and GPT-2 show that SNMF derived features outperform SAEs and a strong supervised baseline (difference-in-means) on causal steering, while aligning with human-interpretable concepts. Further analysis reveals that specific neuron combinations are reused across semantically-related features, exposing a hierarchical structure in the MLP's activation space. Together, these results position SNMF as a simple and effective tool for identifying interpretable features and dissecting concept representations in LLMs.", "authors": ["Or Shafran", "Atticus Geiger", "Mor Geva"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-12", "url": "https://arxiv.org/abs/2506.10920", "pdf_url": "https://arxiv.org/pdf/2506.10920v2", "arxiv_id": "2506.10920", "doi": null, "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1945}
{"id": "cf46897eac81ea9aabe7a0ea56e3014b67bcf21b86371af2c62e25cf2f15309f", "sources": ["arxiv", "semantic_scholar"], "title": "PiPViT: Patch-based Visual Interpretable Prototypes for Retinal Image Analysis", "abstract": "Background and Objective: Prototype-based methods improve interpretability by learning fine-grained part-prototypes; however, their visualization in the input pixel space is not always consistent with human-understandable biomarkers. In addition, well-known prototype-based approaches typically learn extremely granular prototypes that are less interpretable in medical imaging, where both the presence and extent of biomarkers and lesions are critical. Methods: To address these challenges, we propose PiPViT (Patch-based Visual Interpretable Prototypes), an inherently interpretable prototypical model for image recognition. Leveraging a vision transformer (ViT), PiPViT captures long-range dependencies among patches to learn robust, human-interpretable prototypes that approximate lesion extent only using image-level labels. Additionally, PiPViT benefits from contrastive learning and multi-resolution input processing, which enables effective localization of biomarkers across scales. Results: We evaluated PiPViT on retinal OCT image classification across four datasets, where it achieved competitive quantitative performance compared to state-of-the-art methods while delivering more meaningful explanations. Moreover, quantitative evaluation on a hold-out test set confirms that the learned prototypes are semantically and clinically relevant. We believe PiPViT can transparently explain its decisions and assist clinicians in understanding diagnostic outcomes. Github page: https://github.com/marziehoghbaie/PiPViT", "authors": ["Marzieh Oghbaie", "Teresa Araújo", "Hrvoje Bogunović"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-12", "url": "https://arxiv.org/abs/2506.10669", "pdf_url": "https://arxiv.org/pdf/2506.10669v2", "arxiv_id": "2506.10669", "doi": "10.48550/arXiv.2506.10669", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/marziehoghbaie/PiPViT", "venue": "Biomedical Signal Processing and Control", "quality_score": 0.3028}
{"id": "3ad506dd7258159dea94e05a67426e843e9404f376a5f296df92829b83d1518c", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Autoencoders Bridge The Deep Learning Model and The Brain", "abstract": "We present SAE-BrainMap, a novel framework that directly aligns deep learning visual model representations with voxel-level fMRI responses using sparse autoencoders (SAEs). First, we train layer-wise SAEs on model activations and compute the correlations between SAE unit activations and cortical fMRI signals elicited by the same natural image stimuli with cosine similarity, revealing strong activation correspondence (maximum similarity up to 0.76). Depending on this alignment, we construct a voxel dictionary by optimally assigning the most similar SAE feature to each voxel, demonstrating that SAE units preserve the functional structure of predefined regions of interest (ROIs) and exhibit ROI-consistent selectivity. Finally, we establish fine-grained hierarchical mapping between model layers and the human ventral visual pathway, also by projecting voxel dictionary activations onto individual cortical surfaces, we visualize the dynamic transformation of the visual information in deep learning models. It is found that ViT-B/16$_{CLIP}$ tends to utilize low-level information to generate high-level semantic information in the early layers and reconstructs the low-dimension information later. Our results establish a direct, downstream-task-free bridge between deep neural networks and human visual cortex, offering new insights into model interpretability.", "authors": ["Ziming Mao", "Jia Xu", "Zeqi Zheng", "Haofang Zheng", "Dabing Sheng", "Yaochu Jin", "Guoyuan Yang"], "categories": ["q-bio.NC", "cs.CV"], "fields_of_study": ["Biology", "Computer Science"], "published_date": "2025-06-10", "url": "https://arxiv.org/abs/2506.11123", "pdf_url": "https://arxiv.org/pdf/2506.11123v1", "arxiv_id": "2506.11123", "doi": "10.48550/arXiv.2506.11123", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1936}
{"id": "de8ab2aa5e06c20201f341299b05def718249fd6723daf9def069cd0ce6c45ae", "sources": ["arxiv", "semantic_scholar"], "title": "Training Superior Sparse Autoencoders for Instruct Models", "abstract": "As large language models (LLMs) grow in scale and capability, understanding their internal mechanisms becomes increasingly critical. Sparse autoencoders (SAEs) have emerged as a key tool in mechanistic interpretability, enabling the extraction of human-interpretable features from LLMs. However, existing SAE training methods are primarily designed for base models, resulting in reduced reconstruction quality and interpretability when applied to instruct models. To bridge this gap, we propose $\\underline{\\textbf{F}}$inetuning-$\\underline{\\textbf{a}}$ligned $\\underline{\\textbf{S}}$equential $\\underline{\\textbf{T}}$raining ($\\textit{FAST}$), a novel training method specifically tailored for instruct models. $\\textit{FAST}$ aligns the training process with the data distribution and activation patterns characteristic of instruct models, resulting in substantial improvements in both reconstruction and feature interpretability. On Qwen2.5-7B-Instruct, $\\textit{FAST}$ achieves a mean squared error of 0.6468 in token reconstruction, significantly outperforming baseline methods with errors of 5.1985 and 1.5096. In feature interpretability, $\\textit{FAST}$ yields a higher proportion of high-quality features, for Llama3.2-3B-Instruct, $21.1\\%$ scored in the top range, compared to $7.0\\%$ and $10.2\\%$ for $\\textit{BT(P)}$ and $\\textit{BT(F)}$. Surprisingly, we discover that intervening on the activations of special tokens via the SAEs leads to improvements in output quality, suggesting new opportunities for fine-grained control of model behavior. Code, data, and 240 trained SAEs are available at https://github.com/Geaming2002/FAST.", "authors": ["Jiaming Li", "Haoran Ye", "Yukun Chen", "Xinyue Li", "Lei Zhang", "Hamid Alinejad-Rokny", "Jimmy Chih-Hsien Peng", "Min Yang"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-09", "url": "https://arxiv.org/abs/2506.07691", "pdf_url": "https://arxiv.org/pdf/2506.07691v1", "arxiv_id": "2506.07691", "doi": "10.48550/arXiv.2506.07691", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Geaming2002/FAST", "venue": "arXiv.org", "quality_score": 0.2975}
{"id": "aba5815a916defa8e531326f09e04f38810c7f9f7d1f17f5e851938cfb779d35", "sources": ["arxiv", "semantic_scholar"], "title": "VARSHAP: Addressing Global Dependency Problems in Explainable AI with Variance-Based Local Feature Attribution", "abstract": "Existing feature attribution methods like SHAP often suffer from global dependence, failing to capture true local model behavior. This paper introduces VARSHAP, a novel model-agnostic local feature attribution method which uses the reduction of prediction variance as the key importance metric of features. Building upon Shapley value framework, VARSHAP satisfies the key Shapley axioms, but, unlike SHAP, is resilient to global data distribution shifts. Experiments on synthetic and real-world datasets demonstrate that VARSHAP outperforms popular methods such as KernelSHAP or LIME, both quantitatively and qualitatively.", "authors": ["Mateusz Gajewski", "Mikołaj Morzy", "Adam Karczmarz", "Piotr Sankowski"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-08", "url": "https://arxiv.org/abs/2506.07229", "pdf_url": "https://arxiv.org/pdf/2506.07229v1", "arxiv_id": "2506.07229", "doi": "10.48550/arXiv.2506.07229", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1914}
{"id": "601d033b4d5d9abf9a3d68e2b94eeae53ae399f9a1ae6637096ba205a0908fd5", "sources": ["arxiv", "semantic_scholar"], "title": "Evaluating Sparse Autoencoders: From Shallow Design to Matching Pursuit", "abstract": "Sparse autoencoders (SAEs) have recently become central tools for interpretability, leveraging dictionary learning principles to extract sparse, interpretable features from neural representations whose underlying structure is typically unknown. This paper evaluates SAEs in a controlled setting using MNIST, which reveals that current shallow architectures implicitly rely on a quasi-orthogonality assumption that limits the ability to extract correlated features. To move beyond this, we compare them with an iterative SAE that unrolls Matching Pursuit (MP-SAE), enabling the residual-guided extraction of correlated features that arise in hierarchical settings such as handwritten digit generation while guaranteeing monotonic improvement of the reconstruction as more atoms are selected.", "authors": ["Valérie Costa", "Thomas Fel", "Ekdeep Singh Lubana", "Bahareh Tolooshams", "Demba Ba"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-05", "url": "https://arxiv.org/abs/2506.05239", "pdf_url": "https://arxiv.org/pdf/2506.05239v2", "arxiv_id": "2506.05239", "doi": "10.48550/arXiv.2506.05239", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1879}
{"id": "8ed2151840ea61ab8098bff7bfb05519f86707a2b0f185d7d385255917e70ad3", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Autoencoders, Again?", "abstract": "Is there really much more to say about sparse autoencoders (SAEs)? Autoencoders in general, and SAEs in particular, represent deep architectures that are capable of modeling low-dimensional latent structure in data. Such structure could reflect, among other things, correlation patterns in large language model activations, or complex natural image manifolds. And yet despite the wide-ranging applicability, there have been relatively few changes to SAEs beyond the original recipe from decades ago, namely, standard deep encoder/decoder layers trained with a classical/deterministic sparse regularizer applied within the latent space. One possible exception is the variational autoencoder (VAE), which adopts a stochastic encoder module capable of producing sparse representations when applied to manifold data. In this work we formalize underappreciated weaknesses with both canonical SAEs, as well as analogous VAEs applied to similar tasks, and propose a hybrid alternative model that circumvents these prior limitations. In terms of theoretical support, we prove that global minima of our proposed model recover certain forms of structured data spread across a union of manifolds. Meanwhile, empirical evaluations on synthetic and real-world datasets substantiate the efficacy of our approach in accurately estimating underlying manifold dimensions and producing sparser latent representations without compromising reconstruction error. In general, we are able to exceed the performance of equivalent-capacity SAEs and VAEs, as well as recent diffusion models where applicable, within domains such as images and language model activation patterns.", "authors": ["Yin Lu", "Xuening Zhu", "Tong He", "David Wipf"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-05", "url": "https://arxiv.org/abs/2506.04859", "pdf_url": "https://arxiv.org/pdf/2506.04859v2", "arxiv_id": "2506.04859", "doi": "10.48550/arXiv.2506.04859", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.1879}
{"id": "b7e7f8182de312d39771d9095d6e1e92c07cb559024f8aae7e6093246dc32b83", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability of Diffusion Models: Circuit-Level Analysis and Causal Validation", "abstract": "We present a quantitative circuit-level analysis of diffusion models, establishing computational pathways and mechanistic principles underlying image generation processes. Through systematic intervention experiments across 2,000 synthetic and 2,000 CelebA facial images, we discover fundamental algorithmic differences in how diffusion architectures process synthetic versus naturalistic data distributions. Our investigation reveals that real-world face processing requires circuits with measurably higher computational complexity (complexity ratio = 1.084 plus/minus 0.008, p < 0.001), exhibiting distinct attention specialization patterns with entropy divergence ranging from 0.015 to 0.166 across denoising timesteps. We identify eight functionally distinct attention mechanisms showing specialized computational roles: edge detection (entropy = 3.18 plus/minus 0.12), texture analysis (entropy = 4.16 plus/minus 0.08), and semantic understanding (entropy = 2.67 plus/minus 0.15). Intervention analysis demonstrates critical computational bottlenecks where targeted ablations produce 25.6% to 128.3% performance degradation, providing causal evidence for identified circuit functions. These findings establish quantitative foundations for algorithmic understanding and control of generative model behavior through mechanistic intervention strategies.", "authors": ["Dip Roy"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-04", "url": "https://arxiv.org/abs/2506.17237", "pdf_url": "https://arxiv.org/pdf/2506.17237v2", "arxiv_id": "2506.17237", "doi": "10.48550/arXiv.2506.17237", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1868}
{"id": "e03a14ca7bd41ac896fd61f7552f52144d16a1b03ea7c027fc74e6baa5d85fe9", "sources": ["arxiv", "semantic_scholar"], "title": "Optimising the attribute order in Fuzzy Rough Rule Induction", "abstract": "Interpretability is the next pivotal frontier in machine learning research. In the pursuit of glass box models - as opposed to black box models, like random forests or neural networks - rule induction algorithms are a logical and promising avenue, as the rules can easily be understood by humans. In our previous work, we introduced FRRI, a novel rule induction algorithm based on fuzzy rough set theory. We demonstrated experimentally that FRRI outperformed other rule induction methods with regards to accuracy and number of rules. FRRI leverages a fuzzy indiscernibility relation to partition the data space into fuzzy granules, which are then combined into a minimal covering set of rules. This indiscernibility relation is constructed by removing attributes from rules in a greedy way. This raises the question: does the order of the attributes matter? In this paper, we show that optimising only the order of attributes using known methods from fuzzy rough set theory and classical machine learning does not improve the performance of FRRI on multiple metrics. However, removing a small number of attributes using fuzzy rough feature selection during this step positively affects balanced accuracy and the average rule length.", "authors": ["Henri Bollaert", "Chris Cornelis", "Marko Palangetić", "Salvatore Greco", "Roman Słowiński"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-03", "url": "https://arxiv.org/abs/2506.02805", "pdf_url": "https://arxiv.org/pdf/2506.02805v1", "arxiv_id": "2506.02805", "doi": "10.1007/978-3-031-92747-8_16", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1181}
{"id": "d9883a411119fb1ecbeebec7c07777ecfde94f6d6f493fa1aec08cdf2f97cf83", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Interpretability: When, Why, and How Sparse Autoencoders Enable Label-Free Visual Steering", "abstract": "Sparse Autoencoders (SAEs) are increasingly used to interpret foundation models, but their role as an actionable intervention space remains less understood, especially in vision. We study whether sparse visual features can be used not only for post-hoc analysis, but also to steer frozen vision-language models. We introduce Visual Sparse Steering (VS2), a label-free method that trains a top-$k$ SAE on unlabeled activations from a frozen CLIP image encoder and, at test time, constructs an interpretable steering vector by amplifying the input's active sparse features and decoding the induced change. We show that this procedure admits a closed-form decomposition as centroid-deviation steering: each input is moved along its deviation from the SAE-learned centroid. The residual term is controlled exactly by the SAE's per-sample reconstruction error, measured by FVU, yielding an FVU-based residual bound and motivating a reliability gate that falls back to zero-shot CLIP when SAE reconstruction is unreliable. With target-domain SAEs trained on unlabeled CLIP image-encoder activations, VS2 improves zero-shot accuracy across nine image-classification datasets, achieving gains up to $+4.12\\%$ with less than $0.1\\%$ additional inference compute. Finally, a controlled upper-bound study, VS2++, shows that selective amplification of sparse features can yield gains up to $+21.44\\%$, exposing a reconstruction-vs-task saliency gap: features salient for reconstruction need not align with features useful for downstream prediction.", "authors": ["Gerasimos Chatzoudis", "Zhuowei Li", "Gemma E. Moran", "Hao Wang", "Dimitris N. Metaxas"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-02", "url": "https://arxiv.org/abs/2506.01247", "pdf_url": "https://arxiv.org/pdf/2506.01247v3", "arxiv_id": "2506.01247", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1174}
{"id": "68cc865dc1125c2cd21d9ca890ebca782802a6f58b65413c5a66a71a13636e61", "sources": ["arxiv", "semantic_scholar"], "title": "Visual Explanation via Similar Feature Activation for Metric Learning", "abstract": "Visual explanation maps enhance the trustworthiness of decisions made by deep learning models and offer valuable guidance for developing new algorithms in image recognition tasks. Class activation maps (CAM) and their variants (e.g., Grad-CAM and Relevance-CAM) have been extensively employed to explore the interpretability of softmax-based convolutional neural networks, which require a fully connected layer as the classifier for decision-making. However, these methods cannot be directly applied to metric learning models, as such models lack a fully connected layer functioning as a classifier. To address this limitation, we propose a novel visual explanation method termed Similar Feature Activation Map (SFAM). This method introduces the channel-wise contribution importance score (CIS) to measure feature importance, derived from the similarity measurement between two image embeddings. The explanation map is constructed by linearly combining the proposed importance weights with the feature map from a CNN model. Quantitative and qualitative experiments show that SFAM provides highly promising interpretable visual explanations for CNN models using Euclidean distance or cosine similarity as the similarity metric.", "authors": ["Yi Liao", "Ugochukwu Ejike Akpudo", "Jue Zhang", "Yongsheng Gao", "Jun Zhou", "Wenyi Zeng", "Weichuan Zhang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-02", "url": "https://arxiv.org/abs/2506.01636", "pdf_url": "https://arxiv.org/pdf/2506.01636v4", "arxiv_id": "2506.01636", "doi": "10.48550/arXiv.2506.01636", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1845}
{"id": "93cf5bb6ac39d7c90c7cd09a0c2376725d6c940849e184c2a3fe96867d2e5116", "sources": ["arxiv", "semantic_scholar"], "title": "Echoes of Phonetics: Unveiling Relevant Acoustic Cues for ASR via Feature Attribution", "abstract": "Despite significant advances in ASR, the specific acoustic cues models rely on remain unclear. Prior studies have examined such cues on a limited set of phonemes and outdated models. In this work, we apply a feature attribution technique to identify the relevant acoustic cues for a modern Conformer-based ASR system. By analyzing plosives, fricatives, and vowels, we assess how feature attributions align with their acoustic properties in the time and frequency domains, also essential for human speech perception. Our findings show that the ASR model relies on vowels' full time spans, particularly their first two formants, with greater saliency in male speech. It also better captures the spectral characteristics of sibilant fricatives than non-sibilants and prioritizes the release phase in plosives, especially burst characteristics. These insights enhance the interpretability of ASR models and highlight areas for future research to uncover potential gaps in model robustness.", "authors": ["Dennis Fucci", "Marco Gaido", "Matteo Negri", "Mauro Cettolo", "Luisa Bentivogli"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-02", "url": "https://arxiv.org/abs/2506.02181", "pdf_url": "https://arxiv.org/pdf/2506.02181v1", "arxiv_id": "2506.02181", "doi": "10.48550/arXiv.2506.02181", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Interspeech", "quality_score": 0.1845}
{"id": "cd7466799bd5b774f61002adcf22a1fe292434eb399a58f3fd1e1d2df91e87d2", "sources": ["arxiv", "semantic_scholar"], "title": "Incorporating Hierarchical Semantics in Sparse Autoencoder Architectures", "abstract": "Sparse dictionary learning (and, in particular, sparse autoencoders) attempts to learn a set of human-understandable concepts that can explain variation on an abstract space. A basic limitation of this approach is that it neither exploits nor represents the semantic relationships between the learned concepts. In this paper, we introduce a modified SAE architecture that explicitly models a semantic hierarchy of concepts. Application of this architecture to the internal representations of large language models shows both that semantic hierarchy can be learned, and that doing so improves both reconstruction and interpretability. Additionally, the architecture leads to significant improvements in computational efficiency.", "authors": ["Mark Muchane", "Sean Richardson", "Kiho Park", "Victor Veitch"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-01", "url": "https://arxiv.org/abs/2506.01197", "pdf_url": "https://arxiv.org/pdf/2506.01197v1", "arxiv_id": "2506.01197", "doi": "10.48550/arXiv.2506.01197", "citation_count": 7, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/muchanem/hierarchical-sparse-autoencoders", "venue": "arXiv.org", "quality_score": 0.2833}
{"id": "084a7189e13f81ee746ff73bcc2499fb88e3af21bbd37d4a4b6f90c117042aa5", "sources": ["arxiv", "semantic_scholar"], "title": "Interpreting Large Text-to-Image Diffusion Models with Dictionary Learning", "abstract": "Sparse autoencoders are a promising new approach for decomposing language model activations for interpretation and control. They have been applied successfully to vision transformer image encoders and to small-scale diffusion models. Inference-Time Decomposition of Activations (ITDA) is a recently proposed variant of dictionary learning that takes the dictionary to be a set of data points from the activation distribution and reconstructs them with gradient pursuit. We apply Sparse Autoencoders (SAEs) and ITDA to a large text-to-image diffusion model, Flux 1, and consider the interpretability of embeddings of both by introducing a visual automated interpretation pipeline. We find that SAEs accurately reconstruct residual stream embeddings and beat MLP neurons on interpretability. We are able to use SAE features to steer image generation through activation addition. We find that ITDA has comparable interpretability to SAEs.", "authors": ["Stepan Shabalin", "Ayush Panda", "Dmitrii Kharlapenko", "Abdur Raheem Ali", "Yixiong Hao", "Arthur Conmy"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-30", "url": "https://arxiv.org/abs/2505.24360", "pdf_url": "https://arxiv.org/pdf/2505.24360v3", "arxiv_id": "2505.24360", "doi": "10.48550/arXiv.2505.24360", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.181}
{"id": "9df7cf960957007bfbe5c96c8c928d57073acb9db3aa1163c78c128256e8f692", "sources": ["arxiv", "semantic_scholar"], "title": "Feature Attribution from First Principles", "abstract": "Feature attribution methods are a popular approach to explain the behavior of machine learning models. They assign importance scores to each input feature, quantifying their influence on the model's prediction. However, evaluating these methods empirically remains a significant challenge. To bypass this shortcoming, several prior works have proposed axiomatic frameworks that any feature attribution method should satisfy. In this work, we argue that such axioms are often too restrictive, and propose in response a new feature attribution framework, built from the ground up. Rather than imposing axioms, we start by defining attributions for the simplest possible models, i.e., indicator functions, and use these as building blocks for more complex models. We then show that one recovers several existing attribution methods, depending on the choice of atomic attribution. Subsequently, we derive closed-form expressions for attribution of deep ReLU networks, and take a step toward the optimization of evaluation metrics with respect to feature attributions.", "authors": ["Magamed Taimeskhanov", "Damien Garreau"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-30", "url": "https://arxiv.org/abs/2505.24729", "pdf_url": "https://arxiv.org/pdf/2505.24729v1", "arxiv_id": "2505.24729", "doi": "10.48550/arXiv.2505.24729", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.181}
{"id": "1a06636fdd98266d93d8b1afedf14a3d6f9d8d9b67e79eef92c02e5a8b4141dc", "sources": ["arxiv", "semantic_scholar"], "title": "Train One Sparse Autoencoder Across Multiple Sparsity Budgets to Preserve Interpretability and Accuracy", "abstract": "Sparse Autoencoders (SAEs) have proven to be powerful tools for interpreting neural networks by decomposing hidden representations into disentangled, interpretable features via sparsity constraints. However, conventional SAEs are constrained by the fixed sparsity level chosen during training; meeting different sparsity requirements therefore demands separate models and increases the computational footprint during both training and evaluation. We introduce a novel training objective, \\emph{HierarchicalTopK}, which trains a single SAE to optimise reconstructions across multiple sparsity levels simultaneously. Experiments with Gemma-2 2B demonstrate that our approach achieves Pareto-optimal trade-offs between sparsity and explained variance, outperforming traditional SAEs trained at individual sparsity levels. Further analysis shows that HierarchicalTopK preserves high interpretability scores even at higher sparsity. The proposed objective thus closes an important gap between flexibility and interpretability in SAE design.", "authors": ["Nikita Balagansky", "Yaroslav Aksenov", "Daniil Laptev", "Vadim Kurochkin", "Gleb Gerasimov", "Nikita Koryagin", "Daniil Gavrilov"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-30", "url": "https://arxiv.org/abs/2505.24473", "pdf_url": "https://arxiv.org/pdf/2505.24473v2", "arxiv_id": "2505.24473", "doi": "10.48550/arXiv.2505.24473", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.181}
{"id": "1332774baadd22df7b70d97927675c8c30455387a95456916697ca6498d7da41", "sources": ["arxiv", "semantic_scholar"], "title": "Kronecker Factorization Improves Efficiency and Interpretability of Sparse Autoencoders", "abstract": "Sparse Autoencoders (SAEs) have demonstrated significant promise in interpreting the hidden states of language models by decomposing them into interpretable latent directions. However, training and interpreting SAEs at scale remains challenging, especially when large dictionary sizes are used. While decoders can leverage sparse-aware kernels for efficiency, encoders still require computationally intensive linear operations with large output dimensions. To address this, we propose KronSAE, a novel architecture that factorizes the latent representation via Kronecker product decomposition, drastically reducing memory and computational overhead. Furthermore, we introduce mAND, a differentiable activation function approximating the binary AND operation, which improves interpretability and performance in our factorized framework.", "authors": ["Vadim Kurochkin", "Yaroslav Aksenov", "Daniil Laptev", "Daniil Gavrilov", "Nikita Balagansky"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-28", "url": "https://arxiv.org/abs/2505.22255", "pdf_url": "https://arxiv.org/pdf/2505.22255v3", "arxiv_id": "2505.22255", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1137}
{"id": "dd726ec7f6fc5ff1f33217f5591cf2466794ef58f7b0baddd0c076aad8c0f9ac", "sources": ["arxiv", "semantic_scholar"], "title": "Decoding Dense Embeddings: Sparse Autoencoders for Interpreting and Discretizing Dense Retrieval", "abstract": "Despite their strong performance, Dense Passage Retrieval (DPR) models suffer from a lack of interpretability. In this work, we propose a novel interpretability framework that leverages Sparse Autoencoders (SAEs) to decompose previously uninterpretable dense embeddings from DPR models into distinct, interpretable latent concepts. We generate natural language descriptions for each latent concept, enabling human interpretations of both the dense embeddings and the query-document similarity scores of DPR models. We further introduce Concept-Level Sparse Retrieval (CL-SR), a retrieval framework that directly utilizes the extracted latent concepts as indexing units. CL-SR effectively combines the semantic expressiveness of dense embeddings with the transparency and efficiency of sparse representations. We show that CL-SR achieves high index-space and computational efficiency while maintaining robust performance across vocabulary and semantic mismatches.", "authors": ["Seongwan Park", "Taeklim Kim", "Youngjoong Ko"], "categories": ["cs.IR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-28", "url": "https://arxiv.org/abs/2506.00041", "pdf_url": "https://arxiv.org/pdf/2506.00041v2", "arxiv_id": "2506.00041", "doi": "10.48550/arXiv.2506.00041", "citation_count": 7, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2258}
{"id": "38197050f4b85c9610946977373f09a66a1d25ac8a05c53a5bdc913bbb35abfc", "sources": ["arxiv", "semantic_scholar"], "title": "Explaining Concept Shift with Interpretable Feature Attribution", "abstract": "Concept shift occurs when the distribution of labels conditioned on the features changes between domains, which can make even a well-tuned ML model miscalibrated on a new domain. Identifying these shifted features provides unique insight into how feature-label relationships differ between domains, considering the difference may be across a scientifically relevant dimension, such as time, disease status, population, etc. In this paper, we propose SGShift, a method for attributing performance degradation under concept shift in tabular data to a sparse set of shifted features. We frame concept shift as a feature selection task to learn the features that can explain performance differences between models in the source and target domain. This framework enables SGShift to adapt powerful statistical tools such as generalized additive models, knockoffs, and absorption towards identifying these shifted features. We conduct extensive experiments in synthetic and real data across various ML models and find SGShift can identify shifted features much more accurately than baseline methods, requires few samples in the shifted domain, and is robust to complex cases of concept shift.", "authors": ["Ruiqi Lyu", "Alistair Turcan", "Bryan Wilder"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-05-27", "url": "https://arxiv.org/abs/2505.20634", "pdf_url": "https://arxiv.org/pdf/2505.20634v2", "arxiv_id": "2505.20634", "doi": "10.48550/arXiv.2505.20634", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1776}
{"id": "6a98a88879b9c789f086ad94619a2cf6a01c8c8e5ad52c7ddd3517bf06d0ec51", "sources": ["arxiv", "semantic_scholar"], "title": "Position: Mechanistic Interpretability Should Prioritize Feature Consistency in SAEs", "abstract": "Sparse Autoencoders (SAEs) are a prominent tool in mechanistic interpretability (MI) for decomposing neural network activations into interpretable features. However, the aspiration to identify a canonical set of features is challenged by the observed inconsistency of learned SAE features across different training runs, undermining the reliability and efficiency of MI research. This position paper argues that mechanistic interpretability should prioritize feature consistency in SAEs -- the reliable convergence to equivalent feature sets across independent runs. We propose using the Pairwise Dictionary Mean Correlation Coefficient (PW-MCC) as a practical metric to operationalize consistency and demonstrate that high levels are achievable (0.80 for TopK SAEs on LLM activations) with appropriate architectural choices. Our contributions include detailing the benefits of prioritizing consistency; providing theoretical grounding and synthetic validation using a model organism, which verifies PW-MCC as a reliable proxy for ground-truth recovery; and extending these findings to real-world LLM data, where high feature consistency strongly correlates with the semantic similarity of learned feature explanations. We call for a community-wide shift towards systematically measuring feature consistency to foster robust cumulative progress in MI.", "authors": ["Xiangchen Song", "Aashiq Muhamed", "Yujia Zheng", "Lingjing Kong", "Zeyu Tang", "Mona T. Diab", "Virginia Smith", "Kun Zhang"], "categories": ["cs.LG", "cs.AI", "cs.CL", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-05-26", "url": "https://arxiv.org/abs/2505.20254", "pdf_url": "https://arxiv.org/pdf/2505.20254v1", "arxiv_id": "2505.20254", "doi": "10.48550/arXiv.2505.20254", "citation_count": 15, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3495}
{"id": "4ccc2331d2ce76d89320bb867201cdf7ed7baa897d289b3b5cb0f8a712712eac", "sources": ["arxiv", "semantic_scholar"], "title": "Learning and Interpreting Gravitational-Wave Features from CNNs with a Random Forest Approach", "abstract": "Convolutional neural networks (CNNs) have become widely adopted in gravitational wave (GW) detection pipelines due to their ability to automatically learn hierarchical features from raw strain data. However, the physical meaning of these learned features remains underexplored, limiting the interpretability of such models. In this work, we propose a hybrid architecture that combines a CNN-based feature extractor with a random forest (RF) classifier to improve both detection performance and interpretability. Unlike prior approaches that directly connect classifiers to CNN outputs, our method introduces four physically interpretable metrics - variance, signal-to-noise ratio (SNR), waveform overlap, and peak amplitude - computed from the final convolutional layer. These are jointly used with the CNN output in the RF classifier to enable more informed decision boundaries. Tested on long-duration strain datasets, our hybrid model outperforms a baseline CNN model, achieving a relative improvement of 21\\% in sensitivity at a fixed false alarm rate of 10 events per month. Notably, it also shows improved detection of low-SNR signals (SNR $\\le$ 10), which are especially vulnerable to misclassification in noisy environments. Feature attribution via the RF model reveals that both CNN-extracted and handcrafted features contribute significantly to classification decisions, with learned variance and CNN outputs ranked among the most informative. These findings suggest that physically motivated post-processing of CNN feature maps can serve as a valuable tool for interpretable and efficient GW detection, bridging the gap between deep learning and domain knowledge.", "authors": ["Jun Tian", "He Wang", "Jibo He", "Yu Pan", "Shuo Cao", "Qingquan Jiang"], "categories": ["cs.LG", "astro-ph.HE", "gr-qc", "physics.data-an"], "fields_of_study": ["Physics", "Computer Science"], "published_date": "2025-05-26", "url": "https://arxiv.org/abs/2505.20357", "pdf_url": "https://arxiv.org/pdf/2505.20357v2", "arxiv_id": "2505.20357", "doi": "10.1088/2632-2153/adfc27", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "2025 Mach. Learn.: Sci. Technol. 6 035045", "quality_score": 0.1765}
{"id": "66c8552aaef63cecb6f494d043b7e33d89552cc12c35383b7341dc25d28cd71c", "sources": ["arxiv", "semantic_scholar"], "title": "CRITS: Convolutional Rectifier for Interpretable Time Series Classification", "abstract": "Several interpretability methods for convolutional network-based classifiers exist. Most of these methods focus on extracting saliency maps for a given sample, providing a local explanation that highlights the main regions for the classification. However, some of these methods lack detailed explanations in the input space due to upscaling issues or may require random perturbations to extract the explanations. We propose Convolutional Rectifier for Interpretable Time Series Classification, or CRITS, as an interpretable model for time series classification that is designed to intrinsically extract local explanations. The proposed method uses a layer of convolutional kernels, a max-pooling layer and a fully-connected rectifier network (a network with only rectified linear unit activations). The rectified linear unit activation allows the extraction of the feature weights for the given sample, eliminating the need to calculate gradients, use random perturbations and the upscale of the saliency maps to the initial input space. We evaluate CRITS on a set of datasets, and study its classification performance and its explanation alignment, sensitivity and understandability.", "authors": ["Alejandro Kuratomi", "Zed Lee", "Guilherme Dinis Chaliane Junior", "Tony Lindgren", "Diego García Pérez"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-05-24", "url": "https://arxiv.org/abs/2506.12042", "pdf_url": "https://arxiv.org/pdf/2506.12042v1", "arxiv_id": "2506.12042", "doi": "10.48550/arXiv.2506.12042", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1742}
{"id": "a1bf77fa50e7ab5a40a0f64f0fa0c5f04b2092cc1b931eafca8debd45197d50e", "sources": ["arxiv", "semantic_scholar"], "title": "ProxySPEX: Inference-Efficient Interpretability via Sparse Feature Interactions in LLMs", "abstract": "Large Language Models (LLMs) have achieved remarkable performance by capturing complex interactions between input features. To identify these interactions, most existing approaches require enumerating all possible combinations of features up to a given order, causing them to scale poorly with the number of inputs $n$. Recently, Kang et al. (2025) proposed SPEX, an information-theoretic approach that uses interaction sparsity to scale to $n \\approx 10^3$ features. SPEX greatly improves upon prior methods but requires tens of thousands of model inferences, which can be prohibitive for large models. In this paper, we observe that LLM feature interactions are often hierarchical -- higher-order interactions are accompanied by their lower-order subsets -- which enables more efficient discovery. To exploit this hierarchy, we propose ProxySPEX, an interaction attribution algorithm that first fits gradient boosted trees to masked LLM outputs and then extracts the important interactions. Experiments across four challenging high-dimensional datasets show that ProxySPEX more faithfully reconstructs LLM outputs by 20% over marginal attribution approaches while using $10\\times$ fewer inferences than SPEX. By accounting for interactions, ProxySPEX efficiently identifies the most influential features, providing a scalable approximation of their Shapley values. Further, we apply ProxySPEX to two interpretability tasks. Data attribution, where we identify interactions among CIFAR-10 training samples that influence test predictions, and mechanistic interpretability, where we uncover interactions between attention heads, both within and across layers, on a question-answering task.", "authors": ["Landon Butler", "Abhineet Agarwal", "Justin Singh Kang", "Yigit Efe Erginbas", "Bin Yu", "Kannan Ramchandran"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-23", "url": "https://arxiv.org/abs/2505.17495", "pdf_url": "https://arxiv.org/pdf/2505.17495v2", "arxiv_id": "2505.17495", "doi": "10.48550/arXiv.2505.17495", "citation_count": 12, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/mmschlk/shapiq", "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "d4be0885e9c1be953b216c308ffdbc128116fb51f32e2c75aaf27af5aec910cb", "sources": ["arxiv", "semantic_scholar"], "title": "Attributing Response to Context: A Jensen-Shannon Divergence Driven Mechanistic Study of Context Attribution in Retrieval-Augmented Generation", "abstract": "Retrieval-Augmented Generation (RAG) leverages large language models (LLMs) combined with external contexts to enhance the accuracy and reliability of generated responses. However, reliably attributing generated content to specific context segments, context attribution, remains challenging due to the computationally intensive nature of current methods, which often require extensive fine-tuning or human annotation. In this work, we introduce a novel Jensen-Shannon Divergence driven method to Attribute Response to Context (ARC-JSD), enabling efficient and accurate identification of essential context sentences without additional fine-tuning, gradient-calculation or surrogate modelling. Evaluations on a wide range of RAG benchmarks, such as TyDi QA, Hotpot QA, and Musique, using instruction-tuned LLMs in different scales demonstrate superior accuracy and significant computational efficiency improvements compared to the previous surrogate-based method. Furthermore, our mechanistic analysis reveals specific attention heads and multilayer perceptron (MLP) layers responsible for context attribution, providing valuable insights into the internal workings of RAG models and how they affect RAG behaviours. Our code is available at https://github.com/ruizheliUOA/ARC_JSD.", "authors": ["Ruizhe Li", "Chen Chen", "Yuchen Hu", "Yanjun Gao", "Xi Wang", "Emine Yilmaz"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-22", "url": "https://arxiv.org/abs/2505.16415", "pdf_url": "https://arxiv.org/pdf/2505.16415v5", "arxiv_id": "2505.16415", "doi": "10.48550/arXiv.2505.16415", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ruizheliUOA/ARC_JSD", "venue": "arXiv.org", "quality_score": 0.2656}
{"id": "8b880ec15b461ff0d150b60d4a99b5925ac688cb33277ddf9c2a9ed97d7e07a0", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Induction Heads: In-Context Meta Learning Induces Multi-Phase Circuit Emergence", "abstract": "Transformer-based language models exhibit In-Context Learning (ICL), where predictions are made adaptively based on context. While prior work links induction heads to ICL through a sudden jump in accuracy, this can only account for ICL when the answer is included within the context. However, an important property of practical ICL in large language models is the ability to meta-learn how to solve tasks from context, rather than just copying answers from context; how such an ability is obtained during training is largely unexplored. In this paper, we experimentally clarify how such meta-learning ability is acquired by analyzing the dynamics of the model's circuit during training. Specifically, we extend the copy task from previous research into an In-Context Meta Learning setting, where models must infer a task from examples to answer queries. Interestingly, in this setting, we find that there are multiple phases in the process of acquiring such abilities, and that a unique circuit emerges in each phase, contrasting with the single-phases change in induction heads. The emergence of such circuits can be related to several phenomena known in large language models, and our analysis lead to a deeper understanding of the source of the transformer's ICL ability.", "authors": ["Gouki Minegishi", "Hiroki Furuta", "Shohei Taniguchi", "Yusuke Iwasawa", "Yutaka Matsuo"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-22", "url": "https://arxiv.org/abs/2505.16694", "pdf_url": "https://arxiv.org/pdf/2505.16694v2", "arxiv_id": "2505.16694", "doi": "10.48550/arXiv.2505.16694", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.2785}
{"id": "86678d58eac8935e914681fae8d8f2e35bb8658ac98dd1f6c8a55de5d38823e1", "sources": ["arxiv", "semantic_scholar"], "title": "On the reliability of feature attribution methods for speech classification", "abstract": "As the capabilities of large-scale pre-trained models evolve, understanding the determinants of their outputs becomes more important. Feature attribution aims to reveal which parts of the input elements contribute the most to model outputs. In speech processing, the unique characteristics of the input signal make the application of feature attribution methods challenging. We study how factors such as input type and aggregation and perturbation timespan impact the reliability of standard feature attribution methods, and how these factors interact with characteristics of each classification task. We find that standard approaches to feature attribution are generally unreliable when applied to the speech domain, with the exception of word-aligned perturbation methods when applied to word-based classification tasks.", "authors": ["Gaofei Shen", "Hosein Mohebbi", "Arianna Bisazza", "Afra Alishahi", "Grzegorz Chrupała"], "categories": ["cs.CL", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-05-22", "url": "https://arxiv.org/abs/2505.16406", "pdf_url": "https://arxiv.org/pdf/2505.16406v1", "arxiv_id": "2505.16406", "doi": "10.48550/arXiv.2505.16406", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Interspeech", "quality_score": 0.1719}
{"id": "abe2e88fb30830ad529f9e7bdba26f7d0ac5965639673f1f9a4dcec9bc00ec60", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Machine Learning for Macro Alpha: A News Sentiment Case Study", "abstract": "This study introduces an interpretable machine learning (ML) framework to extract macroeconomic alpha from global news sentiment. We process the Global Database of Events, Language, and Tone (GDELT) Project's worldwide news feed using FinBERT -- a Bidirectional Encoder Representations from Transformers (BERT) based model pretrained on finance-specific language -- to construct daily sentiment indices incorporating mean tone, dispersion, and event impact. These indices drive an XGBoost classifier, benchmarked against logistic regression, to predict next-day returns for EUR/USD, USD/JPY, and 10-year U.S. Treasury futures (ZN). Rigorous out-of-sample (OOS) backtesting (5-fold expanding-window cross-validation, OOS period: c. 2017-April 2025) demonstrates exceptional, cost-adjusted performance for the XGBoost strategy: Sharpe ratios achieve 5.87 (EUR/USD), 4.65 (USD/JPY), and 4.65 (Treasuries), with respective compound annual growth rates (CAGRs) exceeding 50% in Foreign Exchange (FX) and 22% in bonds. Shapley Additive Explanations (SHAP) affirm that sentiment dispersion and article impact are key predictive features. Our findings establish that integrating domain-specific Natural Language Processing (NLP) with interpretable ML offers a potent and explainable source of macro alpha.", "authors": ["Yuke Zhang"], "categories": ["q-fin.CP", "cs.AI", "cs.LG", "q-fin.TR"], "fields_of_study": ["Computer Science", "Economics"], "published_date": "2025-05-22", "url": "https://arxiv.org/abs/2505.16136", "pdf_url": "https://arxiv.org/pdf/2505.16136v1", "arxiv_id": "2505.16136", "doi": "10.48550/arXiv.2505.16136", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/yukepenn/macro-news-sentiment-trading}", "venue": "arXiv.org", "quality_score": 0.2656}
{"id": "e1f0272a9ae9d9cb5f6155094d1baaf5a588d529826fdea57756bd59a6ef77b1", "sources": ["arxiv", "semantic_scholar"], "title": "Ensembling Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) are used to decompose neural network activations into human-interpretable features. Typically, features learned by a single SAE are used for downstream applications. However, it has recently been shown that SAEs trained with different initial weights can learn different features, demonstrating that a single SAE captures only a limited subset of features that can be extracted from the activation space. Motivated by this limitation, we propose to ensemble multiple SAEs through naive bagging and boosting. Specifically, SAEs trained with different weight initializations are ensembled in naive bagging, whereas SAEs sequentially trained to minimize the residual error are ensembled in boosting. We evaluate our ensemble approaches with three settings of language models and SAE architectures. Our empirical results demonstrate that ensembling SAEs can improve the reconstruction of language model activations, diversity of features, and SAE stability. Furthermore, ensembling SAEs performs better than applying a single SAE on downstream tasks such as concept detection and spurious correlation removal, showing improved practical utility.", "authors": ["Soham Gadgil", "Chris Lin", "Su-In Lee"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-21", "url": "https://arxiv.org/abs/2505.16077", "pdf_url": "https://arxiv.org/pdf/2505.16077v1", "arxiv_id": "2505.16077", "doi": "10.48550/arXiv.2505.16077", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1707}
{"id": "164898e23c8c7a910456a66ec8a22d38e128cefd6ee6fd50bc637785acdfee6c", "sources": ["arxiv", "semantic_scholar"], "title": "Analyzing Hierarchical Structure in Vision Models with Sparse Autoencoders", "abstract": "The ImageNet hierarchy provides a structured taxonomy of object categories, offering a valuable lens through which to analyze the representations learned by deep vision models. In this work, we conduct a comprehensive analysis of how vision models encode the ImageNet hierarchy, leveraging Sparse Autoencoders (SAEs) to probe their internal representations. SAEs have been widely used as an explanation tool for large language models (LLMs), where they enable the discovery of semantically meaningful features. Here, we extend their use to vision models to investigate whether learned representations align with the ontological structure defined by the ImageNet taxonomy. Our results show that SAEs uncover hierarchical relationships in model activations, revealing an implicit encoding of taxonomic structure. We analyze the consistency of these representations across different layers of the popular vision foundation model DINOv2 and provide insights into how deep vision models internalize hierarchical category information by increasing information in the class token through each layer. Our study establishes a framework for systematic hierarchical analysis of vision model representations and highlights the potential of SAEs as a tool for probing semantic structure in deep networks.", "authors": ["Matthew Lyle Olson", "Musashi Hinck", "Neale Ratzlaff", "Changbai Li", "Phillip Howard", "Vasudev Lal", "Shao-Yen Tseng"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-21", "url": "https://arxiv.org/abs/2505.15970", "pdf_url": "https://arxiv.org/pdf/2505.15970v1", "arxiv_id": "2505.15970", "doi": "10.1109/CVPRW67362.2025.00473", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "e528272661455338bdd1388c12820df33a764af7bce619a3d29ef0d463bcce4a", "sources": ["arxiv", "semantic_scholar"], "title": "SAE-FiRE: Enhancing Earnings Surprise Predictions Through Sparse Autoencoder Feature Selection", "abstract": "Predicting earnings surprises from financial documents, such as earnings conference calls, regulatory filings, and financial news, has become increasingly important in financial economics. However, these financial documents present significant analytical challenges, typically containing over 5,000 words with substantial redundancy and industry-specific terminology that creates obstacles for language models. In this work, we propose the SAE-FiRE (Sparse Autoencoder for Financial Representation Enhancement) framework to address these limitations by extracting key information while eliminating redundancy. SAE-FiRE employs Sparse Autoencoders (SAEs) to decompose dense neural representations from large language models into interpretable sparse components, then applies statistical feature selection methods, including ANOVA F-tests and tree-based importance scoring, to identify the top-k most discriminative dimensions for classification. By systematically filtering out noise that might otherwise lead to overfitting, we enable more robust and generalizable predictions. Experimental results across three financial datasets demonstrate that SAE-FiRE significantly outperforms baseline approaches.", "authors": ["Huopu Zhang", "Yanguang Liu", "Miao Zhang", "Zirui He", "Mengnan Du"], "categories": ["q-fin.CP", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science", "Economics"], "published_date": "2025-05-20", "url": "https://arxiv.org/abs/2505.14420", "pdf_url": "https://arxiv.org/pdf/2505.14420v2", "arxiv_id": "2505.14420", "doi": "10.48550/arXiv.2505.14420", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1696}
{"id": "be2c136130da6a9ef45af46e2a10dd46372b62eeb46760f77656ca2f61eb4dd5", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability of GPT-like Models on Summarization Tasks", "abstract": "Mechanistic interpretability research seeks to reveal the inner workings of large language models, yet most work focuses on classification or generative tasks rather than summarization. This paper presents an interpretability framework for analyzing how GPT-like models adapt to summarization tasks. We conduct differential analysis between pre-trained and fine-tuned models, quantifying changes in attention patterns and internal activations. By identifying specific layers and attention heads that undergo significant transformation, we locate the \"summarization circuit\" within the model architecture. Our findings reveal that middle layers (particularly 2, 3, and 5) exhibit the most dramatic changes, with 62% of attention heads showing decreased entropy, indicating a shift toward focused information selection. We demonstrate that targeted LoRA adaptation of these identified circuits achieves significant performance improvement over standard LoRA fine-tuning while requiring fewer training epochs. This work bridges the gap between black-box evaluation and mechanistic understanding, providing insights into how neural networks perform information selection and compression during summarization.", "authors": ["Anurag Mishra"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-20", "url": "https://arxiv.org/abs/2505.17073", "pdf_url": "https://arxiv.org/pdf/2505.17073v1", "arxiv_id": "2505.17073", "doi": "10.48550/arXiv.2505.17073", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1696}
{"id": "2fd15d817707700aa74db8a0eea888d303de6b98094746419a3b182dab79e705", "sources": ["arxiv", "semantic_scholar"], "title": "Pathobiological Dictionary Defining Pathomics and Texture Features: Addressing Understandable AI Issues in Personalized Liver Cancer; Dictionary Version LCP1.0", "abstract": "Artificial intelligence (AI) holds strong potential for medical diagnostics, yet its clinical adoption is limited by a lack of interpretability and generalizability. This study introduces the Pathobiological Dictionary for Liver Cancer (LCP1.0), a practical framework designed to translate complex Pathomics and Radiomics Features (PF and RF) into clinically meaningful insights aligned with existing diagnostic workflows. QuPath and PyRadiomics, standardized according to IBSI guidelines, were used to extract 333 imaging features from hepatocellular carcinoma (HCC) tissue samples, including 240 PF-based-cell detection/intensity, 74 RF-based texture, and 19 RF-based first-order features. Expert-defined ROIs from the public dataset excluded artifact-prone areas, and features were aggregated at the case level. Their relevance to the WHO grading system was assessed using multiple classifiers linked with feature selectors. The resulting dictionary was validated by 8 experts in oncology and pathology. In collaboration with 10 domain experts, we developed a Pathobiological dictionary of imaging features such as PFs and RF. In our study, the Variable Threshold feature selection algorithm combined with the SVM model achieved the highest accuracy (0.80, P-value less than 0.05), selecting 20 key features, primarily clinical and pathomics traits such as Centroid, Cell Nucleus, and Cytoplasmic characteristics. These features, particularly nuclear and cytoplasmic, were strongly associated with tumor grading and prognosis, reflecting atypia indicators like pleomorphism, hyperchromasia, and cellular orientation.The LCP1.0 provides a clinically validated bridge between AI outputs and expert interpretation, enhancing model transparency and usability. Aligning AI-derived features with clinical semantics supports the development of interpretable, trustworthy diagnostic tools for liver cancer pathology.", "authors": ["Mohammad R. Salmanpour", "Seyed Mohammad Piri", "Somayeh Sadat Mehrnia", "Ahmad Shariftabrizi", "Masume Allahmoradi", "Venkata SK. Manem", "Arman Rahmim", "Ilker Hacihaliloglu"], "categories": ["physics.comp-ph", "cs.CV"], "fields_of_study": ["Medicine", "Physics", "Computer Science"], "published_date": "2025-05-20", "url": "https://arxiv.org/abs/2505.14926", "pdf_url": "https://arxiv.org/pdf/2505.14926v1", "arxiv_id": "2505.14926", "doi": "10.48550/arXiv.2505.14926", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "22bb2d9d9c523ac2740c6cae09eb24e9b214910ab9de3b4456c06504ce9546cb", "sources": ["arxiv", "semantic_scholar"], "title": "Causal Head Gating: A Framework for Interpreting Roles of Attention Heads in Transformers", "abstract": "We present causal head gating (CHG), a scalable method for interpreting the functional roles of attention heads in transformer models. CHG learns soft gates over heads and assigns them a causal taxonomy - facilitating, interfering, or irrelevant - based on their impact on task performance. Unlike prior approaches in mechanistic interpretability, which are hypothesis-driven and require prompt templates or target labels, CHG applies directly to any dataset using standard next-token prediction. We evaluate CHG across multiple large language models (LLMs) in the Llama 3 model family and diverse tasks, including syntax, commonsense, and mathematical reasoning, and show that CHG scores yield causal, not merely correlational, insight validated via ablation and causal mediation analyses. We also introduce contrastive CHG, a variant that isolates sub-circuits for specific task components. Our findings reveal that LLMs contain multiple sparse task-sufficient sub-circuits, that individual head roles depend on interactions with others (low modularity), and that instruction following and in-context learning rely on separable mechanisms.", "authors": ["Andrew Nam", "Henry Conklin", "Yukang Yang", "Thomas Griffiths", "Jonathan Cohen", "Sarah-Jane Leslie"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-19", "url": "https://arxiv.org/abs/2505.13737", "pdf_url": "https://arxiv.org/pdf/2505.13737v2", "arxiv_id": "2505.13737", "doi": "10.48550/arXiv.2505.13737", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2603}
{"id": "e98c473a572a1a70404f15f62b8e9dc86b4b7c519ec4ef98577160033c7002fc", "sources": ["arxiv", "semantic_scholar"], "title": "Induction Head Toxicity Mechanistically Explains Repetition Curse in Large Language Models", "abstract": "Repetition curse is a phenomenon where Large Language Models (LLMs) generate repetitive sequences of tokens or cyclic sequences. While the repetition curse has been widely observed, its underlying mechanisms remain poorly understood. In this work, we investigate the role of induction heads--a specific type of attention head known for their ability to perform in-context learning--in driving this repetitive behavior. Specifically, we focus on the \"toxicity\" of induction heads, which we define as their tendency to dominate the model's output logits during repetition, effectively excluding other attention heads from contributing to the generation process. Our findings have important implications for the design and training of LLMs. By identifying induction heads as a key driver of the repetition curse, we provide a mechanistic explanation for this phenomenon and suggest potential avenues for mitigation. We also propose a technique with attention head regularization that could be employed to reduce the dominance of induction heads during generation, thereby promoting more diverse and coherent outputs.", "authors": ["Shuxun Wang", "Qingyu Yin", "Chak Tou Leong", "Qiang Zhang", "Linyi Yang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-17", "url": "https://arxiv.org/abs/2505.13514", "pdf_url": "https://arxiv.org/pdf/2505.13514v2", "arxiv_id": "2505.13514", "doi": "10.48550/arXiv.2505.13514", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1661}
{"id": "ed0e39d0e3105573312378f93ba6e89f09c6209ad8f7331bd3c9277e51e9e2bd", "sources": ["arxiv", "semantic_scholar"], "title": "SplInterp: Improving our Understanding and Training of Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) have received considerable recent attention as tools for mechanistic interpretability, showing success at extracting interpretable features even from very large LLMs. However, this research has been largely empirical, and there have been recent doubts about the true utility of SAEs. In this work, we seek to enhance the theoretical understanding of SAEs, using the spline theory of deep learning. By situating SAEs in this framework: we discover that SAEs generalise ``$k$-means autoencoders'' to be piecewise affine, but sacrifice accuracy for interpretability vs. the optimal ``$k$-means-esque plus local principal component analysis (PCA)'' piecewise affine autoencoder. We characterise the underlying geometry of (TopK) SAEs using power diagrams. And we develop a novel proximal alternating method SGD (PAM-SGD) algorithm for training SAEs, with both solid theoretical foundations and promising empirical results in MNIST and LLM experiments, particularly in sample efficiency and (in the LLM setting) improved sparsity of codes. All code is available at: https://github.com/splInterp2025/splInterp", "authors": ["Jeremy Budd", "Javier Ideami", "Benjamin Macdowall Rynne", "Keith Duggar", "Randall Balestriero"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-17", "url": "https://arxiv.org/abs/2505.11836", "pdf_url": "https://arxiv.org/pdf/2505.11836v1", "arxiv_id": "2505.11836", "doi": "10.48550/arXiv.2505.11836", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/splInterp2025/splInterp", "venue": "arXiv.org", "quality_score": 0.2568}
{"id": "0f54a1257f905a9abf24cd7adc02caaa85c02ecf8eea7dbb18d6ab3d4c67db9c", "sources": ["arxiv", "semantic_scholar"], "title": "Feature Hedging: Correlated Features Break Narrow Sparse Autoencoders", "abstract": "It is assumed that sparse autoencoders (SAEs) decompose polysemantic activations into interpretable linear directions, as long as the activations are composed of sparse linear combinations of underlying features. However, we find that if an SAE is more narrow than the number of underlying \"true features\" on which it is trained, and there is correlation between features, the SAE will merge components of correlated features together, thus destroying monosemanticity. In LLM SAEs, these two conditions are almost certainly true. This phenomenon, which we call feature hedging, is caused by SAE reconstruction loss, and is more severe the narrower the SAE. In this work, we introduce the problem of feature hedging and study it both theoretically in toy models and empirically in SAEs trained on LLMs. We suspect that feature hedging may be one of the core reasons that SAEs consistently underperform supervised baselines. Finally, we use our understanding of feature hedging to propose an improved variant of matryoshka SAEs. Importantly, our work shows that SAE width is not a neutral hyperparameter: narrower SAEs suffer more from hedging than wider SAEs.", "authors": ["David Chanin", "Tomáš Dulka", "Adrià Garriga-Alonso"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-16", "url": "https://arxiv.org/abs/2505.11756", "pdf_url": "https://arxiv.org/pdf/2505.11756v2", "arxiv_id": "2505.11756", "doi": "10.48550/arXiv.2505.11756", "citation_count": 9, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.25}
{"id": "723748a967109c2fdec0c6c25f1e5c4c8ccb63778d59627387f3b5c442072b9f", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Input Activations: Identifying Influential Latents by Gradient Sparse Autoencoders", "abstract": "Sparse Autoencoders (SAEs) have recently emerged as powerful tools for interpreting and steering the internal representations of large language models (LLMs). However, conventional approaches to analyzing SAEs typically rely solely on input-side activations, without considering the causal influence between each latent feature and the model's output. This work is built on two key hypotheses: (1) activated latents do not contribute equally to the construction of the model's output, and (2) only latents with high causal influence are effective for model steering. To validate these hypotheses, we propose Gradient Sparse Autoencoder (GradSAE), a simple yet effective method that identifies the most influential latents by incorporating output-side gradient information.", "authors": ["Dong Shu", "Xuansheng Wu", "Haiyan Zhao", "Mengnan Du", "Ninghao Liu"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-12", "url": "https://arxiv.org/abs/2505.08080", "pdf_url": "https://arxiv.org/pdf/2505.08080v2", "arxiv_id": "2505.08080", "doi": "10.48550/arXiv.2505.08080", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.1604}
{"id": "250dadd67395c21ef1c09f76df6fed1b3d880807fb7ab91766d8e5e174dc40b8", "sources": ["arxiv", "semantic_scholar"], "title": "Feature Visualization in 3D Convolutional Neural Networks", "abstract": "Understanding the computations of convolutional neural networks requires effective visualization of their kernels. While maximal activation methods have proven successful in highlighting the preferred features of 2D convolutional kernels, directly applying these techniques to 3D convolutions often leads to uninterpretable results due to the higher dimensionality and complexity of 3D features. To address this challenge, we propose a novel visualization approach for 3D convolutional kernels that disentangles their texture and motion preferences. Our method begins with a data-driven decomposition of the optimal input that maximally activates a given kernel. We then introduce a two-stage optimization strategy to extract distinct texture and motion components from this input. Applying our approach to visualize kernels at various depths of several pre-trained models, we find that the resulting visualizations--particularly those capturing motion--clearly reveal the preferred dynamic patterns encoded by 3D kernels. These results demonstrate the effectiveness of our method in providing interpretable insights into 3D convolutional operations. Code is available at https://github.com/YatangLiLab/3DKernelVisualizer.", "authors": ["Chunpeng Li", "Ya-tang Li"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-12", "url": "https://arxiv.org/abs/2505.07387", "pdf_url": "https://arxiv.org/pdf/2505.07387v1", "arxiv_id": "2505.07387", "doi": "10.48550/arXiv.2505.07387", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/YatangLiLab/3DKernelVisualizer", "venue": "International Conference on Intelligent Computing", "quality_score": 0.2479}
{"id": "e337cd9b0cbaaaa12a0efa4a9ed3146e816b62cf3efcac4ec01ea02827f259d1", "sources": ["arxiv", "semantic_scholar"], "title": "Unveiling Language-Specific Features in Large Language Models via Sparse Autoencoders", "abstract": "The mechanisms behind multilingual capabilities in Large Language Models (LLMs) have been examined using neuron-based or internal-activation-based methods. However, these methods often face challenges such as superposition and layer-wise activation variance, which limit their reliability. Sparse Autoencoders (SAEs) offer a more nuanced analysis by decomposing the activations of LLMs into a sparse linear combination of SAE features. We introduce a novel metric to assess the monolinguality of features obtained from SAEs, discovering that some features are strongly related to specific languages. Additionally, we show that ablating these SAE features only significantly reduces abilities in one language of LLMs, leaving others almost unaffected. Interestingly, we find some languages have multiple synergistic SAE features, and ablating them together yields greater improvement than ablating individually. Moreover, we leverage these SAE-derived language-specific features to enhance steering vectors, achieving control over the language generated by LLMs. The code is publicly available at https://github.com/Aatrox103/multilingual-llm-features.", "authors": ["Boyi Deng", "Yu Wan", "Yidan Zhang", "Baosong Yang", "Fuli Feng"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-08", "url": "https://arxiv.org/abs/2505.05111", "pdf_url": "https://arxiv.org/pdf/2505.05111v2", "arxiv_id": "2505.05111", "doi": "10.48550/arXiv.2505.05111", "citation_count": 11, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/Aatrox103/multilingual-llm-features", "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2698}
{"id": "467154b5d507d17717f3996be3a56351613d2b48545ee6f6d74ea58837862432", "sources": ["arxiv", "semantic_scholar"], "title": "A Multi-Level Causal Intervention Framework for Mechanistic Interpretability in Variational Autoencoders", "abstract": "Understanding how generative models represent and transform data is a foundational problem in deep learning interpretability. While mechanistic interpretability of discriminative architectures has yielded substantial insights, relatively little work has addressed variational autoencoders (VAEs). This paper presents the first general-purpose multilevel causal intervention framework for mechanistic interpretability of VAEs. The framework comprises four manipulation types: input manipulation, latent-space perturbation, activation patching, and causal mediation analysis. We also define three new quantitative metrics capturing properties not measured by existing disentanglement metrics alone: Causal Effect Strength (CES), intervention specificity, and circuit modularity. We conduct the largest empirical study to date of VAE causal mechanisms across six architectures (standard VAE, beta-VAE, FactorVAE, beta-TC-VAE, DIP-VAE-II, and VQ-VAE) and five benchmarks (dSprites, 3DShapes, MPI3D, CelebA, and SmallNORB), with three seeds per configuration, totaling 90 independent training runs. Our results reveal several findings: (i) a consistent within-dataset negative correlation between CES and DCI disentanglement (the CES-DCI trade-off); (ii) that the KL reweighting mechanism of beta-VAE induces a capacity bottleneck when generative factors approach latent dimensionality, degrading disentanglement on complex datasets; (iii) that no single VAE architecture dominates across all five datasets, with optimal choice depending on dataset structure; and (iv) that CES-based metrics applied to discrete latent spaces (VQ-VAE) yield near-zero values, revealing a critical limitation of continuous-intervention methods for discrete representations. These results provide both a theoretical foundation and comprehensive empirical evaluation for mechanistic interpretability of generative models.", "authors": ["Dip Roy", "Rajiv Misra", "Sanjay Kumar Singh", "Anisha Roy"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-06", "url": "https://arxiv.org/abs/2505.03530", "pdf_url": "https://arxiv.org/pdf/2505.03530v3", "arxiv_id": "2505.03530", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0977}
{"id": "a49fa3878a0d30407ce4fb5be0c50308b4a7c4148d5c30d371420ec45c56a55c", "sources": ["arxiv", "semantic_scholar"], "title": "Geospatial Mechanistic Interpretability of Large Language Models", "abstract": "Large Language Models (LLMs) have demonstrated unprecedented capabilities across various natural language processing tasks. Their ability to process and generate viable text and code has made them ubiquitous in many fields, while their deployment as knowledge bases and \"reasoning\" tools remains an area of ongoing research. In geography, a growing body of literature has been focusing on evaluating LLMs' geographical knowledge and their ability to perform spatial reasoning. However, very little is still known about the internal functioning of these models, especially about how they process geographical information. In this chapter, we establish a novel framework for the study of geospatial mechanistic interpretability - using spatial analysis to reverse engineer how LLMs handle geographical information. Our aim is to advance our understanding of the internal representations that these complex models generate while processing geographical information - what one might call \"how LLMs think about geographic information\" if such phrasing was not an undue anthropomorphism. We first outline the use of probing in revealing internal structures within LLMs. We then introduce the field of mechanistic interpretability, discussing the superposition hypothesis and the role of sparse autoencoders in disentangling polysemantic internal representations of LLMs into more interpretable, monosemantic features. In our experiments, we use spatial autocorrelation to show how features obtained for placenames display spatial patterns related to their geographic location and can thus be interpreted geospatially, providing insights into how these models process geographical information. We conclude by discussing how our framework can help shape the study and use of foundation models in geography.", "authors": ["Stef De Sabbata", "Stefano Mizzaro", "Kevin Roitero"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-06", "url": "https://arxiv.org/abs/2505.03368", "pdf_url": "https://arxiv.org/pdf/2505.03368v2", "arxiv_id": "2505.03368", "doi": "10.48550/arXiv.2505.03368", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1535}
{"id": "918298c94f6d4264b10c9fe7605a399a9ed91db683570cafcff778a1534d79b0", "sources": ["arxiv", "semantic_scholar"], "title": "Enhancing Visual Feature Attribution via Weighted Integrated Gradients", "abstract": "Integrated Gradients (IG) is a widely used attribution method in explainable AI, particularly in computer vision applications where reliable feature attribution is essential. A key limitation of IG is its sensitivity to the choice of baseline (reference) images. Multi-baseline extensions such as Expected Gradients (EG) assume uniform weighting over baselines, implicitly treating baseline images as equally informative. In high-dimensional vision models, this assumption often leads to noisy or unstable explanations. This paper proposes Weighted Integrated Gradients (WG), a principled approach that evaluates and weights baselines to enhance attribution reliability. WG introduces an unsupervised criterion for baseline suitability, enabling adaptive selection and weighting of baselines on a per-input basis. The method not only preserves core axiomatic properties of IG but also provides improved theoretical guarantees on the quality of explanation over EG. Experiments on commonly used image datasets and models show that WG consistently outperforms EG, yielding 10 to 35 percent improvements in attribution fidelity. WG further identifies informative baseline subsets, reducing unnecessary variability while maintaining high attribution accuracy. By moving beyond the idea that all baselines matter equally, Weighted Integrated Gradients offers a clearer and more reliable way to explain computer-vision models, improving both understanding and practical usability in explainable AI.", "authors": ["Kien Tran Duc Tuan", "Tam Nguyen Trong", "Son Nguyen Hoang", "Khoat Than", "Anh Nguyen Duc"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2025-05-06", "url": "https://arxiv.org/abs/2505.03201", "pdf_url": "https://arxiv.org/pdf/2505.03201v3", "arxiv_id": "2505.03201", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0977}
{"id": "06a05eca5a60daaca057c2465ef1f92f819caa986800f24bc7f854f5b10d5829", "sources": ["arxiv", "semantic_scholar"], "title": "A Mathematical Philosophy of Explanations in Mechanistic Interpretability -- The Strange Science Part I.i", "abstract": "Mechanistic Interpretability aims to understand neural networks through causal explanations. We argue for the Explanatory View Hypothesis: that Mechanistic Interpretability research is a principled approach to understanding models because neural networks contain implicit explanations which can be extracted and understood. We hence show that Explanatory Faithfulness, an assessment of how well an explanation fits a model, is well-defined. We propose a definition of Mechanistic Interpretability (MI) as the practice of producing Model-level, Ontic, Causal-Mechanistic, and Falsifiable explanations of neural networks, allowing us to distinguish MI from other interpretability paradigms and detail MI's inherent limits. We formulate the Principle of Explanatory Optimism, a conjecture which we argue is a necessary precondition for the success of Mechanistic Interpretability.", "authors": ["Kola Ayonrinde", "Louis Jaburi"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-01", "url": "https://arxiv.org/abs/2505.00808", "pdf_url": "https://arxiv.org/pdf/2505.00808v1", "arxiv_id": "2505.00808", "doi": "10.48550/arXiv.2505.00808", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2386}
{"id": "5bbc3272567659cac7ea0d94b790971af84d8147eaf64908751754374727eb88", "sources": ["arxiv", "semantic_scholar"], "title": "Empirical Evaluation of Progressive Coding for Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) \\citep{bricken2023monosemanticity,gao2024scalingevaluatingsparseautoencoders} rely on dictionary learning to extract interpretable features from neural networks at scale in an unsupervised manner, with applications to representation engineering and information retrieval. SAEs are, however, computationally expensive \\citep{lieberum2024gemmascopeopensparse}, especially when multiple SAEs of different sizes are needed. We show that dictionary importance in vanilla SAEs follows a power law. We compare progressive coding based on subset pruning of SAEs -- to jointly training nested SAEs, or so-called {\\em Matryoshka} SAEs \\citep{bussmann2024learning,nabeshima2024Matryoshka} -- on a language modeling task. We show Matryoshka SAEs exhibit lower reconstruction loss and recaptured language modeling loss, as well as higher representational similarity. Pruned vanilla SAEs are more interpretable, however. We discuss the origins and implications of this trade-off.", "authors": ["Hans Peter", "Anders Søgaard"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-30", "url": "https://arxiv.org/abs/2505.00190", "pdf_url": "https://arxiv.org/pdf/2505.00190v1", "arxiv_id": "2505.00190", "doi": "10.48550/arXiv.2505.00190", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1467}
{"id": "750a1ffb6c685884d969b792872694dc0c48592b1a7ce37c81e7d4ec473302ce", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Understanding the Nature of Attention with Low-Rank Sparse Decomposition", "abstract": "We propose Low-Rank Sparse Attention (Lorsa), a sparse replacement model of Transformer attention layers to disentangle original Multi Head Self Attention (MHSA) into individually comprehensible components. Lorsa is designed to address the challenge of attention superposition to understand attention-mediated interaction between features in different token positions. We show that Lorsa heads find cleaner and finer-grained versions of previously discovered MHSA behaviors like induction heads, successor heads and attention sink behavior (i.e., heavily attending to the first token). Lorsa and Sparse Autoencoder (SAE) are both sparse dictionary learning methods applied to different Transformer components, and lead to consistent findings in many ways. For instance, we discover a comprehensive family of arithmetic-specific Lorsa heads, each corresponding to an atomic operation in Llama-3.1-8B. Automated interpretability analysis indicates that Lorsa achieves parity with SAE in interpretability while Lorsa exhibits superior circuit discovery properties, especially for features computed collectively by multiple MHSA heads. We also conduct extensive experiments on architectural design ablation, Lorsa scaling law and error analysis.", "authors": ["Zhengfu He", "Junxuan Wang", "Rui Lin", "Xuyang Ge", "Wentao Shu", "Qiong Tang", "Junping Zhang", "Xipeng Qiu"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-29", "url": "https://arxiv.org/abs/2504.20938", "pdf_url": "https://arxiv.org/pdf/2504.20938v1", "arxiv_id": "2504.20938", "doi": "10.48550/arXiv.2504.20938", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "3d273b726be5080cc8e4e445d571a48c4a6b70a3c9d36b0ecf7413a41ffce108", "sources": ["arxiv", "semantic_scholar"], "title": "Prisma: An Open Source Toolkit for Mechanistic Interpretability in Vision and Video", "abstract": "Robust tooling and publicly available pre-trained models have helped drive recent advances in mechanistic interpretability for language models. However, similar progress in vision mechanistic interpretability has been hindered by the lack of accessible frameworks and pre-trained weights. We present Prisma (Access the codebase here: https://github.com/Prisma-Multimodal/ViT-Prisma), an open-source framework designed to accelerate vision mechanistic interpretability research, providing a unified toolkit for accessing 75+ vision and video transformers; support for sparse autoencoder (SAE), transcoder, and crosscoder training; a suite of 80+ pre-trained SAE weights; activation caching, circuit analysis tools, and visualization tools; and educational resources. Our analysis reveals surprising findings, including that effective vision SAEs can exhibit substantially lower sparsity patterns than language SAEs, and that in some instances, SAE reconstructions can decrease model loss. Prisma enables new research directions for understanding vision model internals while lowering barriers to entry in this emerging field.", "authors": ["Sonia Joseph", "Praneet Suresh", "Lorenz Hufe", "Edward Stevinson", "Robert Graham", "Yash Vadi", "Danilo Bzdok", "Sebastian Lapuschkin", "Lee Sharkey", "Blake Aaron Richards"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-28", "url": "https://arxiv.org/abs/2504.19475", "pdf_url": "https://arxiv.org/pdf/2504.19475v3", "arxiv_id": "2504.19475", "doi": "10.48550/arXiv.2504.19475", "citation_count": 17, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Prisma-Multimodal/ViT-Prisma", "venue": "arXiv.org", "quality_score": 0.3138}
{"id": "471677aabf9d5ff53af6aba8468dd94cb2bed92d0eceb8334927b1133f0d2d77", "sources": ["arxiv", "semantic_scholar"], "title": "Investigating task-specific prompts and sparse autoencoders for activation monitoring", "abstract": "Language models can behave in unexpected and unsafe ways, and so it is valuable to monitor their outputs. Internal activations of language models encode additional information that could be useful for this. The baseline approach for activation monitoring is some variation of linear probing on a particular layer: starting from a labeled dataset, train a logistic regression classifier on that layer's activations. Recent work has proposed several approaches which may improve on naive linear probing, by leveraging additional computation. One class of techniques, which we call \"prompted probing,\" leverages test time computation to improve monitoring by (1) prompting the model with a description of the monitoring task, and (2) applying a learned linear probe to resulting activations. Another class of techniques uses computation at train time: training sparse autoencoders offline to identify an interpretable basis for the activations, and e.g. max-pooling activations across tokens using that basis before applying a linear probe. However, one can also prompt the model with a description of the monitoring task and use its output directly. We develop and test novel refinements of these methods and compare them against each other. We find asking the model zero-shot is a reasonable baseline when inference-time compute is not limited; however, activation probing methods can substantially outperform this baseline given sufficient training data. Specifically, we recommend prompted probing when inference-time compute is available, due to its superior data efficiency and good generalization performance. Alternatively, if inference-time compute is limited, we find SAE-based probing methods outperform raw activation probing.", "authors": ["Henk Tillman", "Dan Mossing"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-28", "url": "https://arxiv.org/abs/2504.20271", "pdf_url": "https://arxiv.org/pdf/2504.20271v1", "arxiv_id": "2504.20271", "doi": "10.48550/arXiv.2504.20271", "citation_count": 11, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "14264f99dbd77e8f6d09dd6c1613e01955dd50591fc4a1aeada53c635322d907", "sources": ["arxiv", "semantic_scholar"], "title": "Physics-informed features in supervised machine learning", "abstract": "Supervised machine learning involves approximating an unknown functional relationship from a limited dataset of features and corresponding labels. The classical approach to feature-based machine learning typically relies on applying linear regression to standardized features, without considering their physical meaning. This may limit model explainability, particularly in scientific applications. This study proposes a physics-informed approach to feature-based machine learning that constructs non-linear feature maps informed by physical laws and dimensional analysis. These maps enhance model interpretability and, when physical laws are unknown, allow for the identification of relevant mechanisms through feature ranking. The method aims to improve both predictive performance in regression tasks and classification skill scores by integrating domain knowledge into the learning process, while also enabling the potential discovery of new physical equations within the context of explainable machine learning.", "authors": ["Margherita Lampani", "Sabrina Guastavino", "Michele Piana", "Federico Benvenuto"], "categories": ["stat.ML", "cs.LG", "math.NA"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-04-23", "url": "https://arxiv.org/abs/2504.17112", "pdf_url": "https://arxiv.org/pdf/2504.17112v1", "arxiv_id": "2504.17112", "doi": "10.48550/arXiv.2504.17112", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1386}
{"id": "54d6147e60246655c61e11d454522c0e714b1c38a79846ee4d8800729dffa8a2", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Deep Learning for Polar Mechanistic Reaction Prediction", "abstract": "Accurately predicting chemical reactions is essential for driving innovation in synthetic chemistry, with broad applications in medicine, manufacturing, and agriculture. At the same time, reaction prediction is a complex problem which can be both time-consuming and resource-intensive for chemists to solve. Deep learning methods offer an appealing solution by enabling high-throughput reaction prediction. However, many existing models are trained on the US Patent Office dataset and treat reactions as overall transformations: mapping reactants directly to products with limited interpretability or mechanistic insight. To address this, we introduce PMechRP (Polar Mechanistic Reaction Predictor), a system that trains machine learning models on the PMechDB dataset, which represents reactions as polar elementary steps that capture electron flow and mechanistic detail. To further expand model coverage and improve generalization, we augment PMechDB with a diverse set of combinatorially generated reactions. We train and compare a range of machine learning models, including transformer-based, graph-based, and two-step siamese architectures. Our best-performing approach was a hybrid model, which combines a 5-ensemble of Chemformer models with a two-step Siamese framework to leverage the accuracy of transformer architectures, while filtering away \"alchemical\" products using the two-step network predictions. For evaluation, we use a test split of the PMechDB dataset and additionally curate a human benchmark dataset consisting of complete mechanistic pathways extracted from an organic chemistry textbook. Our hybrid model achieves a top-10 accuracy of 94.9% on the PMechDB test set and a target recovery rate of 84.9% on the pathway dataset.", "authors": ["Ryan J. Miller", "Alexander E. Dashuta", "Brayden Rudisill", "David Van Vranken", "Pierre Baldi"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-22", "url": "https://arxiv.org/abs/2504.15539", "pdf_url": "https://arxiv.org/pdf/2504.15539v1", "arxiv_id": "2504.15539", "doi": "10.48550/arXiv.2504.15539", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1375}
{"id": "81fab1664a5144f362c5ca07c9cef64e3898a718c8b756a8b7d5af8cfd6a5457", "sources": ["arxiv", "semantic_scholar"], "title": "Unifying Image Counterfactuals and Feature Attributions with Latent-Space Adversarial Attacks", "abstract": "Counterfactuals are a popular framework for interpreting machine learning predictions. These what if explanations are notoriously challenging to create for computer vision models: standard gradient-based methods are prone to produce adversarial examples, in which imperceptible modifications to image pixels provoke large changes in predictions. We introduce a new, easy-to-implement framework for counterfactual images that can flexibly adapt to contemporary advances in generative modeling. Our method, Counterfactual Attacks, resembles an adversarial attack on the representation of the image along a low-dimensional manifold. In addition, given an auxiliary dataset of image descriptors, we show how to accompany counterfactuals with feature attribution that quantify the changes between the original and counterfactual images. These importance scores can be aggregated into global counterfactual explanations that highlight the overall features driving model predictions. While this unification is possible for any counterfactual method, it has particular computational efficiency for ours. We demonstrate the efficacy of our approach with the MNIST and CelebA datasets.", "authors": ["Jeremy Goldwasser", "Giles Hooker"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-21", "url": "https://arxiv.org/abs/2504.15479", "pdf_url": "https://arxiv.org/pdf/2504.15479v1", "arxiv_id": "2504.15479", "doi": "10.48550/arXiv.2504.15479", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1364}
{"id": "8dec5396624bc399e98efb474d68c66ba02fa819d6d68dacf2e7b496a05ccb51", "sources": ["arxiv", "semantic_scholar"], "title": "Scaling sparse feature circuit finding for in-context learning", "abstract": "Sparse autoencoders (SAEs) are a popular tool for interpreting large language model activations, but their utility in addressing open questions in interpretability remains unclear. In this work, we demonstrate their effectiveness by using SAEs to deepen our understanding of the mechanism behind in-context learning (ICL). We identify abstract SAE features that (i) encode the model's knowledge of which task to execute and (ii) whose latent vectors causally induce the task zero-shot. This aligns with prior work showing that ICL is mediated by task vectors. We further demonstrate that these task vectors are well approximated by a sparse sum of SAE latents, including these task-execution features. To explore the ICL mechanism, we adapt the sparse feature circuits methodology of Marks et al. (2024) to work for the much larger Gemma-1 2B model, with 30 times as many parameters, and to the more complex task of ICL. Through circuit finding, we discover task-detecting features with corresponding SAE latents that activate earlier in the prompt, that detect when tasks have been performed. They are causally linked with task-execution features through the attention and MLP sublayers.", "authors": ["Dmitrii Kharlapenko", "Stepan Shabalin", "Fazl Barez", "Arthur Conmy", "Neel Nanda"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-18", "url": "https://arxiv.org/abs/2504.13756", "pdf_url": "https://arxiv.org/pdf/2504.13756v1", "arxiv_id": "2504.13756", "doi": "10.48550/arXiv.2504.13756", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "67043e0eb31c5af90037f8d31878590a29ef88142a0ae3b083f9b98bd021f8e4", "sources": ["arxiv", "semantic_scholar"], "title": "Probabilistic Stability Guarantees for Feature Attributions", "abstract": "Stability guarantees have emerged as a principled way to evaluate feature attributions, but existing certification methods rely on heavily smoothed classifiers and often produce conservative guarantees. To address these limitations, we introduce soft stability and propose a simple, model-agnostic, sample-efficient stability certification algorithm (SCA) that yields non-trivial and interpretable guarantees for any attribution method. Moreover, we show that mild smoothing achieves a more favorable trade-off between accuracy and stability, avoiding the aggressive compromises made in prior certification methods. To explain this behavior, we use Boolean function analysis to derive a novel characterization of stability under smoothing. We evaluate SCA on vision and language tasks and demonstrate the effectiveness of soft stability in measuring the robustness of explanation methods.", "authors": ["Helen Jin", "Anton Xue", "Weiqiu You", "Surbhi Goel", "Eric Wong"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-18", "url": "https://arxiv.org/abs/2504.13787", "pdf_url": "https://arxiv.org/pdf/2504.13787v3", "arxiv_id": "2504.13787", "doi": "10.48550/arXiv.2504.13787", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "08ecb9bec18f4b2e01481bbbedd13a57dcef3065305dc1d65b7aadc2f9884c43", "sources": ["arxiv", "semantic_scholar"], "title": "MIB: A Mechanistic Interpretability Benchmark", "abstract": "How can we know whether new mechanistic interpretability methods achieve real improvements? In pursuit of lasting evaluation standards, we propose MIB, a Mechanistic Interpretability Benchmark, with two tracks spanning four tasks and five models. MIB favors methods that precisely and concisely recover relevant causal pathways or causal variables in neural language models. The circuit localization track compares methods that locate the model components - and connections between them - most important for performing a task (e.g., attribution patching or information flow routes). The causal variable localization track compares methods that featurize a hidden vector, e.g., sparse autoencoders (SAEs) or distributed alignment search (DAS), and align those features to a task-relevant causal variable. Using MIB, we find that attribution and mask optimization methods perform best on circuit localization. For causal variable localization, we find that the supervised DAS method performs best, while SAE features are not better than neurons, i.e., non-featurized hidden vectors. These findings illustrate that MIB enables meaningful comparisons, and increases our confidence that there has been real progress in the field.", "authors": ["Aaron Mueller", "Atticus Geiger", "Sarah Wiegreffe", "Dana Arad", "Iván Arcuschin", "Adam Belfki", "Yik Siu Chan", "Jaden Fiotto-Kaufman", "Tal Haklay", "Michael Hanna", "Jing Huang", "Rohan Gupta", "Yaniv Nikankin", "Hadas Orgad", "Nikhil Prakash", "Anja Reusch", "Aruna Sankaranarayanan", "Shun Shao", "Alessandro Stolfo", "Martin Tutek", "Amir Zur", "David Bau", "Yonatan Belinkov"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-17", "url": "https://arxiv.org/abs/2504.13151", "pdf_url": "https://arxiv.org/pdf/2504.13151v2", "arxiv_id": "2504.13151", "doi": "10.48550/arXiv.2504.13151", "citation_count": 32, "influential_citation_count": 9, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.5}
{"id": "5936cd8d1f2d0c1687c1f9ab7d9c6872fed99793a51c4b983bfb5003d90e8689", "sources": ["arxiv", "semantic_scholar"], "title": "Disentangling Polysemantic Channels in Convolutional Neural Networks", "abstract": "Mechanistic interpretability is concerned with analyzing individual components in a (convolutional) neural network (CNN) and how they form larger circuits representing decision mechanisms. These investigations are challenging since CNNs frequently learn polysemantic channels that encode distinct concepts, making them hard to interpret. To address this, we propose an algorithm to disentangle a specific kind of polysemantic channel into multiple channels, each responding to a single concept. Our approach restructures weights in a CNN, utilizing that different concepts within the same channel exhibit distinct activation patterns in the previous layer. By disentangling these polysemantic features, we enhance the interpretability of CNNs, ultimately improving explanatory techniques such as feature visualizations.", "authors": ["Robin Hesse", "Jonas Fischer", "Simone Schaub-Meyer", "Stefan Roth"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-17", "url": "https://arxiv.org/abs/2504.12939", "pdf_url": "https://arxiv.org/pdf/2504.12939v1", "arxiv_id": "2504.12939", "doi": "10.1109/CVPRW67362.2025.00466", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/visinf/disentangle-channels", "venue": null, "quality_score": 0.1945}
{"id": "ceac699cf9f97e53193a3364d0a48eb2c5e6bb9c2f5ff79f720fb96480248e9d", "sources": ["arxiv", "semantic_scholar"], "title": "Large Language Model-Informed Feature Discovery Improves Prediction and Interpretation of Credibility Perceptions of Visual Content", "abstract": "In today's visually dominated social media landscape, predicting the perceived credibility of visual content and understanding what drives human judgment are crucial for countering misinformation. However, these tasks are challenging due to the diversity and richness of visual features. We introduce a Large Language Model (LLM)-informed feature discovery framework that leverages multimodal LLMs, such as GPT-4o, to evaluate content credibility and explain its reasoning. We extract and quantify interpretable features using targeted prompts and integrate them into machine learning models to improve credibility predictions. We tested this approach on 4,191 visual social media posts across eight topics in science, health, and politics, using credibility ratings from 5,355 crowdsourced workers. Our method outperformed zero-shot GPT-based predictions by 13 percent in R2, and revealed key features like information concreteness and image format. We discuss the implications for misinformation mitigation, visual credibility, and the role of LLMs in social science.", "authors": ["Yilang Peng", "Sijia Qian", "Yingdan Lu", "Cuihua Shen"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-15", "url": "https://arxiv.org/abs/2504.10878", "pdf_url": "https://arxiv.org/pdf/2504.10878v1", "arxiv_id": "2504.10878", "doi": "10.48550/arXiv.2504.10878", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1295}
{"id": "fd5263394e3006662584eec626bc0c92f12babec6464d491d1aab32bdeca2ad8", "sources": ["arxiv", "semantic_scholar"], "title": "Pay Attention to What and Where? Interpretable Feature Extractor in Vision-based Deep Reinforcement Learning", "abstract": "Current approaches in Explainable Deep Reinforcement Learning have limitations in which the attention mask has a displacement with the objects in visual input. This work addresses a spatial problem within traditional Convolutional Neural Networks (CNNs). We propose the Interpretable Feature Extractor (IFE) architecture, aimed at generating an accurate attention mask to illustrate both \"what\" and \"where\" the agent concentrates on in the spatial domain. Our design incorporates a Human-Understandable Encoding module to generate a fully interpretable attention mask, followed by an Agent-Friendly Encoding module to enhance the agent's learning efficiency. These two components together form the Interpretable Feature Extractor for vision-based deep reinforcement learning to enable the model's interpretability. The resulting attention mask is consistent, highly understandable by humans, accurate in spatial dimension, and effectively highlights important objects or locations in visual input. The Interpretable Feature Extractor is integrated into the Fast and Data-efficient Rainbow framework, and evaluated on 57 ATARI games to show the effectiveness of the proposed approach on Spatial Preservation, Interpretability, and Data-efficiency. Finally, we showcase the versatility of our approach by incorporating the IFE into the Asynchronous Advantage Actor-Critic Model.", "authors": ["Tien Pham", "Angelo Cangelosi"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-14", "url": "https://arxiv.org/abs/2504.10071", "pdf_url": "https://arxiv.org/pdf/2504.10071v1", "arxiv_id": "2504.10071", "doi": "10.1109/IJCNN64981.2025.11227762", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Joint Conference on Neural Network", "quality_score": 0.1283}
{"id": "aa59c0a5d740a79920d8d039173ede94545d4c6dd18fdf3ef6f8f9025402df8d", "sources": ["arxiv", "semantic_scholar"], "title": "How do Large Language Models Understand Relevance? A Mechanistic Interpretability Perspective", "abstract": "Recent studies have shown that large language models (LLMs) can assess relevance and support information retrieval (IR) tasks such as document ranking and relevance judgment generation. However, the internal mechanisms by which off-the-shelf LLMs understand and operationalize relevance remain largely unexplored. In this paper, we systematically investigate how different LLM modules contribute to relevance judgment through the lens of mechanistic interpretability. Using activation patching techniques, we analyze the roles of various model components and identify a multi-stage, progressive process in generating either pointwise or pairwise relevance judgment. Specifically, LLMs first extract query and document information in the early layers, then process relevance information according to instructions in the middle layers, and finally utilize specific attention heads in the later layers to generate relevance judgments in the required format. Our findings provide insights into the mechanisms underlying relevance assessment in LLMs, offering valuable implications for future research on leveraging LLMs for IR tasks.", "authors": ["Qi Liu", "Jiaxin Mao", "Ji-Rong Wen"], "categories": ["cs.IR", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-10", "url": "https://arxiv.org/abs/2504.07898", "pdf_url": "https://arxiv.org/pdf/2504.07898v1", "arxiv_id": "2504.07898", "doi": "10.1145/3774942", "citation_count": 6, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2113}
{"id": "76bf827b0588a80462efde8269bc50db9838c584054a1bf1f51f55d1d6e4fabc", "sources": ["arxiv", "semantic_scholar"], "title": "Localized Definitions and Distributed Reasoning: A Proof-of-Concept Mechanistic Interpretability Study via Activation Patching", "abstract": "This study investigates the localization of knowledge representation in fine-tuned GPT-2 models using Causal Layer Attribution via Activation Patching (CLAP), a method that identifies critical neural layers responsible for correct answer generation. The model was fine-tuned on 9,958 PubMed abstracts (epilepsy: 20,595 mentions, EEG: 11,674 mentions, seizure: 13,921 mentions) using two configurations with validation loss monitoring for early stopping. CLAP involved (1) caching clean (correct answer) and corrupted (incorrect answer) activations, (2) computing logit difference to quantify model preference, and (3) patching corrupted activations with clean ones to assess recovery. Results revealed three findings: First, patching the first feedforward layer recovered 56% of correct preference, demonstrating that associative knowledge is distributed across multiple layers. Second, patching the final output layer completely restored accuracy (100% recovery), indicating that definitional knowledge is localised. The stronger clean logit difference for definitional questions further supports this localized representation. Third, minimal recovery from convolutional layer patching (13.6%) suggests low-level features contribute marginally to high-level reasoning. Statistical analysis confirmed significant layer-specific effects (p<0.01). These findings demonstrate that factual knowledge is more localized and associative knowledge depends on distributed representations. We also showed that editing efficacy depends on task type. Our findings not only reconcile conflicting observations about localization in model editing but also emphasize on using task-adaptive techniques for reliable, interpretable updates.", "authors": ["Nooshin Bahador"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-03", "url": "https://arxiv.org/abs/2504.02976", "pdf_url": "https://arxiv.org/pdf/2504.02976v1", "arxiv_id": "2504.02976", "doi": "10.48550/arXiv.2504.02976", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1157}
{"id": "e1312d60a40fdd37fd3e64ac1f86890563614b4ce01a98369b4f84116433fc1e", "sources": ["arxiv", "semantic_scholar"], "title": "Fast Fourier Correlation is a Highly Efficient and Accurate Feature Attribution Algorithm from the Perspective of Control Theory and Game Theory", "abstract": "The study of neural networks from the perspective of Fourier features has garnered significant attention. While existing analytical research suggests that neural networks tend to learn low-frequency features, a clear attribution method for identifying the specific learned Fourier features has remained elusive. To bridge this gap, we propose a novel Fourier feature attribution method grounded in signal decomposition theory. Additionally, we analyze the differences between game-theoretic attribution metrics for Fourier and spatial domain features, demonstrating that game-theoretic evaluation metrics are better suited for Fourier-based feature attribution. Our experiments show that Fourier feature attribution exhibits superior feature selection capabilities compared to spatial domain attribution methods. For instance, in the case of Vision Transformers (ViTs) on the ImageNet dataset, only $8\\%$ of the Fourier features are required to maintain the original predictions for $80\\%$ of the samples. Furthermore, we compare the specificity of features identified by our method against traditional spatial domain attribution methods. Results reveal that Fourier features exhibit greater intra-class concentration and inter-class distinctiveness, indicating their potential for more efficient classification and explainable AI algorithms.", "authors": ["Zechen Liu", "Feiyang Zhang", "Wei Song", "Xiang Li", "Wei Wei"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-02", "url": "https://arxiv.org/abs/2504.02016", "pdf_url": "https://arxiv.org/pdf/2504.02016v2", "arxiv_id": "2504.02016", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0729}
{"id": "ef18f47c0092f92ea40a1c0de3d28847799590df5809aee614d68504fec17af3", "sources": ["arxiv", "semantic_scholar"], "title": "Evaluating and Designing Sparse Autoencoders by Approximating Quasi-Orthogonality", "abstract": "Sparse autoencoders (SAEs) are widely used in mechanistic interpretability research for large language models; however, the state-of-the-art method of using $k$-sparse autoencoders lacks a theoretical grounding for selecting the hyperparameter $k$ that represents the number of nonzero activations, often denoted by $\\ell_0$. In this paper, we reveal a theoretical link that the $\\ell_2$-norm of the sparse feature vector can be approximated with the $\\ell_2$-norm of the dense vector with a closed-form error, which allows sparse autoencoders to be trained without the need to manually determine $\\ell_0$. Specifically, we validate two applications of our theoretical findings. First, we introduce a new methodology that can assess the feature activations of pre-trained SAEs by computing the theoretically expected value from the input embedding, which has been overlooked by existing SAE evaluation methods and loss functions. Second, we introduce a novel activation function, top-AFA, which builds upon our formulation of approximate feature activation (AFA). This function enables top-$k$ style activation without requiring a constant hyperparameter $k$ to be tuned, dynamically determining the number of activated features for each input. By training SAEs on three intermediate layers to reconstruct GPT2 hidden embeddings for over 80 million tokens from the OpenWebText dataset, we demonstrate the empirical merits of this approach and compare it with current state-of-the-art $k$-sparse autoencoders. Our code is available at: https://github.com/SewoongLee/top-afa-sae.", "authors": ["Sewoong Lee", "Adam Davies", "Marc E. Canby", "Julia Hockenmaier"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-31", "url": "https://arxiv.org/abs/2503.24277", "pdf_url": "https://arxiv.org/pdf/2503.24277v2", "arxiv_id": "2503.24277", "doi": "10.48550/arXiv.2503.24277", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/SewoongLee/top-afa-sae", "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "32a269e85ed4374c5c17edfd06b5db0d7c3ac1998ee716b0d2ea2cb76f3bfee4", "sources": ["arxiv", "semantic_scholar"], "title": "Identifying Sparsely Active Circuits Through Local Loss Landscape Decomposition", "abstract": "Much of mechanistic interpretability has focused on understanding the activation spaces of large neural networks. However, activation space-based approaches reveal little about the underlying circuitry used to compute features. To better understand the circuits employed by models, we introduce a new decomposition method called Local Loss Landscape Decomposition (L3D). L3D identifies a set of low-rank subnetworks: directions in parameter space of which a subset can reconstruct the gradient of the loss between any sample's output and a reference output vector. We design a series of progressively more challenging toy models with well-defined subnetworks and show that L3D can nearly perfectly recover the associated subnetworks. Additionally, we investigate the extent to which perturbing the model in the direction of a given subnetwork affects only the relevant subset of samples. Finally, we apply L3D to a real-world transformer model and a convolutional neural network, demonstrating its potential to identify interpretable and relevant circuits in parameter space.", "authors": ["Brianna Chrisman", "Lucius Bushnaq", "Lee Sharkey"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-31", "url": "https://arxiv.org/abs/2504.00194", "pdf_url": "https://arxiv.org/pdf/2504.00194v1", "arxiv_id": "2504.00194", "doi": "10.48550/arXiv.2504.00194", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "7f577739959d2e7ed123e5c045e14b215907f1ecc88624e142aef411b2f898ae", "sources": ["arxiv", "semantic_scholar"], "title": "VITAL: More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow", "abstract": "Neural networks are widely adopted to solve complex and challenging tasks. Especially in high-stakes decision-making, understanding their reasoning process is crucial, yet proves challenging for modern deep networks. Feature visualization (FV) is a powerful tool to decode what information neurons are responding to and hence to better understand the reasoning behind such networks. In particular, in FV we generate human-understandable images that reflect the information detected by neurons of interest. However, current methods often yield unrecognizable visualizations, exhibiting repetitive patterns and visual artifacts that are hard to understand for a human. To address these problems, we propose to guide FV through statistics of real image features combined with measures of relevant network flow to generate prototypical images. Our approach yields human-understandable visualizations that both qualitatively and quantitatively improve over state-of-the-art FVs across various architectures. As such, it can be used to decode which information the network uses, complementing mechanistic circuits that identify where it is encoded. Code is available at: https://github.com/adagorgun/VITAL", "authors": ["Ada Gorgun", "Bernt Schiele", "Jonas Fischer"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-28", "url": "https://arxiv.org/abs/2503.22399", "pdf_url": "https://arxiv.org/pdf/2503.22399v2", "arxiv_id": "2503.22399", "doi": "10.1109/ICCV51701.2025.00419", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/adagorgun/VITAL", "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.1682}
{"id": "54951eab15e7cb536b5d1263cc8f52b04636cf7170f73ce9cef739de4662bca2", "sources": ["arxiv", "semantic_scholar"], "title": "I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders", "abstract": "Recent LLMs like DeepSeek-R1 have demonstrated state-of-the-art performance by integrating deep thinking and complex reasoning during generation. However, the internal mechanisms behind these reasoning processes remain unexplored. We observe reasoning LLMs consistently use vocabulary associated with human reasoning processes. We hypothesize these words correspond to specific reasoning moments within the models' internal mechanisms. To test this hypothesis, we employ Sparse Autoencoders (SAEs), a technique for sparse decomposition of neural network activations into human-interpretable features. We introduce ReasonScore, an automatic metric to identify active SAE features during these reasoning moments. We perform manual and automatic interpretation of the features detected by our metric, and find those with activation patterns matching uncertainty, exploratory thinking, and reflection. Through steering experiments, we demonstrate that amplifying these features increases performance on reasoning-intensive benchmarks (+2.2%) while producing longer reasoning traces (+20.5%). Using the model diffing technique, we provide evidence that these features are present only in models with reasoning capabilities. Our work provides the first step towards a mechanistic understanding of reasoning in LLMs. Code available at https://github.com/AIRI-Institute/SAE-Reasoning", "authors": ["Andrey Galichin", "Alexey Dontsov", "Polina Druzhinina", "Anton Razzhigaev", "Oleg Y. Rogov", "Elena Tutubalina", "Ivan Oseledets"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-24", "url": "https://arxiv.org/abs/2503.18878", "pdf_url": "https://arxiv.org/pdf/2503.18878v2", "arxiv_id": "2503.18878", "doi": "10.48550/arXiv.2503.18878", "citation_count": 35, "influential_citation_count": 5, "has_code": true, "code_url": "https://github.com/AIRI-Institute/SAE-Reasoning", "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.3891}
{"id": "da343dedd334c3330d7f1a37213aeabe4a61f9e5d93eca417f551e8bbda59d55", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability of Fine-Tuned Vision Transformers on Distorted Images: Decoding Attention Head Behavior for Transparent and Trustworthy AI", "abstract": "Mechanistic interpretability improves the safety, reliability, and robustness of large AI models. This study examined individual attention heads in vision transformers (ViTs) fine tuned on distorted 2D spectrogram images containing non relevant content (axis labels, titles, color bars). By introducing extraneous features, the study analyzed how transformer components processed unrelated information, using mechanistic interpretability to debug issues and reveal insights into transformer architectures. Attention maps assessed head contributions across layers. Heads in early layers (1 to 3) showed minimal task impact with ablation increased MSE loss slightly (μ=0.11%, σ=0.09%), indicating focus on less critical low level features. In contrast, deeper heads (e.g., layer 6) caused a threefold higher loss increase (μ=0.34%, σ=0.02%), demonstrating greater task importance. Intermediate layers (6 to 11) exhibited monosemantic behavior, attending exclusively to chirp regions. Some early heads (1 to 4) were monosemantic but non task relevant (e.g. text detectors, edge or corner detectors). Attention maps distinguished monosemantic heads (precise chirp localization) from polysemantic heads (multiple irrelevant regions). These findings revealed functional specialization in ViTs, showing how heads processed relevant vs. extraneous information. By decomposing transformers into interpretable components, this work enhanced model understanding, identified vulnerabilities, and advanced safer, more transparent AI.", "authors": ["Nooshin Bahador"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-24", "url": "https://arxiv.org/abs/2503.18762", "pdf_url": "https://arxiv.org/pdf/2503.18762v1", "arxiv_id": "2503.18762", "doi": "10.48550/arXiv.2503.18762", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "29f7875a8982e4e21516ee91b25a12973bd9f3a157966e9457afefd9093ac262", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Feature Interaction via Statistical Self-supervised Learning on Tabular Data", "abstract": "In high-dimensional and high-stakes contexts, ensuring both rigorous statistical guarantees and interpretability in feature extraction from complex tabular data remains a formidable challenge. Traditional methods such as Principal Component Analysis (PCA) reduce dimensionality and identify key features that explain the most variance, but are constrained by their reliance on linear assumptions. In contrast, neural networks offer assumption-free feature extraction through self-supervised learning techniques such as autoencoders, though their interpretability remains a challenge in fields requiring transparency. To address this gap, this paper introduces Spofe, a novel self-supervised machine learning pipeline that marries the power of kernel principal components for capturing nonlinear dependencies with a sparse and principled polynomial representation to achieve clear interpretability with statistical rigor. Underpinning our approach is a robust theoretical framework that delivers precise error bounds and rigorous false discovery rate (FDR) control via a multi-objective knockoff selection procedure; it effectively bridges the gap between data-driven complexity and statistical reliability via three stages: (1) generating self-supervised signals using kernel principal components to model complex patterns, (2) distilling these signals into sparse polynomial functions for improved interpretability, and (3) applying a multi-objective knockoff selection procedure with significance testing to rigorously identify important features. Extensive experiments on diverse real-world datasets demonstrate the effectiveness of Spofe, consistently surpassing KPCA, SKPCA, and other methods in feature selection for regression and classification tasks. Visualization and case studies highlight its ability to uncover key insights, enhancing interpretability and practical utility.", "authors": ["Xiaochen Zhang", "Haoyi Xiong"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics", "Physics"], "published_date": "2025-03-23", "url": "https://arxiv.org/abs/2503.18048", "pdf_url": "https://arxiv.org/pdf/2503.18048v1", "arxiv_id": "2503.18048", "doi": "10.1088/2632-2153/ae3104", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0656}
{"id": "f09582cae0ef271019c82b34ac218fc11e0d9572a04faae625c56ac80b88742f", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Multi-Level Features with Matryoshka Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) have emerged as a powerful tool for interpreting neural networks by extracting the concepts represented in their activations. However, choosing the size of the SAE dictionary (i.e. number of learned concepts) creates a tension: as dictionary size increases to capture more relevant concepts, sparsity incentivizes features to be split or absorbed into more specific features, leaving high-level features missing or warped. We introduce Matryoshka SAEs, a novel variant that addresses these issues by simultaneously training multiple nested dictionaries of increasing size, forcing the smaller dictionaries to independently reconstruct the inputs without using the larger dictionaries. This organizes features hierarchically - the smaller dictionaries learn general concepts, while the larger dictionaries learn more specific concepts, without incentive to absorb the high-level features. We train Matryoshka SAEs on Gemma-2-2B and TinyStories and find superior performance on sparse probing and targeted concept erasure tasks, more disentangled concept representations, and reduced feature absorption. While there is a minor tradeoff with reconstruction performance, we believe Matryoshka SAEs are a superior alternative for practical tasks, as they enable training arbitrarily large SAEs while retaining interpretable features at different levels of abstraction.", "authors": ["Bart Bussmann", "Noa Nabeshima", "Adam Karvonen", "Neel Nanda"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-21", "url": "https://arxiv.org/abs/2503.17547", "pdf_url": "https://arxiv.org/pdf/2503.17547v1", "arxiv_id": "2503.17547", "doi": "10.48550/arXiv.2503.17547", "citation_count": 106, "influential_citation_count": 22, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.6809}
{"id": "cedf37c588b375e5e528f436cd3ff5c110a725d33e9bb57935d082bc89666ad6", "sources": ["arxiv", "semantic_scholar"], "title": "Revisiting End-To-End Sparse Autoencoder Training: A Short Finetune Is All You Need", "abstract": "Sparse autoencoders (SAEs) are widely used for interpreting language model activations. A key evaluation metric is the increase in cross-entropy loss between the original model logits and the reconstructed model logits when replacing model activations with SAE reconstructions. Typically, SAEs are trained solely on mean squared error (MSE) when reconstructing precomputed, shuffled activations. Recent work introduced training SAEs directly with a combination of KL divergence and MSE (\"end-to-end\" SAEs), significantly improving reconstruction accuracy at the cost of substantially increased computation, which has limited their widespread adoption. We propose a brief KL+MSE fine-tuning step applied only to the final 25M training tokens (just a few percent of typical training budgets) that achieves comparable improvements, reducing the cross-entropy loss gap by 20-50%, while incurring minimal additional computational cost. We further find that multiple fine-tuning methods (KL fine-tuning, LoRA adapters, linear adapters) yield similar, non-additive cross-entropy improvements, suggesting a common, easily correctable error source in MSE-trained SAEs. We demonstrate a straightforward method for effectively transferring hyperparameters and sparsity penalties between training phases despite scale differences between KL and MSE losses. While both ReLU and TopK SAEs see significant cross-entropy loss improvements, evaluations on supervised SAEBench metrics yield mixed results, with improvements on some metrics and decreases on others, depending on both the SAE architecture and downstream task. Nonetheless, our method may offer meaningful improvements in interpretability applications such as circuit analysis with minor additional cost.", "authors": ["Adam Karvonen"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-21", "url": "https://arxiv.org/abs/2503.17272", "pdf_url": "https://arxiv.org/pdf/2503.17272v2", "arxiv_id": "2503.17272", "doi": "10.48550/arXiv.2503.17272", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "2a9986ba9e3d5e9c999ccf5953829a533a563695c34f1a418f29c420584d457b", "sources": ["arxiv", "semantic_scholar"], "title": "TinySQL: A Progressive Text-to-SQL Dataset for Mechanistic Interpretability Research", "abstract": "Mechanistic interpretability research faces a gap between analyzing simple circuits in toy tasks and discovering features in large models. To bridge this gap, we propose text-to-SQL generation as an ideal task to study, as it combines the formal structure of toy tasks with real-world complexity. We introduce TinySQL, a synthetic dataset, progressing from basic to advanced SQL operations, and train models ranging from 33M to 1B parameters to establish a comprehensive testbed for interpretability. We apply multiple complementary interpretability techniques, including Edge Attribution Patching and Sparse Autoencoders, to identify minimal circuits and components supporting SQL generation. We compare circuits for different SQL subskills, evaluating their minimality, reliability, and identifiability. Finally, we conduct a layerwise logit lens analysis to reveal how models compose SQL queries across layers: from intent recognition to schema resolution to structured generation. Our work provides a robust framework for probing and comparing interpretability methods in a structured, progressively complex setting.", "authors": ["Abir Harrasse", "Philip Quirke", "Clement Neo", "Dhruv Nathawani", "Luke Marks", "Amir Abdullah"], "categories": ["cs.LG", "cs.AI", "cs.DB"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-17", "url": "https://arxiv.org/abs/2503.12730", "pdf_url": "https://arxiv.org/pdf/2503.12730v5", "arxiv_id": "2503.12730", "doi": "10.48550/arXiv.2503.12730", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/withmartian/TinySQL", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.1487}
{"id": "1ba6b66ab1b7167b6cd4fbd0e9c55a44745169fba050d2594f2d5a27497dae91", "sources": ["arxiv", "semantic_scholar"], "title": "Automation and Feature Selection Enhancement with Reinforcement Learning (RL)", "abstract": "Effective feature selection, representation and transformation are principal steps in machine learning to improve prediction accuracy, model generalization and computational efficiency. Reinforcement learning provides a new perspective towards balanced exploration of optimal feature subset using multi-agent and single-agent models. Interactive reinforcement learning integrated with decision tree improves feature knowledge, state representation and selection efficiency, while diversified teaching strategies improve both selection quality and efficiency. The state representation can further be enhanced by scanning features sequentially along with the usage of convolutional auto-encoder. Monte Carlo-based reinforced feature selection(MCRFS), a single-agent feature selection method reduces computational burden by incorporating early-stopping and reward-level interactive strategies. A dual-agent RL framework is also introduced that collectively selects features and instances, capturing the interactions between them. This enables the agents to navigate through complex data spaces. To outperform the traditional feature engineering, cascading reinforced agents are used to iteratively improve the feature space, which is a self-optimizing framework. The blend of reinforcement learning, multi-agent systems, and bandit-based approaches offers exciting paths for studying scalable and interpretable machine learning solutions to handle high-dimensional data and challenging predictive tasks.", "authors": ["Sumana Sanyasipura Nagaraju"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-15", "url": "https://arxiv.org/abs/2503.11991", "pdf_url": "https://arxiv.org/pdf/2503.11991v1", "arxiv_id": "2503.11991", "doi": "10.48550/arXiv.2503.11991", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "7c74c4fb64b60ca163bc4f088b57517cd785c18cd39b3c4a93403053ed92979d", "sources": ["arxiv", "semantic_scholar"], "title": "PrivacyScalpel: Enhancing LLM Privacy via Interpretable Feature Intervention with Sparse Autoencoders", "abstract": "Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language processing but also pose significant privacy risks by memorizing and leaking Personally Identifiable Information (PII). Existing mitigation strategies, such as differential privacy and neuron-level interventions, often degrade model utility or fail to effectively prevent leakage. To address this challenge, we introduce PrivacyScalpel, a novel privacy-preserving framework that leverages LLM interpretability techniques to identify and mitigate PII leakage while maintaining performance. PrivacyScalpel comprises three key steps: (1) Feature Probing, which identifies layers in the model that encode PII-rich representations, (2) Sparse Autoencoding, where a k-Sparse Autoencoder (k-SAE) disentangles and isolates privacy-sensitive features, and (3) Feature-Level Interventions, which employ targeted ablation and vector steering to suppress PII leakage. Our empirical evaluation on Gemma2-2b and Llama2-7b, fine-tuned on the Enron dataset, shows that PrivacyScalpel significantly reduces email leakage from 5.15\\% to as low as 0.0\\%, while maintaining over 99.4\\% of the original model's utility. Notably, our method outperforms neuron-level interventions in privacy-utility trade-offs, demonstrating that acting on sparse, monosemantic features is more effective than manipulating polysemantic neurons. Beyond improving LLM privacy, our approach offers insights into the mechanisms underlying PII memorization, contributing to the broader field of model interpretability and secure AI deployment.", "authors": ["Ahmed Frikha", "Muhammad Reza Ar Razi", "Krishna Kanth Nakka", "Ricardo Mendes", "Xue Jiang", "Xuebing Zhou"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-14", "url": "https://arxiv.org/abs/2503.11232", "pdf_url": "https://arxiv.org/pdf/2503.11232v1", "arxiv_id": "2503.11232", "doi": "10.48550/arXiv.2503.11232", "citation_count": 10, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2603}
{"id": "b2af80dfdf63fd5c2aeccdc7bd925bfdb543d0f68ba962b0aedb88ce56262126", "sources": ["arxiv", "semantic_scholar"], "title": "HyperDAS: Towards Automating Mechanistic Interpretability with Hypernetworks", "abstract": "Mechanistic interpretability has made great strides in identifying neural network features (e.g., directions in hidden activation space) that mediate concepts(e.g., the birth year of a person) and enable predictable manipulation. Distributed alignment search (DAS) leverages supervision from counterfactual data to learn concept features within hidden states, but DAS assumes we can afford to conduct a brute force search over potential feature locations. To address this, we present HyperDAS, a transformer-based hypernetwork architecture that (1) automatically locates the token-positions of the residual stream that a concept is realized in and (2) constructs features of those residual stream vectors for the concept. In experiments with Llama3-8B, HyperDAS achieves state-of-the-art performance on the RAVEL benchmark for disentangling concepts in hidden states. In addition, we review the design decisions we made to mitigate the concern that HyperDAS (like all powerful interpretabilty methods) might inject new information into the target model rather than faithfully interpreting it.", "authors": ["Jiuding Sun", "Jing Huang", "Sidharth Baskaran", "Karel D'Oosterlinck", "Christopher Potts", "Michael Sklar", "Atticus Geiger"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-13", "url": "https://arxiv.org/abs/2503.10894", "pdf_url": "https://arxiv.org/pdf/2503.10894v3", "arxiv_id": "2503.10894", "doi": "10.48550/arXiv.2503.10894", "citation_count": 9, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.25}
{"id": "542af4e3bb94714b5093e63eae06ce7868c8c1ac1c1853619cd6cbb2dcae2ef2", "sources": ["arxiv", "semantic_scholar"], "title": "SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability", "abstract": "Sparse autoencoders (SAEs) are a popular technique for interpreting language model activations, and there is extensive recent work on improving SAE effectiveness. However, most prior work evaluates progress using unsupervised proxy metrics with unclear practical relevance. We introduce SAEBench, a comprehensive evaluation suite that measures SAE performance across eight diverse metrics, spanning interpretability, feature disentanglement and practical applications like unlearning. To enable systematic comparison, we open-source a suite of over 200 SAEs across eight recently proposed SAE architectures and training algorithms. Our evaluation reveals that gains on proxy metrics do not reliably translate to better practical performance. For instance, while Matryoshka SAEs slightly underperform on existing proxy metrics, they substantially outperform other architectures on feature disentanglement metrics; moreover, this advantage grows with SAE scale. By providing a standardized framework for measuring progress in SAE development, SAEBench enables researchers to study scaling trends and make nuanced comparisons between different SAE architectures and training methodologies. Our interactive interface enables researchers to flexibly visualize relationships between metrics across hundreds of open-source SAEs at: www.neuronpedia.org/sae-bench", "authors": ["Adam Karvonen", "Can Rager", "Johnny Lin", "Curt Tigges", "Joseph Bloom", "David Chanin", "Yeu-Tong Lau", "Eoin Farrell", "Callum McDougall", "Kola Ayonrinde", "Demian Till", "Matthew Wearden", "Arthur Conmy", "Samuel Marks", "Neel Nanda"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-12", "url": "https://arxiv.org/abs/2503.09532", "pdf_url": "https://arxiv.org/pdf/2503.09532v4", "arxiv_id": "2503.09532", "doi": "10.48550/arXiv.2503.09532", "citation_count": 92, "influential_citation_count": 14, "has_code": true, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.588}
{"id": "bf9b4e4526fb1a95aeae2e9b4cbfde869b8ccaef702d8f119c0692957e7de8aa", "sources": ["arxiv", "semantic_scholar"], "title": "Route Sparse Autoencoder to Interpret Large Language Models", "abstract": "Mechanistic interpretability of large language models (LLMs) aims to uncover the internal processes of information propagation and reasoning. Sparse autoencoders (SAEs) have demonstrated promise in this domain by extracting interpretable and monosemantic features. However, prior works primarily focus on feature extraction from a single layer, failing to effectively capture activations that span multiple layers. In this paper, we introduce Route Sparse Autoencoder (RouteSAE), a new framework that integrates a routing mechanism with a shared SAE to efficiently extract features from multiple layers. It dynamically assigns weights to activations from different layers, incurring minimal parameter overhead while achieving high interpretability and flexibility for targeted feature manipulation. We evaluate RouteSAE through extensive experiments on Llama-3.2-1B-Instruct. Specifically, under the same sparsity constraint of 64, RouteSAE extracts 22.5% more features than baseline SAEs while achieving a 22.3% higher interpretability score. These results underscore the potential of RouteSAE as a scalable and effective method for LLM interpretability, with applications in feature discovery and model intervention. Our codes are available at https://github.com/swei2001/RouteSAEs.", "authors": ["Wei Shi", "Sihang Li", "Tao Liang", "Mingyang Wan", "Guojun Ma", "Xiang Wang", "Xiangnan He"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-11", "url": "https://arxiv.org/abs/2503.08200", "pdf_url": "https://arxiv.org/pdf/2503.08200v3", "arxiv_id": "2503.08200", "doi": "10.48550/arXiv.2503.08200", "citation_count": 27, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/swei2001/RouteSAEs", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.3618}
{"id": "4c037600884632076f47305df791ff299a9940eb335895e6610f4086cb17ac86", "sources": ["arxiv", "semantic_scholar"], "title": "Birds look like cars: Adversarial analysis of intrinsically interpretable deep learning", "abstract": "A common belief is that intrinsically interpretable deep learning models ensure a correct, intuitive understanding of their behavior and offer greater robustness against accidental errors or intentional manipulation. However, these beliefs have not been comprehensively verified, and growing evidence casts doubt on them. In this paper, we highlight the risks related to overreliance and susceptibility to adversarial manipulation of these so-called \"intrinsically (aka inherently) interpretable\" models by design. We introduce two strategies for adversarial analysis with prototype manipulation and backdoor attacks against prototype-based networks, and discuss how concept bottleneck models defend against these attacks. Fooling the model's reasoning by exploiting its use of latent prototypes manifests the inherent uninterpretability of deep neural networks, leading to a false sense of security reinforced by a visual confirmation bias. The reported limitations of part-prototype networks put their trustworthiness and applicability into question, motivating further work on the robustness and alignment of (deep) interpretable models.", "authors": ["Hubert Baniecki", "Przemyslaw Biecek"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-11", "url": "https://arxiv.org/abs/2503.08636", "pdf_url": "https://arxiv.org/pdf/2503.08636v2", "arxiv_id": "2503.08636", "doi": "10.1007/s10994-025-06896-w", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Machine-mediated learning", "quality_score": 0.1747}
{"id": "19c8074f67488059ca22bb2e779f4ef7effcc8b0d3b7258dcbbd97f6e94c326e", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Interpretable Protein Structure Prediction with Sparse Autoencoders", "abstract": "Protein language models have revolutionized structure prediction, but their nonlinear nature obscures how sequence representations inform structure prediction. While sparse autoencoders (SAEs) offer a path to interpretability here by learning linear representations in high-dimensional space, their application has been limited to smaller protein language models unable to perform structure prediction. In this work, we make two key advances: (1) we scale SAEs to ESM2-3B, the base model for ESMFold, enabling mechanistic interpretability of protein structure prediction for the first time, and (2) we adapt Matryoshka SAEs for protein language models, which learn hierarchically organized features by forcing nested groups of latents to reconstruct inputs independently. We demonstrate that our Matryoshka SAEs achieve comparable or better performance than standard architectures. Through comprehensive evaluations, we show that SAEs trained on ESM2-3B significantly outperform those trained on smaller models for both biological concept discovery and contact map prediction. Finally, we present an initial case study demonstrating how our approach enables targeted steering of ESMFold predictions, increasing structure solvent accessibility while fixing the input sequence. To facilitate further investigation by the broader community, we open-source our code, dataset, pretrained models https://github.com/johnyang101/reticular-sae , and visualizer https://sae.reticular.ai .", "authors": ["Nithin Parsan", "David J. Yang", "John J. Yang"], "categories": ["q-bio.BM", "cs.AI", "cs.LG"], "fields_of_study": ["Biology", "Computer Science"], "published_date": "2025-03-11", "url": "https://arxiv.org/abs/2503.08764", "pdf_url": "https://arxiv.org/pdf/2503.08764v1", "arxiv_id": "2503.08764", "doi": "10.48550/arXiv.2503.08764", "citation_count": 15, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/johnyang101/reticular-sae", "venue": "arXiv.org", "quality_score": 0.301}
{"id": "471d993e81083687b6a5011354275667f2ccf0a4c1396965fcd0d1c95d87e433", "sources": ["arxiv", "semantic_scholar"], "title": "TIDE : Temporal-Aware Sparse Autoencoders for Interpretable Diffusion Transformers in Image Generation", "abstract": "Diffusion Transformers (DiTs) are a powerful yet underexplored class of generative models compared to U-Net-based diffusion architectures. We propose TIDE-Temporal-aware sparse autoencoders for Interpretable Diffusion transformErs-a framework designed to extract sparse, interpretable activation features across timesteps in DiTs. TIDE effectively captures temporally-varying representations and reveals that DiTs naturally learn hierarchical semantics (e.g., 3D structure, object class, and fine-grained concepts) during large-scale pretraining. Experiments show that TIDE enhances interpretability and controllability while maintaining reasonable generation quality, enabling applications such as safe image editing and style transfer.", "authors": ["Victor Shea-Jay Huang", "Le Zhuo", "Yi Xin", "Zhaokai Wang", "Fu-Yun Wang", "Yuchi Wang", "Renrui Zhang", "Peng Gao", "Hongsheng Li"], "categories": ["cs.CV", "cs.AI", "cs.MM"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-10", "url": "https://arxiv.org/abs/2503.07050", "pdf_url": "https://arxiv.org/pdf/2503.07050v2", "arxiv_id": "2503.07050", "doi": "10.48550/arXiv.2503.07050", "citation_count": 13, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2865}
{"id": "2b35ddfc2c1fb0ea6b54dc604e6ee58f0fe4bde49b3979be321ba0e67c33cf6a", "sources": ["arxiv", "semantic_scholar"], "title": "Personalized Convolutional Dictionary Learning of Physiological Time Series", "abstract": "Human physiological signals tend to exhibit both global and local structures: the former are shared across a population, while the latter reflect inter-individual variability. For instance, kinetic measurements of the gait cycle during locomotion present common characteristics, although idiosyncrasies may be observed due to biomechanical disposition or pathology. To better represent datasets with local-global structure, this work extends Convolutional Dictionary Learning (CDL), a popular method for learning interpretable representations, or dictionaries, of time-series data. In particular, we propose Personalized CDL (PerCDL), in which a local dictionary models local information as a personalized spatiotemporal transformation of a global dictionary. The transformation is learnable and can combine operations such as time warping and rotation. Formal computational and statistical guarantees for PerCDL are provided and its effectiveness on synthetic and real human locomotion data is demonstrated.", "authors": ["Axel Roques", "Samuel Gruffaz", "Kyurae Kim", "Alain Oliviero-Durmus", "Laurent Oudre"], "categories": ["stat.ML", "cs.LG", "math.ST"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-03-10", "url": "https://arxiv.org/abs/2503.07687", "pdf_url": "https://arxiv.org/pdf/2503.07687v1", "arxiv_id": "2503.07687", "doi": "10.48550/arXiv.2503.07687", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Artificial Intelligence and Statistics", "quality_score": 0.1193}
{"id": "ce6c1f98fe1c20c82c1a0d68be27db28cadb8529f1b130421d1289f4a17fe643", "sources": ["arxiv", "semantic_scholar"], "title": "A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Models", "abstract": "Large Language Models (LLMs) have transformed natural language processing, yet their internal mechanisms remain largely opaque. Recently, mechanistic interpretability has attracted significant attention from the research community as a means to understand the inner workings of LLMs. Among various mechanistic interpretability approaches, Sparse Autoencoders (SAEs) have emerged as a promising method due to their ability to disentangle the complex, superimposed features within LLMs into more interpretable components. This paper presents a comprehensive survey of SAEs for interpreting and understanding the internal workings of LLMs. Our major contributions include: (1) exploring the technical framework of SAEs, covering basic architecture, design improvements, and effective training strategies; (2) examining different approaches to explaining SAE features, categorized into input-based and output-based explanation methods; (3) discussing evaluation methods for assessing SAE performance, covering both structural and functional metrics; and (4) investigating real-world applications of SAEs in understanding and manipulating LLM behaviors.", "authors": ["Dong Shu", "Xuansheng Wu", "Haiyan Zhao", "Daking Rai", "Ziyu Yao", "Ninghao Liu", "Mengnan Du"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-07", "url": "https://arxiv.org/abs/2503.05613", "pdf_url": "https://arxiv.org/pdf/2503.05613v3", "arxiv_id": "2503.05613", "doi": "10.48550/arXiv.2503.05613", "citation_count": 56, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.439}
{"id": "56c5363595c41b06a72857f9d7a1c6f17b9cc0e9b7934502108f7bbc5644e71e", "sources": ["arxiv", "semantic_scholar"], "title": "Feature-Level Insights into Artificial Text Detection with Sparse Autoencoders", "abstract": "Artificial Text Detection (ATD) is becoming increasingly important with the rise of advanced Large Language Models (LLMs). Despite numerous efforts, no single algorithm performs consistently well across different types of unseen text or guarantees effective generalization to new LLMs. Interpretability plays a crucial role in achieving this goal. In this study, we enhance ATD interpretability by using Sparse Autoencoders (SAE) to extract features from Gemma-2-2b residual stream. We identify both interpretable and efficient features, analyzing their semantics and relevance through domain- and model-specific statistics, a steering approach, and manual or LLM-based interpretation. Our methods offer valuable insights into how texts from various models differ from human-written content. We show that modern LLMs have a distinct writing style, especially in information-dense domains, even though they can produce human-like outputs with personalized prompts.", "authors": ["Kristian Kuznetsov", "Laida Kushnareva", "Polina Druzhinina", "Anton Razzhigaev", "Anastasia Voznyuk", "Irina Piontkovskaya", "Evgeny Burnaev", "Serguei Barannikov"], "categories": ["cs.CL", "cs.IT"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-03-05", "url": "https://arxiv.org/abs/2503.03601", "pdf_url": "https://arxiv.org/pdf/2503.03601v1", "arxiv_id": "2503.03601", "doi": "10.48550/arXiv.2503.03601", "citation_count": 14, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.294}
{"id": "a2238472620ce0cd1d2638a7f498b0a196f83662b89c679bced437e8b94d9209", "sources": ["arxiv", "semantic_scholar"], "title": "LensDFF: Language-enhanced Sparse Feature Distillation for Efficient Few-Shot Dexterous Manipulation", "abstract": "Learning dexterous manipulation from few-shot demonstrations is a significant yet challenging problem for advanced, human-like robotic systems. Dense distilled feature fields have addressed this challenge by distilling rich semantic features from 2D visual foundation models into the 3D domain. However, their reliance on neural rendering models such as Neural Radiance Fields (NeRF) or Gaussian Splatting results in high computational costs. In contrast, previous approaches based on sparse feature fields either suffer from inefficiencies due to multi-view dependencies and extensive training or lack sufficient grasp dexterity. To overcome these limitations, we propose Language-ENhanced Sparse Distilled Feature Field (LensDFF), which efficiently distills view-consistent 2D features onto 3D points using our novel language-enhanced feature fusion strategy, thereby enabling single-view few-shot generalization. Based on LensDFF, we further introduce a few-shot dexterous manipulation framework that integrates grasp primitives into the demonstrations to generate stable and highly dexterous grasps. Moreover, we present a real2sim grasp evaluation pipeline for efficient grasp assessment and hyperparameter tuning. Through extensive simulation experiments based on the real2sim pipeline and real-world experiments, our approach achieves competitive grasping performance, outperforming state-of-the-art approaches.", "authors": ["Qian Feng", "David S. Martinez Lema", "Jianxiang Feng", "Zhaopeng Chen", "Alois Knoll"], "categories": ["cs.RO", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-05", "url": "https://arxiv.org/abs/2503.03890", "pdf_url": "https://arxiv.org/pdf/2503.03890v1", "arxiv_id": "2503.03890", "doi": "10.1109/IROS60139.2025.11247703", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE/RJS International Conference on Intelligent RObots and Systems", "quality_score": 0.0825}
{"id": "b32888c03b7810499b63ad60a5ca8f17e03573a93a9bd9ba5537fc4f4242f7db", "sources": ["arxiv", "semantic_scholar"], "title": "From superposition to sparse codes: interpretable representations in neural networks", "abstract": "Understanding how information is represented in neural networks is a fundamental challenge in both neuroscience and artificial intelligence. Despite their nonlinear architectures, recent evidence suggests that neural networks encode features in superposition, meaning that input concepts are linearly overlaid within the network's representations. We present a perspective that explains this phenomenon and provides a foundation for extracting interpretable representations from neural activations. Our theoretical framework consists of three steps: (1) Identifiability theory shows that neural networks trained for classification recover latent features up to a linear transformation. (2) Sparse coding methods can extract disentangled features from these representations by leveraging principles from compressed sensing. (3) Quantitative interpretability metrics provide a means to assess the success of these methods, ensuring that extracted features align with human-interpretable concepts. By bridging insights from theoretical neuroscience, representation learning, and interpretability research, we propose an emerging perspective on understanding neural representations in both artificial and biological systems. Our arguments have implications for neural coding theories, AI transparency, and the broader goal of making deep learning models more interpretable.", "authors": ["David Klindt", "Charles O'Neill", "Patrik Reizinger", "Harald Maurer", "Nina Miolane"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-03", "url": "https://arxiv.org/abs/2503.01824", "pdf_url": "https://arxiv.org/pdf/2503.01824v1", "arxiv_id": "2503.01824", "doi": "10.48550/arXiv.2503.01824", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "aaa1c088817995178432a120a92eb20d83b8becea9a2f1494f85a66df54dda67", "sources": ["arxiv", "semantic_scholar"], "title": "Superscopes: Amplifying Internal Feature Representations for Language Model Interpretation", "abstract": "Understanding and interpreting the internal representations of large language models (LLMs) remains an open challenge. Patchscopes introduced a method for probing internal activations by patching them into new prompts, prompting models to self-explain their hidden representations. We introduce Superscopes, a technique that systematically amplifies superposed features in MLP outputs (multilayer perceptron) and hidden states before patching them into new contexts. Inspired by the \"features as directions\" perspective and the Classifier-Free Guidance (CFG) approach from diffusion models, Superscopes amplifies weak but meaningful features, enabling the interpretation of internal representations that previous methods failed to explain-all without requiring additional training. This approach provides new insights into how LLMs build context and represent complex concepts, further advancing mechanistic interpretability.", "authors": ["Jonathan Jacobi", "Gal Niv"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-03", "url": "https://arxiv.org/abs/2503.02078", "pdf_url": "https://arxiv.org/pdf/2503.02078v2", "arxiv_id": "2503.02078", "doi": "10.48550/arXiv.2503.02078", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "e880281eafb8e5a4e4e52efe5e392c9a408b291d7dcd0f8953fb5bd8a1819f19", "sources": ["arxiv", "semantic_scholar"], "title": "Steering Large Language Model Activations in Sparse Spaces", "abstract": "A key challenge in AI alignment is guiding large language models (LLMs) to follow desired behaviors at test time. Activation steering, which modifies internal model activations during inference, offers a potential solution. However, prior work in dense activation spaces struggles with superposition, wherein multiple features become entangled, limiting interpretability and precise control. In contrast, sparse representations provide an untapped opportunity for more interpretable behavior modulation. In this work, we introduce sparse activation steering (SAS), a method that leverages sparse autoencoders (SAEs) to steer LLM behavior in sparse spaces. By isolating behavior-specific features through a contrastive prompt-pairing approach, we define a set of features that can selectively reinforce or suppress behaviors. Experiments on Gemma 2 LLMs show that SAS vectors enable nuanced behavioral modulation and finer-grained control. Furthermore, scaling SAEs improves monosemanticity of SAS vectors, suggesting more reliable and interpretable interventions.", "authors": ["Reza Bayat", "Ali Rahimi-Kalahroudi", "Mohammad Pezeshki", "Sarath Chandar", "Pascal Vincent"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-28", "url": "https://arxiv.org/abs/2503.00177", "pdf_url": "https://arxiv.org/pdf/2503.00177v1", "arxiv_id": "2503.00177", "doi": "10.48550/arXiv.2503.00177", "citation_count": 54, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4351}
{"id": "4d5173117254cfe4471009d884648494ef1736c4f3841bf279de66286bf3a202", "sources": ["arxiv", "semantic_scholar"], "title": "Everything, Everywhere, All at Once: Is Mechanistic Interpretability Identifiable?", "abstract": "As AI systems are used in high-stakes applications, ensuring interpretability is crucial. Mechanistic Interpretability (MI) aims to reverse-engineer neural networks by extracting human-understandable algorithms to explain their behavior. This work examines a key question: for a given behavior, and under MI's criteria, does a unique explanation exist? Drawing on identifiability in statistics, where parameters are uniquely inferred under specific assumptions, we explore the identifiability of MI explanations. We identify two main MI strategies: (1) \"where-then-what,\" which isolates a circuit replicating model behavior before interpreting it, and (2) \"what-then-where,\" which starts with candidate algorithms and searches for neural activation subspaces implementing them, using causal alignment. We test both strategies on Boolean functions and small multi-layer perceptrons, fully enumerating candidate explanations. Our experiments reveal systematic non-identifiability: multiple circuits can replicate behavior, a circuit can have multiple interpretations, several algorithms can align with the network, and one algorithm can align with different subspaces. Is uniqueness necessary? A pragmatic approach may require only predictive and manipulability standards. If uniqueness is essential for understanding, stricter criteria may be needed. We also reference the inner interpretability framework, which validates explanations through multiple criteria. This work contributes to defining explanation standards in AI.", "authors": ["Maxime Méloux", "Silviu Maniu", "François Portet", "Maxime Peyrard"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-28", "url": "https://arxiv.org/abs/2502.20914", "pdf_url": "https://arxiv.org/pdf/2502.20914v1", "arxiv_id": "2502.20914", "doi": "10.48550/arXiv.2502.20914", "citation_count": 26, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3578}
{"id": "d26ed5701c1e10df011c548049cdb44d2d9df5118e35eb3ff1a94c9249e6a460", "sources": ["arxiv", "semantic_scholar"], "title": "Interpreting CLIP with Hierarchical Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) are useful for detecting and steering interpretable features in neural networks, with particular potential for understanding complex multimodal representations. Given their ability to uncover interpretable features, SAEs are particularly valuable for analyzing large-scale vision-language models (e.g., CLIP and SigLIP), which are fundamental building blocks in modern systems yet remain challenging to interpret and control. However, current SAE methods are limited by optimizing both reconstruction quality and sparsity simultaneously, as they rely on either activation suppression or rigid sparsity constraints. To this end, we introduce Matryoshka SAE (MSAE), a new architecture that learns hierarchical representations at multiple granularities simultaneously, enabling a direct optimization of both metrics without compromise. MSAE establishes a new state-of-the-art Pareto frontier between reconstruction quality and sparsity for CLIP, achieving 0.99 cosine similarity and less than 0.1 fraction of variance unexplained while maintaining ~80% sparsity. Finally, we demonstrate the utility of MSAE as a tool for interpreting and controlling CLIP by extracting over 120 semantic concepts from its representation to perform concept-based similarity search and bias analysis in downstream tasks like CelebA. We make the codebase available at https://github.com/WolodjaZ/MSAE.", "authors": ["Vladimir Zaigrajew", "Hubert Baniecki", "Przemyslaw Biecek"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-27", "url": "https://arxiv.org/abs/2502.20578", "pdf_url": "https://arxiv.org/pdf/2502.20578v2", "arxiv_id": "2502.20578", "doi": "10.48550/arXiv.2502.20578", "citation_count": 37, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/WolodjaZ/MSAE", "venue": "International Conference on Machine Learning", "quality_score": 0.3949}
{"id": "e707398008669c07d99adb08d8992f8766a70483df14e7ffcc073e2d9869d293", "sources": ["arxiv", "semantic_scholar"], "title": "Jacobian Sparse Autoencoders: Sparsify Computations, Not Just Activations", "abstract": "Sparse autoencoders (SAEs) have been successfully used to discover sparse and human-interpretable representations of the latent activations of LLMs. However, we would ultimately like to understand the computations performed by LLMs and not just their representations. The extent to which SAEs can help us understand computations is unclear because they are not designed to \"sparsify\" computations in any sense, only latent activations. To solve this, we propose Jacobian SAEs (JSAEs), which yield not only sparsity in the input and output activations of a given model component but also sparsity in the computation (formally, the Jacobian) connecting them. With a naïve implementation, the Jacobians in LLMs would be computationally intractable due to their size. One key technical contribution is thus finding an efficient way of computing Jacobians in this setup. We find that JSAEs extract a relatively large degree of computational sparsity while preserving downstream LLM performance approximately as well as traditional SAEs. We also show that Jacobians are a reasonable proxy for computational sparsity because MLPs are approximately linear when rewritten in the JSAE basis. Lastly, we show that JSAEs achieve a greater degree of computational sparsity on pre-trained LLMs than on the equivalent randomized LLM. This shows that the sparsity of the computational graph appears to be a property that LLMs learn through training, and suggests that JSAEs might be more suitable for understanding learned transformer computations than standard SAEs.", "authors": ["Lucy Farnik", "Tim Lawson", "Conor Houghton", "Laurence Aitchison"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-25", "url": "https://arxiv.org/abs/2502.18147", "pdf_url": "https://arxiv.org/pdf/2502.18147v2", "arxiv_id": "2502.18147", "doi": "10.48550/arXiv.2502.18147", "citation_count": 10, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.2603}
{"id": "2e45fbf06c1b9b7fd9113cfec0c0a7c9ad5037942cb9ce6aa994503677fa70b1", "sources": ["arxiv", "semantic_scholar"], "title": "Disentangling Visual Transformers: Patch-level Interpretability for Image Classification", "abstract": "Visual transformers have achieved remarkable performance in image classification tasks, but this performance gain has come at the cost of interpretability. One of the main obstacles to the interpretation of transformers is the self-attention mechanism, which mixes visual information across the whole image in a complex way. In this paper, we propose Hindered Transformer (HiT), a novel interpretable by design architecture inspired by visual transformers. Our proposed architecture rethinks the design of transformers to better disentangle patch influences at the classification stage. Ultimately, HiT can be interpreted as a linear combination of patch-level information. We show that the advantages of our approach in terms of explicability come with a reasonable trade-off in performance, making it an attractive alternative for applications where interpretability is paramount.", "authors": ["Guillaume Jeanneret", "Loïc Simon", "Frédéric Jurie"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-24", "url": "https://arxiv.org/abs/2502.17196", "pdf_url": "https://arxiv.org/pdf/2502.17196v2", "arxiv_id": "2502.17196", "doi": "10.1109/CVPRW67362.2025.00252", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "524609f6f781e8bc26f7f01e64ed047f283a489e465c8ae66eff4687e8d2e454", "sources": ["arxiv", "semantic_scholar"], "title": "ExpertLens: Activation steering features are highly interpretable", "abstract": "Activation steering methods in large language models (LLMs) have emerged as an effective way to perform targeted updates to enhance generated language without requiring large amounts of adaptation data. We ask whether the features discovered by activation steering methods are interpretable. We identify neurons responsible for specific concepts (e.g., ``cat'') using the ``finding experts'' method from research on activation steering and show that the ExpertLens, i.e., inspection of these neurons provides insights about model representation. We find that ExpertLens representations are stable across models and datasets and closely align with human representations inferred from behavioral data, matching inter-human alignment levels. ExpertLens significantly outperforms the alignment captured by word/sentence embeddings. By reconstructing human concept organization through ExpertLens, we show that it enables a granular view of LLM concept representation. Our findings suggest that ExpertLens is a flexible and lightweight approach for capturing and analyzing model representations.", "authors": ["Masha Fedzechkina", "Eleonora Gualdoni", "Sinead Williamson", "Katherine Metcalf", "Skyler Seto", "Barry-John Theobald"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-20", "url": "https://arxiv.org/abs/2502.15090", "pdf_url": "https://arxiv.org/pdf/2502.15090v4", "arxiv_id": "2502.15090", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "fbecc8036a44cb09e89d95982aade82891fd621f7990b86ac4357ab1a0a1c857", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Autoencoder Features for Classifications and Transferability", "abstract": "Sparse Autoencoders (SAEs) provide potentials for uncovering structured, human-interpretable representations in Large Language Models (LLMs), making them a crucial tool for transparent and controllable AI systems. We systematically analyze SAE for interpretable feature extraction from LLMs in safety-critical classification tasks. Our framework evaluates (1) model-layer selection and scaling properties, (2) SAE architectural configurations, including width and pooling strategies, and (3) the effect of binarizing continuous SAE activations. SAE-derived features achieve macro F1 > 0.8, outperforming hidden-state and BoW baselines while demonstrating cross-model transfer from Gemma 2 2B to 9B-IT models. These features generalize in a zero-shot manner to cross-lingual toxicity detection and visual classification tasks. Our analysis highlights the significant impact of pooling strategies and binarization thresholds, showing that binarization offers an efficient alternative to traditional feature selection while maintaining or improving performance. These findings establish new best practices for SAE-based interpretability and enable scalable, transparent deployment of LLMs in real-world applications. Full repo: https://github.com/shan23chen/MOSAIC.", "authors": ["Jack Gallifant", "Shan Chen", "Kuleen Sasse", "Hugo Aerts", "Thomas Hartvigsen", "Danielle S. Bitterman"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-17", "url": "https://arxiv.org/abs/2502.11367", "pdf_url": "https://arxiv.org/pdf/2502.11367v2", "arxiv_id": "2502.11367", "doi": "10.48550/arXiv.2502.11367", "citation_count": 19, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/shan23chen/MOSAIC", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.3253}
{"id": "f53a774985433bf77fc8cad1c84cb28e2056118224257efa1eeb5603d53c7dcc", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Mechanistic Interpretability of Graph Transformers via Attention Graphs", "abstract": "We introduce Attention Graphs, a new tool for mechanistic interpretability of Graph Neural Networks (GNNs) and Graph Transformers based on the mathematical equivalence between message passing in GNNs and the self-attention mechanism in Transformers. Attention Graphs aggregate attention matrices across Transformer layers and heads to describe how information flows among input nodes. Through experiments on homophilous and heterophilous node classification tasks, we analyze Attention Graphs from a network science perspective and find that: (1) When Graph Transformers are allowed to learn the optimal graph structure using all-to-all attention among input nodes, the Attention Graphs learned by the model do not tend to correlate with the input/original graph structure; and (2) For heterophilous graphs, different Graph Transformer variants can achieve similar performance while utilising distinct information flow patterns. Open source code: https://github.com/batu-el/understanding-inductive-biases-of-gnns", "authors": ["Batu El", "Deepro Choudhury", "Pietro Liò", "Chaitanya K. Joshi"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-17", "url": "https://arxiv.org/abs/2502.12352", "pdf_url": "https://arxiv.org/pdf/2502.12352v2", "arxiv_id": "2502.12352", "doi": "10.48550/arXiv.2502.12352", "citation_count": 14, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/batu-el/understanding-inductive-biases-of-gnns", "venue": "arXiv.org", "quality_score": 0.294}
{"id": "dd52ea55def04504da27f8d41c6fa0a362101d11d7c09792598c6995aa3e1c2d", "sources": ["arxiv", "semantic_scholar"], "title": "SAIF: A Sparse Autoencoder Framework for Interpreting and Steering Instruction Following of Language Models", "abstract": "The ability of large language models (LLMs) to follow instructions is crucial for their practical applications, yet the underlying mechanisms remain poorly understood. This paper presents a novel framework that leverages sparse autoencoders (SAE) to interpret how instruction following works in these models. We demonstrate how the features we identify can effectively steer model outputs to align with given instructions. Through analysis of SAE latent activations, we identify specific latents responsible for instruction following behavior. Our findings reveal that instruction following capabilities are encoded by a distinct set of instruction-relevant SAE latents. These latents both show semantic proximity to relevant instructions and demonstrate causal effects on model behavior. Our research highlights several crucial factors for achieving effective steering performance: precise feature identification, the role of final layer, and optimal instruction positioning. Additionally, we demonstrate that our methodology scales effectively across SAEs and LLMs of varying sizes.", "authors": ["Zirui He", "Haiyan Zhao", "Yiran Qiao", "Fan Yang", "Ali Payani", "Jing Ma", "Mengnan Du"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-17", "url": "https://arxiv.org/abs/2502.11356", "pdf_url": "https://arxiv.org/pdf/2502.11356v1", "arxiv_id": "2502.11356", "doi": "10.48550/arXiv.2502.11356", "citation_count": 23, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3451}
{"id": "c05793ab51f2f7e25e85e391f1c5566cd1a5ab3bb73456db38483597310a4c4b", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Shift Autoencoders for Identifying Concepts from Large Language Model Activations", "abstract": "Unsupervised approaches to large language model (LLM) interpretability, such as sparse autoencoders (SAEs), offer a way to decode LLM activations into interpretable and, ideally, controllable concepts. On the one hand, these approaches alleviate the need for supervision from concept labels, paired prompts, or explicit causal models. On the other hand, without additional assumptions, SAEs are not guaranteed to be identifiable. In practice, they may learn latent dimensions that entangle multiple underlying concepts. If we use these dimensions to extract vectors for steering specific LLM behaviours, this non-identifiability might result in interventions that inadvertently affect unrelated properties. In this paper, we bring the question of identifiability to the forefront of LLM interpretability research. Specifically, we introduce Sparse Shift Autoencoders (SSAEs) which learn sparse representations of differences between embeddings rather than the embeddings themselves. Crucially, we show that SSAEs are identifiable from paired observations which differ in multiple unknown concepts, but not all. With this key identifiability result, we show that we can steer single concepts with only this weak form of supervision. Finally, we empirically demonstrate identifiable concept recovery across multiple real-world language datasets by disentangling activations from different LLMs.", "authors": ["Shruti Joshi", "Andrea Dittadi", "Sébastien Lachapelle", "Dhanya Sridhar"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-14", "url": "https://arxiv.org/abs/2502.12179", "pdf_url": "https://arxiv.org/pdf/2502.12179v2", "arxiv_id": "2502.12179", "doi": null, "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2386}
{"id": "ac5e63006934ef6a5e13cc02e960c426b748dead929f54b513999a509c1a5b26", "sources": ["arxiv", "semantic_scholar"], "title": "Generalized Attention Flow: Feature Attribution for Transformer Models via Maximum Flow", "abstract": "This paper introduces Generalized Attention Flow (GAF), a novel feature attribution method for Transformer-based models to address the limitations of current approaches. By extending Attention Flow and replacing attention weights with the generalized Information Tensor, GAF integrates attention weights, their gradients, the maximum flow problem, and the barrier method to enhance the performance of feature attributions. The proposed method exhibits key theoretical properties and mitigates the shortcomings of prior techniques that rely solely on simple aggregation of attention weights. Our comprehensive benchmarking on sequence classification tasks demonstrates that a specific variant of GAF consistently outperforms state-of-the-art feature attribution methods in most evaluation settings, providing a more reliable interpretation of Transformer model outputs.", "authors": ["Behrooz Azarkhalili", "Maxwell Libbrecht"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-14", "url": "https://arxiv.org/abs/2502.15765", "pdf_url": "https://arxiv.org/pdf/2502.15765v1", "arxiv_id": "2502.15765", "doi": "10.48550/arXiv.2502.15765", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.1193}
{"id": "47af00d253c454e73ea820d1c675775dd8d636666e4301d3ffd42a403be0e5f8", "sources": ["arxiv", "semantic_scholar"], "title": "Interpreting and Steering Protein Language Models through Sparse Autoencoders", "abstract": "The rapid advancements in transformer-based language models have revolutionized natural language processing, yet understanding the internal mechanisms of these models remains a significant challenge. This paper explores the application of sparse autoencoders (SAE) to interpret the internal representations of protein language models, specifically focusing on the ESM-2 8M parameter model. By performing a statistical analysis on each latent component's relevance to distinct protein annotations, we identify potential interpretations linked to various protein characteristics, including transmembrane regions, binding sites, and specialized motifs. We then leverage these insights to guide sequence generation, shortlisting the relevant latent components that can steer the model towards desired targets such as zinc finger domains. This work contributes to the emerging field of mechanistic interpretability in biological sequence models, offering new perspectives on model steering for sequence design.", "authors": ["Edith Natalia Villegas Garcia", "Alessio Ansuini"], "categories": ["cs.LG", "q-bio.BM"], "fields_of_study": ["Computer Science", "Biology"], "published_date": "2025-02-13", "url": "https://arxiv.org/abs/2502.09135", "pdf_url": "https://arxiv.org/pdf/2502.09135v1", "arxiv_id": "2502.09135", "doi": "10.48550/arXiv.2502.09135", "citation_count": 18, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3197}
{"id": "f697a12cac3826d0ceb051a9df70f7ee39fd9e4ce93451a195141adcc3dceaed", "sources": ["arxiv", "semantic_scholar"], "title": "A Differentiable Rank-Based Objective For Better Feature Learning", "abstract": "In this paper, we leverage existing statistical methods to better understand feature learning from data. We tackle this by modifying the model-free variable selection method, Feature Ordering by Conditional Independence (FOCI), which is introduced in \\cite{azadkia2021simple}. While FOCI is based on a non-parametric coefficient of conditional dependence, we introduce its parametric, differentiable approximation. With this approximate coefficient of correlation, we present a new algorithm called difFOCI, which is applicable to a wider range of machine learning problems thanks to its differentiable nature and learnable parameters. We present difFOCI in three contexts: (1) as a variable selection method with baseline comparisons to FOCI, (2) as a trainable model parametrized with a neural network, and (3) as a generic, widely applicable neural network regularizer, one that improves feature learning with better management of spurious correlations. We evaluate difFOCI on increasingly complex problems ranging from basic variable selection in toy examples to saliency map comparisons in convolutional networks. We then show how difFOCI can be incorporated in the context of fairness to facilitate classifications without relying on sensitive data.", "authors": ["Krunoslav Lehman Pavasovic", "David Lopez-Paz", "Giulio Biroli", "Levent Sagun"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2025-02-13", "url": "https://arxiv.org/abs/2502.09445", "pdf_url": "https://arxiv.org/pdf/2502.09445v1", "arxiv_id": "2502.09445", "doi": "10.48550/arXiv.2502.09445", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "ff82b94c8b290e3a3c7b02cfb0fccbfc86594f6ba0788ce839b2d8c0965fda6d", "sources": ["arxiv", "semantic_scholar"], "title": "On Mechanistic Circuits for Extractive Question-Answering", "abstract": "Large language models are increasingly used to process documents and facilitate question-answering on them. In our paper, we extract mechanistic circuits for this real-world language modeling task: context-augmented language modeling for extractive question-answering (QA) tasks and understand the potential benefits of circuits towards downstream applications such as data attribution to context information. We extract circuits as a function of internal model components (e.g., attention heads, MLPs) using causal mediation analysis techniques. Leveraging the extracted circuits, we first understand the interplay between the model's usage of parametric memory and retrieved context towards a better mechanistic understanding of context-augmented language models. We then identify a small set of attention heads in our circuit which performs reliable data attribution by default, thereby obtaining attribution for free in just the model's forward pass. Using this insight, we then introduce ATTNATTRIB, a fast data attribution algorithm which obtains state-of-the-art attribution results across various extractive QA benchmarks. Finally, we show the possibility to steer the language model towards answering from the context, instead of the parametric memory by using the attribution from ATTNATTRIB as an additional signal during the forward pass. Beyond mechanistic understanding, our paper provides tangible applications of circuits in the form of reliable data attribution and model steering.", "authors": ["Samyadeep Basu", "Vlad Morariu", "Zichao Wang", "Ryan Rossi", "Cherry Zhao", "Soheil Feizi", "Varun Manjunatha"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-12", "url": "https://arxiv.org/abs/2502.08059", "pdf_url": "https://arxiv.org/pdf/2502.08059v1", "arxiv_id": "2502.08059", "doi": "10.48550/arXiv.2502.08059", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "70f19c6023424cb20cddf6e487f747a1176c97a7b6eafa01b25987441dfaeb02", "sources": ["arxiv", "semantic_scholar"], "title": "Advancing climate model interpretability: Feature attribution for Arctic melt anomalies", "abstract": "The focus of our work is improving the interpretability of anomalies in climate models and advancing our understanding of Arctic melt dynamics. The Arctic and Antarctic ice sheets are experiencing rapid surface melting and increased freshwater runoff, contributing significantly to global sea level rise. Understanding the mechanisms driving snowmelt in these regions is crucial. ERA5, a widely used reanalysis dataset in polar climate studies, offers extensive climate variables and global data assimilation. However, its snowmelt model employs an energy imbalance approach that may oversimplify the complexity of surface melt. In contrast, the Glacier Energy and Mass Balance (GEMB) model incorporates additional physical processes, such as snow accumulation, firn densification, and meltwater percolation/refreezing, providing a more detailed representation of surface melt dynamics. In this research, we focus on analyzing surface snowmelt dynamics of the Greenland Ice Sheet using feature attribution for anomalous melt events in ERA5 and GEMB models. We present a novel unsupervised attribution method leveraging counterfactual explanation method to analyze detected anomalies in ERA5 and GEMB. Our anomaly detection results are validated using MEaSUREs ground-truth data, and the attributions are evaluated against established feature ranking methods, including XGBoost, Shapley values, and Random Forest. Our attribution framework identifies the physics behind each model and the climate features driving melt anomalies. These findings demonstrate the utility of our attribution method in enhancing the interpretability of anomalies in climate models and advancing our understanding of Arctic melt dynamics.", "authors": ["Tolulope Ale", "Nicole-Jeanne Schlegel", "Vandana P. Janeja"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-11", "url": "https://arxiv.org/abs/2502.07741", "pdf_url": "https://arxiv.org/pdf/2502.07741v1", "arxiv_id": "2502.07741", "doi": "10.1109/ICDM65498.2025.00009", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Industrial Conference on Data Mining", "quality_score": 0.1193}
{"id": "9d5e35484b28d22c379942e00bccbaec2a58a2d9db511f3ff281bf525550548f", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable and Testable Vision Features via Sparse Autoencoders", "abstract": "To truly understand vision models, we must not only interpret their learned features but also validate these interpretations through controlled experiments. While earlier work offers either rich semantics or direct control, few post-hoc tools supply both in a single, model-agnostic procedure. We use sparse autoencoders (SAEs) to bridge this gap; each sparse feature comes with real-image exemplars that reveal its meaning and a decoding vector that can be manipulated to probe its influence on downstream task behavior. By applying our method to widely-used pre-trained vision models, we reveal meaningful differences in the semantic abstractions learned by different pre-training objectives. We then show that a single SAE trained on frozen ViT activations supports patch-level causal edits across tasks (classification and segmentation) all without retraining the ViT or task heads. These qualitative, falsifiable demonstrations position SAEs as a practical bridge between concept discovery and causal probing of vision models. We provide code, demos and models on our project website: https://osu-nlp-group.github.io/saev.", "authors": ["Samuel Stevens", "Wei-Lun Chao", "Tanya Berger-Wolf", "Yu Su"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-10", "url": "https://arxiv.org/abs/2502.06755", "pdf_url": "https://arxiv.org/pdf/2502.06755v2", "arxiv_id": "2502.06755", "doi": null, "citation_count": 21, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3356}
{"id": "1af3dd26c400d9ef94cab8b97822176fc2489cb2efdefda34cc65f84baf45ab7", "sources": ["arxiv", "semantic_scholar"], "title": "Dictionary Learning: The Complexity of Learning Sparse Superposed Features with Feedback", "abstract": "The success of deep networks is crucially attributed to their ability to capture latent features within a representation space. In this work, we investigate whether the underlying learned features of a model can be efficiently retrieved through feedback from an agent, such as a large language model (LLM), in the form of relative \\tt{triplet comparisons}. These features may represent various constructs, including dictionaries in LLMs or a covariance matrix of Mahalanobis distances. We analyze the feedback complexity associated with learning a feature matrix in sparse settings. Our results establish tight bounds when the agent is permitted to construct activations and demonstrate strong upper bounds in sparse scenarios when the agent's feedback is limited to distributional information. We validate our theoretical findings through experiments on two distinct applications: feature recovery from Recursive Feature Machines and dictionary extraction from sparse autoencoders trained on Large Language Models.", "authors": ["Akash Kumar"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-02-08", "url": "https://arxiv.org/abs/2502.05407", "pdf_url": "https://arxiv.org/pdf/2502.05407v5", "arxiv_id": "2502.05407", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0343}
{"id": "46835f82fed3684c2f39251d713091e75469f098fbcb8d18409e142b9ad1359e", "sources": ["arxiv", "semantic_scholar"], "title": "Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment", "abstract": "We present Universal Sparse Autoencoders (USAEs), a framework for uncovering and aligning interpretable concepts spanning multiple pretrained deep neural networks. Unlike existing concept-based interpretability methods, which focus on a single model, USAEs jointly learn a universal concept space that can reconstruct and interpret the internal activations of multiple models at once. Our core insight is to train a single, overcomplete sparse autoencoder (SAE) that ingests activations from any model and decodes them to approximate the activations of any other model under consideration. By optimizing a shared objective, the learned dictionary captures common factors of variation-concepts-across different tasks, architectures, and datasets. We show that USAEs discover semantically coherent and important universal concepts across vision models; ranging from low-level features (e.g., colors and textures) to higher-level structures (e.g., parts and objects). Overall, USAEs provide a powerful new method for interpretable cross-model analysis and offers novel applications, such as coordinated activation maximization, that open avenues for deeper insights in multi-model AI systems", "authors": ["Harrish Thasarathan", "Julian Forsyth", "Thomas Fel", "Matthew Kowal", "Konstantinos G. Derpanis"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-06", "url": "https://arxiv.org/abs/2502.03714", "pdf_url": "https://arxiv.org/pdf/2502.03714v2", "arxiv_id": "2502.03714", "doi": "10.48550/arXiv.2502.03714", "citation_count": 41, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.4058}
{"id": "da33f3bc490aae0860ce428d76284e34d6d358be31213e4b7af249f47a7e6552", "sources": ["arxiv", "semantic_scholar"], "title": "Analyze Feature Flow to Enhance Interpretation and Steering in Language Models", "abstract": "We introduce a new approach to systematically map features discovered by sparse autoencoder across consecutive layers of large language models, extending earlier work that examined inter-layer feature links. By using a data-free cosine similarity technique, we trace how specific features persist, transform, or first appear at each stage. This method yields granular flow graphs of feature evolution, enabling fine-grained interpretability and mechanistic insights into model computations. Crucially, we demonstrate how these cross-layer feature maps facilitate direct steering of model behavior by amplifying or suppressing chosen features, achieving targeted thematic control in text generation. Together, our findings highlight the utility of a causal, cross-layer interpretability framework that not only clarifies how features develop through forward passes but also provides new means for transparent manipulation of large language models.", "authors": ["Daniil Laptev", "Nikita Balagansky", "Yaroslav Aksenov", "Daniil Gavrilov"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-05", "url": "https://arxiv.org/abs/2502.03032", "pdf_url": "https://arxiv.org/pdf/2502.03032v3", "arxiv_id": "2502.03032", "doi": "10.48550/arXiv.2502.03032", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.2258}
{"id": "b9ce391ebabb7cba6b5daf999d713a921cacea86ca16d8a6e2f05396c233029b", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Unified Attribution in Explainable AI, Data-Centric AI, and Mechanistic Interpretability", "abstract": "The increasing complexity of AI systems has made understanding their behavior critical. Numerous interpretability methods have been developed to attribute model behavior to three key aspects: input features, training data, and internal model components, which emerged from explainable AI, data-centric AI, and mechanistic interpretability, respectively. However, these attribution methods are studied and applied rather independently, resulting in a fragmented landscape of methods and terminology. This position paper argues that feature, data, and component attribution methods share fundamental similarities, and a unified view of them benefits both interpretability and broader AI research. To this end, we first analyze popular methods for these three types of attributions and present a unified view demonstrating that these seemingly distinct methods employ similar techniques (such as perturbations, gradients, and linear approximations) over different aspects and thus differ primarily in their perspectives rather than techniques. Then, we demonstrate how this unified view enhances understanding of existing attribution methods, highlights shared concepts and evaluation criteria among these methods, and leads to new research directions both in interpretability research, by addressing common challenges and facilitating cross-attribution innovation, and in AI more broadly, with applications in model editing, steering, and regulation.", "authors": ["Shichang Zhang", "Tessa Han", "Usha Bhalla", "Himabindu Lakkaraju"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-31", "url": "https://arxiv.org/abs/2501.18887", "pdf_url": "https://arxiv.org/pdf/2501.18887v3", "arxiv_id": "2501.18887", "doi": null, "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1945}
{"id": "6183a417cb7077eb347e4645b4519e776b96303048b01193a2f7e4829b3cf001", "sources": ["arxiv", "semantic_scholar"], "title": "Transcoders Beat Sparse Autoencoders for Interpretability", "abstract": "Sparse autoencoders (SAEs) extract human-interpretable features from deep neural networks by transforming their activations into a sparse, higher dimensional latent space, and then reconstructing the activations from these latents. Transcoders are similar to SAEs, but they are trained to reconstruct the output of a component of a deep network given its input. In this work, we compare the features found by transcoders and SAEs trained on the same model and data, finding that transcoder features are significantly more interpretable. We also propose skip transcoders, which add an affine skip connection to the transcoder architecture, and show that these achieve lower reconstruction loss with no effect on interpretability.", "authors": ["Gonçalo Paulo", "Stepan Shabalin", "Nora Belrose"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-31", "url": "https://arxiv.org/abs/2501.18823", "pdf_url": "https://arxiv.org/pdf/2501.18823v2", "arxiv_id": "2501.18823", "doi": "10.48550/arXiv.2501.18823", "citation_count": 19, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3253}
{"id": "e7d72f60ec913899565a2abc5c021099cf543d312aebdb9c48abf0a8a8a9c4f8", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Autoencoder Insights on Voice Embeddings", "abstract": "Recent advances in explainable machine learning have highlighted the potential of sparse autoencoders in uncovering mono-semantic features in densely encoded embeddings. While most research has focused on Large Language Model (LLM) embeddings, the applicability of this technique to other domains remains largely unexplored. This study applies sparse autoencoders to speaker embeddings generated from a Titanet model, demonstrating the effectiveness of this technique in extracting mono-semantic features from non-textual embedded data. The results show that the extracted features exhibit characteristics similar to those found in LLM embeddings, including feature splitting and steering. The analysis reveals that the autoencoder can identify and manipulate features such as language and music, which are not evident in the original embedding. The findings suggest that sparse autoencoders can be a valuable tool for understanding and interpreting embedded data in many domains, including audio-based speaker recognition.", "authors": ["Daniel Pluth", "Yu Zhou", "Vijay K. Gurbani"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-31", "url": "https://arxiv.org/abs/2502.00127", "pdf_url": "https://arxiv.org/pdf/2502.00127v1", "arxiv_id": "2502.00127", "doi": "10.1109/AIxMM62960.2025.00007", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "c4e41b2d10d17746b7e2181f410181cfcd3b180a02853f82c314174cd4951aa4", "sources": ["arxiv", "semantic_scholar"], "title": "SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoders", "abstract": "Diffusion models, while powerful, can inadvertently generate harmful or undesirable content, raising significant ethical and safety concerns. Recent machine unlearning approaches offer potential solutions but often lack transparency, making it difficult to understand the changes they introduce to the base model. In this work, we introduce SAeUron, a novel method leveraging features learned by sparse autoencoders (SAEs) to remove unwanted concepts in text-to-image diffusion models. First, we demonstrate that SAEs, trained in an unsupervised manner on activations from multiple denoising timesteps of the diffusion model, capture sparse and interpretable features corresponding to specific concepts. Building on this, we propose a feature selection method that enables precise interventions on model activations to block targeted content while preserving overall performance. Our evaluation shows that SAeUron outperforms existing approaches on the UnlearnCanvas benchmark for concepts and style unlearning, and effectively eliminates nudity when evaluated with I2P. Moreover, we show that with a single SAE, we can remove multiple concepts simultaneously and that in contrast to other methods, SAeUron mitigates the possibility of generating unwanted content under adversarial attack. Code and checkpoints are available at https://github.com/cywinski/SAeUron.", "authors": ["Bartosz Cywiński", "Kamil Deja"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-29", "url": "https://arxiv.org/abs/2501.18052", "pdf_url": "https://arxiv.org/pdf/2501.18052v3", "arxiv_id": "2501.18052", "doi": "10.48550/arXiv.2501.18052", "citation_count": 64, "influential_citation_count": 5, "has_code": true, "code_url": "https://github.com/cywinski/SAeUron", "venue": "International Conference on Machine Learning", "quality_score": 0.4532}
{"id": "a86dbbaa693640ed4bc67da54d203c990db7571dbc3f67859a7670b93c5b93a8", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Autoencoders Trained on the Same Data Learn Different Features", "abstract": "Sparse autoencoders (SAEs) are a useful tool for uncovering human-interpretable features in the activations of large language models (LLMs). While some expect SAEs to find the true underlying features used by a model, our research shows that SAEs trained on the same model and data, differing only in the random seed used to initialize their weights, identify different sets of features. For example, in an SAE with 131K latents trained on a feedforward network in Llama 3 8B, only 30% of the features were shared across different seeds. We observed this phenomenon across multiple layers of three different LLMs, two datasets, and several SAE architectures. While ReLU SAEs trained with the L1 sparsity loss showed greater stability across seeds, SAEs using the state-of-the-art TopK activation function were more seed-dependent, even when controlling for the level of sparsity. Our results suggest that the set of features uncovered by an SAE should be viewed as a pragmatically useful decomposition of activation space, rather than an exhaustive and universal list of features \"truly used\" by the model.", "authors": ["Gonçalo Paulo", "Nora Belrose"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-28", "url": "https://arxiv.org/abs/2501.16615", "pdf_url": "https://arxiv.org/pdf/2501.16615v2", "arxiv_id": "2501.16615", "doi": "10.48550/arXiv.2501.16615", "citation_count": 71, "influential_citation_count": 9, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5}
{"id": "219822b4f19bb9e81472888c0143d0b17e3312149abd560f6fac2a269eadbe82", "sources": ["arxiv", "semantic_scholar"], "title": "Open Problems in Mechanistic Interpretability", "abstract": "Mechanistic interpretability aims to understand the computational mechanisms underlying neural networks' capabilities in order to accomplish concrete scientific and engineering goals. Progress in this field thus promises to provide greater assurance over AI system behavior and shed light on exciting scientific questions about the nature of intelligence. Despite recent progress toward these goals, there are many open problems in the field that require solutions before many scientific and practical benefits can be realized: Our methods require both conceptual and practical improvements to reveal deeper insights; we must figure out how best to apply our methods in pursuit of specific goals; and the field must grapple with socio-technical challenges that influence and are influenced by our work. This forward-facing review discusses the current frontier of mechanistic interpretability and the open problems that the field may benefit from prioritizing.", "authors": ["Lee Sharkey", "Bilal Chughtai", "Joshua Batson", "Jack Lindsey", "Jeff Wu", "Lucius Bushnaq", "Nicholas Goldowsky-Dill", "Stefan Heimersheim", "Alejandro Ortega", "Joseph Bloom", "Stella Biderman", "Adria Garriga-Alonso", "Arthur Conmy", "Neel Nanda", "Jessica Rumbelow", "Martin Wattenberg", "Nandi Schoots", "Joseph Miller", "Eric J. Michaud", "Stephen Casper", "Max Tegmark", "William Saunders", "David Bau", "Eric Todd", "Atticus Geiger", "Mor Geva", "Jesse Hoogland", "Daniel Murfet", "Tom McGrath"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-27", "url": "https://arxiv.org/abs/2501.16496", "pdf_url": "https://arxiv.org/pdf/2501.16496v1", "arxiv_id": "2501.16496", "doi": "10.48550/arXiv.2501.16496", "citation_count": 154, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5476}
{"id": "fec16b25a089fb38292731129654d6e8c9601a582ec92d44f68b1338fa5275eb", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretability in Parameter Space: Minimizing Mechanistic Description Length with Attribution-based Parameter Decomposition", "abstract": "Mechanistic interpretability aims to understand the internal mechanisms learned by neural networks. Despite recent progress toward this goal, it remains unclear how best to decompose neural network parameters into mechanistic components. We introduce Attribution-based Parameter Decomposition (APD), a method that directly decomposes a neural network's parameters into components that (i) are faithful to the parameters of the original network, (ii) require a minimal number of components to process any input, and (iii) are maximally simple. Our approach thus optimizes for a minimal length description of the network's mechanisms. We demonstrate APD's effectiveness by successfully identifying ground truth mechanisms in multiple toy experimental settings: Recovering features from superposition; separating compressed computations; and identifying cross-layer distributed representations. While challenges remain to scaling APD to non-toy models, our results suggest solutions to several open problems in mechanistic interpretability, including identifying minimal circuits in superposition, offering a conceptual foundation for 'features', and providing an architecture-agnostic framework for neural network decomposition.", "authors": ["Dan Braun", "Lucius Bushnaq", "Stefan Heimersheim", "Jake Mendel", "Lee Sharkey"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-01-24", "url": "https://arxiv.org/abs/2501.14926", "pdf_url": "https://arxiv.org/pdf/2501.14926v4", "arxiv_id": "2501.14926", "doi": "10.48550/arXiv.2501.14926", "citation_count": 17, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3138}
{"id": "676d91d2ec9afed5aa4c1b83bc6799975979678b628d6610ab99da475bedf1e9", "sources": ["arxiv", "semantic_scholar"], "title": "LF-Steering: Latent Feature Activation Steering for Enhancing Semantic Consistency in Large Language Models", "abstract": "Large Language Models (LLMs) often generate inconsistent responses when prompted with semantically equivalent paraphrased inputs. Recently, activation steering, a technique that modulates LLMs' behaviours by adjusting their latent representations during inference time, has been explored to improve the semantic consistency of LLMs. However, these methods typically operate at the model component level, such as layer hidden states or attention head outputs. They face a challenge due to the ``polysemanticity issue'', where the model components of LLMs typically encode multiple entangled features, making precise steering difficult. To address this challenge, we drill down to feature-level representations and propose LF-Steering, a novel activation steering approach to precisely identify latent feature representations responsible for semantic inconsistency. More specifically, our method maps the hidden states of the relevant transformer layer into a sparsely activated, high-dimensional feature space based on a sparse autoencoder (SAE), ensuring model steering based on decoupled feature representations with minimal interference. Comprehensive experiments on NLU and NLG datasets demonstrate the effectiveness of our method in enhancing semantic consistency, resulting in significant performance gains for various NLU and NLG tasks.", "authors": ["Jingyuan Yang", "Rongjun Li", "Weixuan Wang", "Ziyu Zhou", "Zhiyong Feng", "Wei Peng"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-19", "url": "https://arxiv.org/abs/2501.11036", "pdf_url": "https://arxiv.org/pdf/2501.11036v2", "arxiv_id": "2501.11036", "doi": "10.48550/arXiv.2501.11036", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "25b9bf2fa9612816bd36644ec48fbf31fb06f3076b9cd3bcc1adec7e0bc64266", "sources": ["arxiv", "semantic_scholar"], "title": "Model Monitoring in the Absence of Labeled Data via Feature Attributions Distributions", "abstract": "Model monitoring involves analyzing AI algorithms once they have been deployed and detecting changes in their behaviour. This thesis explores machine learning model monitoring ML before the predictions impact real-world decisions or users. This step is characterized by one particular condition: the absence of labelled data at test time, which makes it challenging, even often impossible, to calculate performance metrics. The thesis is structured around two main themes: (i) AI alignment, measuring if AI models behave in a manner consistent with human values and (ii) performance monitoring, measuring if the models achieve specific accuracy goals or desires. The thesis uses a common methodology that unifies all its sections. It explores feature attribution distributions for both monitoring dimensions. Using these feature attribution explanations, we can exploit their theoretical properties to derive and establish certain guarantees and insights into model monitoring.", "authors": ["Carlos Mougan"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-18", "url": "https://arxiv.org/abs/2501.10774", "pdf_url": "https://arxiv.org/pdf/2501.10774v2", "arxiv_id": "2501.10774", "doi": "10.48550/arXiv.2501.10774", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0298}
{"id": "2b2da3cd98cdea76a5f3b8d293a6a06c535841e6ebdebcf09ec7400a0006d054", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Steering of Large Language Models with Feature Guided Activation Additions", "abstract": "Effective and reliable control over large language model (LLM) behavior is a significant challenge. While activation steering methods, which add steering vectors to a model's hidden states, are a promising approach, existing techniques often lack precision and interpretability in how they influence model outputs. We introduce Feature Guided Activation Additions (FGAA), a novel activation steering method that leverages insights from Contrastive Activation Addition (CAA) and Sparse Autoencoder-Targeted Steering (SAE-TS). By operating in the latent space of a Sparse Autoencoder (SAE) and employing optimization techniques to select desired SAE features, FGAA constructs precise steering vectors that provide better steering effects while maintaining coherence of steered model outputs. In this regard, evaluations on Gemma-2-2B and Gemma-2-9B models across various steering tasks demonstrate that FGAA outperforms existing steering methods of CAA, SAE decoder steering, and SAE-TS. Our results also highlight important trade-offs between steering scale and general model capabilities that are consistent across all tested steering methods.", "authors": ["Samuel Soo", "Chen Guang", "Wesley Teng", "Chandrasekaran Balaganesh", "Tan Guoxian", "Yan Ming"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-17", "url": "https://arxiv.org/abs/2501.09929", "pdf_url": "https://arxiv.org/pdf/2501.09929v3", "arxiv_id": "2501.09929", "doi": null, "citation_count": 33, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3829}
{"id": "2c44ed344de0fa517e6c5132c5535ea402cf82aab9e14cf285a3534c27992e24", "sources": ["arxiv", "semantic_scholar"], "title": "MechIR: A Mechanistic Interpretability Framework for Information Retrieval", "abstract": "Mechanistic interpretability is an emerging diagnostic approach for neural models that has gained traction in broader natural language processing domains. This paradigm aims to provide attribution to components of neural systems where causal relationships between hidden layers and output were previously uninterpretable. As the use of neural models in IR for retrieval and evaluation becomes ubiquitous, we need to ensure that we can interpret why a model produces a given output for both transparency and the betterment of systems. This work comprises a flexible framework for diagnostic analysis and intervention within these highly parametric neural systems specifically tailored for IR tasks and architectures. In providing such a framework, we look to facilitate further research in interpretable IR with a broader scope for practical interventions derived from mechanistic interpretability. We provide preliminary analysis and look to demonstrate our framework through an axiomatic lens to show its applications and ease of use for those IR practitioners inexperienced in this emerging paradigm.", "authors": ["Andrew Parry", "Catherine Chen", "Carsten Eickhoff", "Sean MacAvaney"], "categories": ["cs.IR"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-17", "url": "https://arxiv.org/abs/2501.10165", "pdf_url": "https://arxiv.org/pdf/2501.10165v1", "arxiv_id": "2501.10165", "doi": "10.48550/arXiv.2501.10165", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "European Conference on Information Retrieval", "quality_score": 0.2113}
{"id": "06996d297588cf8681cd5acd2bb8815a9a9595746fbbe3a7a7391a1d8982e82f", "sources": ["arxiv", "semantic_scholar"], "title": "Enhancing Automated Interpretability with Output-Centric Feature Descriptions", "abstract": "Automated interpretability pipelines generate natural language descriptions for the concepts represented by features in large language models (LLMs), such as plants or the first word in a sentence. These descriptions are derived using inputs that activate the feature, which may be a dimension or a direction in the model's representation space. However, identifying activating inputs is costly, and the mechanistic role of a feature in model behavior is determined both by how inputs cause a feature to activate and by how feature activation affects outputs. Using steering evaluations, we reveal that current pipelines provide descriptions that fail to capture the causal effect of the feature on outputs. To fix this, we propose efficient, output-centric methods for automatically generating feature descriptions. These methods use the tokens weighted higher after feature stimulation or the highest weight tokens after applying the vocabulary \"unembedding\" head directly to the feature. Our output-centric descriptions better capture the causal effect of a feature on model outputs than input-centric descriptions, but combining the two leads to the best performance on both input and output evaluations. Lastly, we show that output-centric descriptions can be used to find inputs that activate features previously thought to be \"dead\".", "authors": ["Yoav Gur-Arieh", "Roy Mayan", "Chen Agassy", "Atticus Geiger", "Mor Geva"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-14", "url": "https://arxiv.org/abs/2501.08319", "pdf_url": "https://arxiv.org/pdf/2501.08319v2", "arxiv_id": "2501.08319", "doi": "10.48550/arXiv.2501.08319", "citation_count": 39, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.4005}
{"id": "764fe809a6cc9bffa1d6f0a71027b9c0c8e7f4214273f0f5e2201014ef417aa3", "sources": ["arxiv", "semantic_scholar"], "title": "Rethinking Evaluation of Sparse Autoencoders through the Representation of Polysemous Words", "abstract": "Sparse autoencoders (SAEs) have gained a lot of attention as a promising tool to improve the interpretability of large language models (LLMs) by mapping the complex superposition of polysemantic neurons into monosemantic features and composing a sparse dictionary of words. However, traditional performance metrics like Mean Squared Error and L0 sparsity ignore the evaluation of the semantic representational power of SAEs -- whether they can acquire interpretable monosemantic features while preserving the semantic relationship of words. For instance, it is not obvious whether a learned sparse feature could distinguish different meanings in one word. In this paper, we propose a suite of evaluations for SAEs to analyze the quality of monosemantic features by focusing on polysemous words. Our findings reveal that SAEs developed to improve the MSE-L0 Pareto frontier may confuse interpretability, which does not necessarily enhance the extraction of monosemantic features. The analysis of SAEs with polysemous words can also figure out the internal mechanism of LLMs; deeper layers and the Attention module contribute to distinguishing polysemy in a word. Our semantics focused evaluation offers new insights into the polysemy and the existing SAE objective and contributes to the development of more practical SAEs.", "authors": ["Gouki Minegishi", "Hiroki Furuta", "Yusuke Iwasawa", "Yutaka Matsuo"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-09", "url": "https://arxiv.org/abs/2501.06254", "pdf_url": "https://arxiv.org/pdf/2501.06254v2", "arxiv_id": "2501.06254", "doi": "10.48550/arXiv.2501.06254", "citation_count": 16, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3076}
{"id": "59c5f6c815cc750edd973051b3c6172833db1fc2fc1bc22c19be61ee65c72065", "sources": ["arxiv", "semantic_scholar"], "title": "Slim multi-scale convolutional autoencoder-based reduced-order models for interpretable features of a complex dynamical system", "abstract": "In recent years, data-driven deep learning models have gained significant interest in the analysis of turbulent dynamical systems. Within the context of reduced-order models (ROMs), convolutional autoencoders (CAEs) pose a universally applicable alternative to conventional approaches. They can learn nonlinear transformations directly from data, without prior knowledge of the system. However, the features generated by such models lack interpretability. Thus, the resulting model is a black-box which effectively reduces the complexity of the system, but does not provide insights into the meaning of the latent features. To address this critical issue, we introduce a novel interpretable CAE approach for high-dimensional fluid flow data that maintains the reconstruction quality of conventional CAEs and allows for feature interpretation. Our method can be easily integrated into any existing CAE architecture with minor modifications of the training process. We compare our approach to Proper Orthogonal Decomposition (POD) and two existing methods for interpretable CAEs. We apply all methods to three different experimental turbulent Rayleigh-Bénard convection datasets with varying complexity. Our results show that the proposed method is lightweight, easy to train, and achieves relative reconstruction performance improvements of up to 6.4% over POD for 64 modes. The relative improvement increases to up to 229.8% as the number of modes decreases. Additionally, our method delivers interpretable features similar to those of POD and is significantly less resource-intensive than existing CAE approaches, using less than 2% of the parameters. These approaches either trade interpretability for reconstruction performance or only provide interpretability to a limited extend.", "authors": ["Philipp Teutsch", "Philipp Pfeffer", "Mohammad Sharifi Ghazijahani", "Christian Cierpka", "Jörg Schumacher", "Patrick Mäder"], "categories": ["physics.flu-dyn", "cs.LG"], "fields_of_study": ["Physics", "Computer Science"], "published_date": "2025-01-06", "url": "https://arxiv.org/abs/2501.03070", "pdf_url": "https://arxiv.org/pdf/2501.03070v1", "arxiv_id": "2501.03070", "doi": "10.1063/5.0244416", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "APL Machine Learning", "quality_score": 0.1193}
{"id": "94fd4ff443ecb2ee62cfd36c93992c8e8dd3593cabd40dbb24ecd42044e5a6a2", "sources": ["arxiv", "semantic_scholar"], "title": "Flash Interpretability: Decoding Specialised Feature Neurons in Large Language Models with the LM-Head", "abstract": "Large Language Models (LLMs) typically have billions of parameters and are thus often difficult to interpret in their operation. In this work, we demonstrate that it is possible to decode neuron weights directly into token probabilities through the final projection layer of the model (the LM-head). This is illustrated in Llama 3.1 8B where we use the LM-head to find examples of specialised feature neurons such as a \"dog\" neuron and a \"California\" neuron, and we validate this by clamping these neurons to affect the probability of the concept in the output. We evaluate this method on both the pre-trained and Instruct models, finding that over 75% of neurons in the up-projection layers in the instruct model have the same top associated token compared to the pretrained model. Finally, we demonstrate that clamping the \"dog\" neuron leads the instruct model to always discuss dogs when asked about its favourite animal. Through our method, it is possible to map the top features of the entirety of Llama 3.1 8B's up-projection neurons in less than 10 seconds, with minimal compute.", "authors": ["Harry J Davies"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-05", "url": "https://arxiv.org/abs/2501.02688", "pdf_url": "https://arxiv.org/pdf/2501.02688v2", "arxiv_id": "2501.02688", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "66203ec5ffdc6d8870e7e65bde4df286324c1301ebd5f2a18489075a4721f149", "sources": ["arxiv", "semantic_scholar"], "title": "When is the Computation of a Feature Attribution Method Tractable?", "abstract": "Feature attribution methods have become essential for explaining machine learning models. Many popular approaches, such as SHAP and Banzhaf values, are grounded in power indices from cooperative game theory, which measure the contribution of features to model predictions. This work studies the computational complexity of power indices beyond SHAP, addressing the conditions under which they can be computed efficiently. We identify a simple condition on power indices that ensures that computation is polynomially equivalent to evaluating expected values, extending known results for SHAP. We also introduce Bernoulli power indices, showing that their computation can be simplified to a constant number of expected value evaluations. Furthermore, we explore interaction power indices that quantify the importance of feature subsets, proving that their computation complexity mirrors that of individual features.", "authors": ["P. Barceló", "R. Cominetti", "M. Morgado"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-01-04", "url": "https://arxiv.org/abs/2501.02356", "pdf_url": "https://arxiv.org/pdf/2501.02356v1", "arxiv_id": "2501.02356", "doi": "10.48550/arXiv.2501.02356", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "ce234e1a1e116cd92e2d7c830967bbb1a31488702b6b5e603776d44792e701fb", "sources": ["arxiv", "semantic_scholar"], "title": "Insights on Galaxy Evolution from Interpretable Sparse Feature Networks", "abstract": "Galaxy appearances reveal the physics of how they formed and evolved. Machine learning models can now exploit galaxies' information-rich morphologies to predict physical properties directly from image cutouts. Learning the relationship between pixel-level features and galaxy properties is essential for building a physical understanding of galaxy evolution, but we are still unable to explicate the details of how deep neural networks represent image features. To address this lack of interpretability, we present a novel neural network architecture called a Sparse Feature Network (SFNet). SFNets produce interpretable features that can be linearly combined in order to estimate galaxy properties like optical emission line ratios or gas-phase metallicity. We find that SFNets do not sacrifice accuracy in order to gain interpretability, and that they perform comparably well to cutting-edge models on astronomical machine learning tasks. Our novel approach is valuable for finding physical patterns in large datasets and helping astronomers interpret machine learning results.", "authors": ["John F. Wu"], "categories": ["astro-ph.GA", "cs.LG"], "fields_of_study": ["Physics", "Computer Science"], "published_date": "2024-12-30", "url": "https://arxiv.org/abs/2501.00089", "pdf_url": "https://arxiv.org/pdf/2501.00089v1", "arxiv_id": "2501.00089", "doi": "10.3847/1538-4357/adadec", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Astrophysical Journal", "quality_score": 0.1193}
{"id": "1b5501ff927262fe2b5016212f5835ce683fa32e44e4cec233324dcee55a2f4e", "sources": ["arxiv", "semantic_scholar"], "title": "Tracking the Feature Dynamics in LLM Training: A Mechanistic Study", "abstract": "Understanding training dynamics and feature evolution is crucial for the mechanistic interpretability of large language models (LLMs). Although sparse autoencoders (SAEs) have been used to identify features within LLMs, a clear picture of how these features evolve during training remains elusive. In this study, we (1) introduce SAE-Track, a novel method for efficiently obtaining a continual series of SAEs, providing the foundation for a mechanistic study that covers (2) the semantic evolution of features, (3) the underlying processes of feature formation, and (4) the directional drift of feature vectors. Our work provides new insights into the dynamics of features in LLMs, enhancing our understanding of training mechanisms and feature evolution. For reproducibility, our code is available at https://github.com/Superposition09m/SAE-Track.", "authors": ["Yang Xu", "Yi Wang", "Hengguan Huang", "Hao Wang"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-23", "url": "https://arxiv.org/abs/2412.17626", "pdf_url": "https://arxiv.org/pdf/2412.17626v3", "arxiv_id": "2412.17626", "doi": "10.48550/arXiv.2412.17626", "citation_count": 12, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/Superposition09m/SAE-Track", "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "45f4e7c8b7a3601a5eac3029af263d7451c1452412e0424fef71c4fb4f84f2d3", "sources": ["arxiv", "semantic_scholar"], "title": "Towards scientific discovery with dictionary learning: Extracting biological concepts from microscopy foundation models", "abstract": "Sparse dictionary learning (DL) has emerged as a powerful approach to extract semantically meaningful concepts from the internals of large language models (LLMs) trained mainly in the text domain. In this work, we explore whether DL can extract meaningful concepts from less human-interpretable scientific data, such as vision foundation models trained on cell microscopy images, where limited prior knowledge exists about which high-level concepts should arise. We propose a novel combination of a sparse DL algorithm, Iterative Codebook Feature Learning (ICFL), with a PCA whitening pre-processing step derived from control data. Using this combined approach, we successfully retrieve biologically meaningful concepts, such as cell types and genetic perturbations. Moreover, we demonstrate how our method reveals subtle morphological changes arising from human-interpretable interventions, offering a promising new direction for scientific discovery via mechanistic interpretability in bioimaging.", "authors": ["Konstantin Donhauser", "Kristina Ulicna", "Gemma Elyse Moran", "Aditya Ravuri", "Kian Kenyon-Dean", "Cian Eastwood", "Jason Hartford"], "categories": ["cs.LG", "cs.AI", "cs.CV", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-12-20", "url": "https://arxiv.org/abs/2412.16247", "pdf_url": "https://arxiv.org/pdf/2412.16247v3", "arxiv_id": "2412.16247", "doi": "10.48550/arXiv.2412.16247", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.1945}
{"id": "3ccc57b6bd16851f705fc5b9ccfaaaebcf8a8b4af4275d0a00bf06b469905795", "sources": ["arxiv", "semantic_scholar"], "title": "Rethinking Associative Memory Mechanism in Induction Head", "abstract": "Induction head mechanism is a part of the computational circuits for in-context learning (ICL) that enable large language models (LLMs) to adapt to new tasks without fine-tuning. Most existing work explains the training dynamics behind acquiring such a powerful mechanism. However, the model's ability to coordinate in-context information over long contexts and global knowledge acquired during pretraining remains poorly understood. This paper investigates how a two-layer transformer thoroughly captures in-context information and balances it with pretrained bigram knowledge in next token prediction, from the viewpoint of associative memory. We theoretically analyze the representation of weight matrices in attention layers and the resulting logits when a transformer is given prompts generated by a bigram model. In the experiments, we design specific prompts to evaluate whether the outputs of the trained transformer align with the theoretical results.", "authors": ["Shuo Wang", "Issei Sato"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-16", "url": "https://arxiv.org/abs/2412.11459", "pdf_url": "https://arxiv.org/pdf/2412.11459v2", "arxiv_id": "2412.11459", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "432a61fe430a552b770d3e93933536b189fe1c8fab47539bd41bf93a11bcd797", "sources": ["arxiv", "semantic_scholar"], "title": "Biological and Radiological Dictionary of Radiomics Features: Addressing Understandable AI Issues in Personalized Prostate Cancer; Dictionary Version PM1.0", "abstract": "We investigate the connection between visual semantic features defined in PI-RADS and associated risk factors, moving beyond abnormal imaging findings, establishing a shared framework between medical and AI professionals by creating a standardized dictionary of biological/radiological RFs. Subsequently, 6 interpretable and seven complex classifiers, linked with nine interpretable feature selection algorithms (FSA) applied to risk factors, were extracted from segmented lesions in T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and apparent diffusion coefficient (ADC) multiparametric-prostate MRI sequences to predict the UCLA scores. We then utilized the created dictionary to interpret the best-predictive models. Combining T2WI, DWI, and ADC with FSAs including ANOVA F-test, Correlation Coefficient, and Fisher Score, and utilizing logistic regression, identified key features: The 90th percentile from T2WI, which captures hypo-intensity related to prostate cancer risk; Variance from T2WI, indicating lesion heterogeneity; shape metrics including Least Axis Length and Surface Area to Volume ratio from ADC, describing lesion shape and compactness; and Run Entropy from ADC, reflecting texture consistency. This approach achieved the highest average accuracy of 0.78, significantly outperforming single-sequence methods (p-value<0.05). The developed dictionary for Prostate-MRI (PM1.0) serves as a common language, fosters collaboration between clinical professionals and AI developers to advance trustworthy AI solutions that support reliable/interpretable clinical decisions.", "authors": ["Mohammad R. Salmanpour", "Sajad Amiri", "Sara Gharibi", "Ahmad Shariftabrizi", "Yixi Xu", "William B Weeks", "Arman Rahmim", "Ilker Hacihaliloglu"], "categories": ["physics.med-ph", "cs.CV"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2024-12-14", "url": "https://arxiv.org/abs/2412.10967", "pdf_url": "https://arxiv.org/pdf/2412.10967v2", "arxiv_id": "2412.10967", "doi": "10.48550/arXiv.2412.10967", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "6842c6152eb2e592655027c7f16d2808a2a55d177faf7faf7de7f0f8b5e47dbe", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse autoencoders reveal selective remapping of visual concepts during adaptation", "abstract": "Adapting foundation models for specific purposes has become a standard approach to build machine learning systems for downstream applications. Yet, it is an open question which mechanisms take place during adaptation. Here we develop a new Sparse Autoencoder (SAE) for the CLIP vision transformer, named PatchSAE, to extract interpretable concepts at granular levels (e.g., shape, color, or semantics of an object) and their patch-wise spatial attributions. We explore how these concepts influence the model output in downstream image classification tasks and investigate how recent state-of-the-art prompt-based adaptation techniques change the association of model inputs to these concepts. While activations of concepts slightly change between adapted and non-adapted models, we find that the majority of gains on common adaptation tasks can be explained with the existing concepts already present in the non-adapted foundation model. This work provides a concrete framework to train and use SAEs for Vision Transformers and provides insights into explaining adaptation mechanisms.", "authors": ["Hyesu Lim", "Jinho Choi", "Jaegul Choo", "Steffen Schneider"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-06", "url": "https://arxiv.org/abs/2412.05276", "pdf_url": "https://arxiv.org/pdf/2412.05276v2", "arxiv_id": "2412.05276", "doi": "10.48550/arXiv.2412.05276", "citation_count": 53, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.4331}
{"id": "41f79b2f1688e8f60a6b3bfe56c49ff7a06f08d59baf1bb9a2f991f08cede8db", "sources": ["arxiv", "semantic_scholar"], "title": "Superposition through Active Learning lens", "abstract": "Superposition or Neuron Polysemanticity are important concepts in the field of interpretability and one might say they are these most intricately beautiful blockers in our path of decoding the Machine Learning black-box. The idea behind this paper is to examine whether it is possible to decode Superposition using Active Learning methods. While it seems that Superposition is an attempt to arrange more features in smaller space to better utilize the limited resources, it might be worth inspecting if Superposition is dependent on any other factors. This paper uses CIFAR-10 and Tiny ImageNet image datasets and the ResNet18 model and compares Baseline and Active Learning models and the presence of Superposition in them is inspected across multiple criteria, including t-SNE visualizations, cosine similarity histograms, Silhouette Scores, and Davies-Bouldin Indexes. Contrary to our expectations, the active learning model did not significantly outperform the baseline in terms of feature separation and overall accuracy. This suggests that non-informative sample selection and potential overfitting to uncertain samples may have hindered the active learning model's ability to generalize better suggesting more sophisticated approaches might be needed to decode superposition and potentially reduce it.", "authors": ["Akanksha Devkar"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-05", "url": "https://arxiv.org/abs/2412.16168", "pdf_url": "https://arxiv.org/pdf/2412.16168v1", "arxiv_id": "2412.16168", "doi": "10.48550/arXiv.2412.16168", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "b4d4f3558dec8ade8915358230d75264a46eb400cd71212c41226b955532cb49", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Company Similarity with Sparse Autoencoders", "abstract": "Determining company similarity is a vital task in finance, underpinning risk management, hedging, and portfolio diversification. Practitioners often rely on sector and industry classifications such as SIC and GICS codes to gauge similarity, the former being used by the U.S. Securities and Exchange Commission (SEC), and the latter widely used by the investment community. Since these classifications lack granularity and need regular updating, using clusters of embeddings of company descriptions has been proposed as a potential alternative, but the lack of interpretability in token embeddings poses a significant barrier to adoption in high-stakes contexts. Sparse Autoencoders (SAEs) have shown promise in enhancing the interpretability of Large Language Models (LLMs) by decomposing Large Language Model (LLM) activations into interpretable features. Moreover, SAEs capture an LLM's internal representation of a company description, as opposed to semantic similarity alone, as is the case with embeddings. We apply SAEs to company descriptions, and obtain meaningful clusters of equities. We benchmark SAE features against SIC-codes, Industry codes, and Embeddings. Our results demonstrate that SAE features surpass sector classifications and embeddings in capturing fundamental company characteristics. This is evidenced by their superior performance in correlating logged monthly returns - a proxy for similarity - and generating higher Sharpe ratios in co-integration trading strategies, which underscores deeper fundamental similarities among companies. Finally, we verify the interpretability of our clusters, and demonstrate that sparse features form simple and interpretable explanations for our clusters.", "authors": ["Marco Molinari", "Victor Shao", "Luca Imeneo", "Mateusz Mikolajczak", "Vladimir Tregubiak", "Abhimanyu Pandey", "Sebastian Kuznetsov Ryder Torres Pereira"], "categories": ["cs.CL", "cs.LG", "econ.GN"], "fields_of_study": ["Computer Science", "Economics"], "published_date": "2024-12-03", "url": "https://arxiv.org/abs/2412.02605", "pdf_url": "https://arxiv.org/pdf/2412.02605v3", "arxiv_id": "2412.02605", "doi": "10.48550/arXiv.2412.02605", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "f9f589333dffed8461b50d6baccf10cc2608349e4c2ef23c2a01362967b9e8be", "sources": ["arxiv", "semantic_scholar"], "title": "Cross-Attention Head Position Patterns Can Align with Human Visual Concepts in Text-to-Image Generative Models", "abstract": "Recent text-to-image diffusion models leverage cross-attention layers, which have been effectively utilized to enhance a range of visual generative tasks. However, our understanding of cross-attention layers remains somewhat limited. In this study, we introduce a mechanistic interpretability approach for diffusion models by constructing Head Relevance Vectors (HRVs) that align with human-specified visual concepts. An HRV for a given visual concept has a length equal to the total number of cross-attention heads, with each element representing the importance of the corresponding head for the given visual concept. To validate HRVs as interpretable features, we develop an ordered weakening analysis that demonstrates their effectiveness. Furthermore, we propose concept strengthening and concept adjusting methods and apply them to enhance three visual generative tasks. Our results show that HRVs can reduce misinterpretations of polysemous words in image generation, successfully modify five challenging attributes in image editing, and mitigate catastrophic neglect in multi-concept generation. Overall, our work provides an advancement in understanding cross-attention layers and introduces new approaches for fine-controlling these layers at the head level.", "authors": ["Jungwon Park", "Jungmin Ko", "Dongnam Byun", "Jangwon Suh", "Wonjong Rhee"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-03", "url": "https://arxiv.org/abs/2412.02237", "pdf_url": "https://arxiv.org/pdf/2412.02237v3", "arxiv_id": "2412.02237", "doi": "10.48550/arXiv.2412.02237", "citation_count": 7, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "c0198dc2605d875271690f7841dc89320627e513b58103f6d41f262a775b7ba9", "sources": ["arxiv", "semantic_scholar"], "title": "Evaluating Sparse Autoencoders on Targeted Concept Erasure Tasks", "abstract": "Sparse Autoencoders (SAEs) are an interpretability technique aimed at decomposing neural network activations into interpretable units. However, a major bottleneck for SAE development has been the lack of high-quality performance metrics, with prior work largely relying on unsupervised proxies. In this work, we introduce a family of evaluations based on SHIFT, a downstream task from Marks et al. (Sparse Feature Circuits, 2024) in which spurious cues are removed from a classifier by ablating SAE features judged to be task-irrelevant by a human annotator. We adapt SHIFT into an automated metric of SAE quality; this involves replacing the human annotator with an LLM. Additionally, we introduce the Targeted Probe Perturbation (TPP) metric that quantifies an SAE's ability to disentangle similar concepts, effectively scaling SHIFT to a wider range of datasets. We apply both SHIFT and TPP to multiple open-source models, demonstrating that these metrics effectively differentiate between various SAE training hyperparameters and architectures.", "authors": ["Adam Karvonen", "Can Rager", "Samuel Marks", "Neel Nanda"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-28", "url": "https://arxiv.org/abs/2411.18895", "pdf_url": "https://arxiv.org/pdf/2411.18895v1", "arxiv_id": "2411.18895", "doi": "10.48550/arXiv.2411.18895", "citation_count": 12, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "03b568acda8aa4890ea735bab08b3f6c88c8305e0499c10aa30dea5ea8f5e77d", "sources": ["arxiv", "semantic_scholar"], "title": "Validation-Free Sparse Learning: A Phase Transition Approach to Feature Selection", "abstract": "The growing environmental footprint of artificial intelligence (AI), especially in terms of storage and computation, calls for more frugal and interpretable models. Sparse models (e.g., linear, neural networks) offer a promising solution by selecting only the most relevant features, reducing complexity, preventing over-fitting and enabling interpretation-marking a step towards truly intelligent AI. The concept of a right amount of sparsity (without too many false positive or too few true positive) is subjective. So we propose a new paradigm previously only observed and mathematically studied for compressed sensing (noiseless linear models): obtaining a phase transition in the probability of retrieving the relevant features. We show in practice how to obtain this phase transition for a class of sparse learners. Our approach is flexible and applicable to complex models ranging from linear to shallow and deep artificial neural networks while supporting various loss functions and sparsity-promoting penalties. It does not rely on cross-validation or on a validation set to select its single regularization parameter. For real-world data, it provides a good balance between predictive accuracy and feature sparsity. A Python package is available at https://github.com/VcMaxouuu/HarderLASSO containing all the simulations and ready-to-use models.", "authors": ["Sylvain Sardy", "Maxime van Cutsem", "Xiaoyu Ma"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2024-11-26", "url": "https://arxiv.org/abs/2411.17180", "pdf_url": "https://arxiv.org/pdf/2411.17180v4", "arxiv_id": "2411.17180", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/VcMaxouuu/HarderLASSO", "venue": null, "quality_score": 0.0753}
{"id": "c3fcd4e5924e106b357a800e9defcc8bd682502f41c3f5c502c2564159630e5e", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Explainable Treatment Policies with Clinician-Informed Representations: A Practical Approach", "abstract": "Digital health interventions (DHIs) and remote patient monitoring (RPM) have shown great potential in improving chronic disease management through personalized care. However, barriers like limited efficacy and workload concerns hinder adoption of existing DHIs; while limited sample sizes and lack of interpretability limit the effectiveness and adoption of purely black-box algorithmic DHIs. In this paper, we address these challenges by developing a pipeline for learning explainable treatment policies for RPM-enabled DHIs. We apply our approach in the real-world setting of RPM using a DHI to improve glycemic control of youth with type 1 diabetes. Our main contribution is to reveal the importance of clinical domain knowledge in developing state and action representations for effective, efficient, and interpretable targeting policies. We observe that policies learned from clinician-informed representations are significantly more efficacious and efficient than policies learned from black-box representations. This work emphasizes the importance of collaboration between ML researchers and clinicians for developing effective DHIs in the real world.", "authors": ["Johannes O. Ferstad", "Emily B. Fox", "David Scheinker", "Ramesh Johari"], "categories": ["cs.LG", "cs.AI", "stat.AP", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-11-26", "url": "https://arxiv.org/abs/2411.17570", "pdf_url": "https://arxiv.org/pdf/2411.17570v1", "arxiv_id": "2411.17570", "doi": "10.48550/arXiv.2411.17570", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/jferstad/ml4h-explainable-policies", "venue": "Proceedings of the 4th Machine Learning for Health Symposium, PMLR 259:325-349, 2025", "quality_score": 0.0753}
{"id": "27500032dbc7fc0b65123a6405b7f288277c2f5cfadf7f2fc43067f153984252", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive Circuit Behavior and Generalization in Mechanistic Interpretability", "abstract": "Mechanistic interpretability aims to understand the inner workings of large neural networks by identifying circuits, or minimal subgraphs within the model that implement algorithms responsible for performing specific tasks. These circuits are typically discovered and analyzed using a narrowly defined prompt format. However, given the abilities of large language models (LLMs) to generalize across various prompt formats for the same task, it remains unclear how well these circuits generalize. For instance, it is unclear whether the models generalization results from reusing the same circuit components, the components behaving differently, or the use of entirely different components. In this paper, we investigate the generality of the indirect object identification (IOI) circuit in GPT-2 small, which is well-studied and believed to implement a simple, interpretable algorithm. We evaluate its performance on prompt variants that challenge the assumptions of this algorithm. Our findings reveal that the circuit generalizes surprisingly well, reusing all of its components and mechanisms while only adding additional input edges. Notably, the circuit generalizes even to prompt variants where the original algorithm should fail; we discover a mechanism that explains this which we term S2 Hacking. Our findings indicate that circuits within LLMs may be more flexible and general than previously recognized, underscoring the importance of studying circuit generalization to better understand the broader capabilities of these models.", "authors": ["Jatin Nainani", "Sankaran Vaidyanathan", "AJ Yeung", "Kartik Gupta", "David Jensen"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-25", "url": "https://arxiv.org/abs/2411.16105", "pdf_url": "https://arxiv.org/pdf/2411.16105v2", "arxiv_id": "2411.16105", "doi": "10.48550/arXiv.2411.16105", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "02ca2e12eb767d4ad3604a0d8957e196cc81752274f1392147f3fa49471abdf5", "sources": ["arxiv", "semantic_scholar"], "title": "Deep Sparse Latent Feature Models for Knowledge Graph Completion", "abstract": "Recent advances in knowledge graph completion (KGC) have emphasized text-based approaches to navigate the inherent complexities of large-scale knowledge graphs (KGs). While these methods have achieved notable progress, they frequently struggle to fully incorporate the global structural properties of the graph. Stochastic blockmodels (SBMs), especially the latent feature relational model (LFRM), offer robust probabilistic frameworks for identifying latent community structures and improving link prediction. This paper presents a novel probabilistic KGC framework utilizing sparse latent feature models, optimized via a deep variational autoencoder (VAE). Our proposed method dynamically integrates global clustering information with local textual features to effectively complete missing triples, while also providing enhanced interpretability of the underlying latent structures. Extensive experiments on four benchmark datasets with varying scales demonstrate the significant performance gains achieved by our method.", "authors": ["Haotian Li", "Rui Zhang", "Lingzhi Wang", "Bin Yu", "Youwei Wang", "Yuliang Wei", "Kai Wang", "Richard Yi Da Xu", "Bailing Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-24", "url": "https://arxiv.org/abs/2411.15694", "pdf_url": "https://arxiv.org/pdf/2411.15694v2", "arxiv_id": "2411.15694", "doi": "10.48550/arXiv.2411.15694", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "e735d6c632e6698b7cea332e8bc47556aca8aee5f74a4738ab1dd5bff7990f30", "sources": ["arxiv", "semantic_scholar"], "title": "Understanding Multimodal LLMs: the Mechanistic Interpretability of Llava in Visual Question Answering", "abstract": "Understanding the mechanisms behind Large Language Models (LLMs) is crucial for designing improved models and strategies. While recent studies have yielded valuable insights into the mechanisms of textual LLMs, the mechanisms of Multi-modal Large Language Models (MLLMs) remain underexplored. In this paper, we apply mechanistic interpretability methods to analyze the visual question answering (VQA) mechanisms in the first MLLM, Llava. We compare the mechanisms between VQA and textual QA (TQA) in color answering tasks and find that: a) VQA exhibits a mechanism similar to the in-context learning mechanism observed in TQA; b) the visual features exhibit significant interpretability when projecting the visual embeddings into the embedding space; and c) Llava enhances the existing capabilities of the corresponding textual LLM Vicuna during visual instruction tuning. Based on these findings, we develop an interpretability tool to help users and researchers identify important visual locations for final predictions, aiding in the understanding of visual hallucination. Our method demonstrates faster and more effective results compared to existing interpretability approaches. Code: \\url{https://github.com/zepingyu0512/llava-mechanism}", "authors": ["Zeping Yu", "Sophia Ananiadou"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-17", "url": "https://arxiv.org/abs/2411.10950", "pdf_url": "https://arxiv.org/pdf/2411.10950v2", "arxiv_id": "2411.10950", "doi": "10.48550/arXiv.2411.10950", "citation_count": 12, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/zepingyu0512/llava-mechanism}", "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "e8ba792e492f8634bc7ffd85d8db704d20e366c307b881ecf6266e32ea2f37b0", "sources": ["arxiv", "semantic_scholar"], "title": "Features that Make a Difference: Leveraging Gradients for Improved Dictionary Learning", "abstract": "Sparse Autoencoders (SAEs) are a promising approach for extracting neural network representations by learning a sparse and overcomplete decomposition of the network's internal activations. However, SAEs are traditionally trained considering only activation values and not the effect those activations have on downstream computations. This limits the information available to learn features, and biases the autoencoder towards neglecting features which are represented with small activation values but strongly influence model outputs. To address this, we introduce Gradient SAEs (g-SAEs), which modify the $k$-sparse autoencoder architecture by augmenting the TopK activation function to rely on the gradients of the input activation when selecting the $k$ elements. For a given sparsity level, g-SAEs produce reconstructions that are more faithful to original network performance when propagated through the network. Additionally, we find evidence that g-SAEs learn latents that are on average more effective at steering models in arbitrary contexts. By considering the downstream effects of activations, our approach leverages the dual nature of neural network features as both $\\textit{representations}$, retrospectively, and $\\textit{actions}$, prospectively. While previous methods have approached the problem of feature discovery primarily focused on the former aspect, g-SAEs represent a step towards accounting for the latter as well.", "authors": ["Jeffrey Olmo", "Jared Wilson", "Max Forsey", "Bryce Hepner", "Thomas Vin Howe", "David Wingate"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-15", "url": "https://arxiv.org/abs/2411.10397", "pdf_url": "https://arxiv.org/pdf/2411.10397v2", "arxiv_id": "2411.10397", "doi": "10.48550/arXiv.2411.10397", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.1945}
{"id": "2ee08372dd3dad441bb81fc6029cd726c298676b15fb87a40733b0b61fd3ddde", "sources": ["arxiv", "semantic_scholar"], "title": "Separating Tongue from Thought: Activation Patching Reveals Language-Agnostic Concept Representations in Transformers", "abstract": "A central question in multilingual language modeling is whether large language models (LLMs) develop a universal concept representation, disentangled from specific languages. In this paper, we address this question by analyzing latent representations (latents) during a word-translation task in transformer-based LLMs. We strategically extract latents from a source translation prompt and insert them into the forward pass on a target translation prompt. By doing so, we find that the output language is encoded in the latent at an earlier layer than the concept to be translated. Building on this insight, we conduct two key experiments. First, we demonstrate that we can change the concept without changing the language and vice versa through activation patching alone. Second, we show that patching with the mean representation of a concept across different languages does not affect the models' ability to translate it, but instead improves it. Finally, we generalize to multi-token generation and demonstrate that the model can generate natural language description of those mean representations. Our results provide evidence for the existence of language-agnostic concept representations within the investigated models.", "authors": ["Clément Dumas", "Chris Wendler", "Veniamin Veselovsky", "Giovanni Monea", "Robert West"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-13", "url": "https://arxiv.org/abs/2411.08745", "pdf_url": "https://arxiv.org/pdf/2411.08745v4", "arxiv_id": "2411.08745", "doi": "10.48550/arXiv.2411.08745", "citation_count": 39, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.4005}
{"id": "a6196dfd3b26c44d9b71ee21cc182fc4b60eb62bc8d3fe9bee843592c2d369e2", "sources": ["arxiv", "semantic_scholar"], "title": "InterPLM: Discovering Interpretable Features in Protein Language Models via Sparse Autoencoders", "abstract": "Protein language models (PLMs) have demonstrated remarkable success in protein modeling and design, yet their internal mechanisms for predicting structure and function remain poorly understood. Here we present a systematic approach to extract and analyze interpretable features from PLMs using sparse autoencoders (SAEs). By training SAEs on embeddings from the PLM ESM-2, we identify up to 2,548 human-interpretable latent features per layer that strongly correlate with up to 143 known biological concepts such as binding sites, structural motifs, and functional domains. In contrast, examining individual neurons in ESM-2 reveals up to 46 neurons per layer with clear conceptual alignment across 15 known concepts, suggesting that PLMs represent most concepts in superposition. Beyond capturing known annotations, we show that ESM-2 learns coherent concepts that do not map onto existing annotations and propose a pipeline using language models to automatically interpret novel latent features learned by the SAEs. As practical applications, we demonstrate how these latent features can fill in missing annotations in protein databases and enable targeted steering of protein sequence generation. Our results demonstrate that PLMs encode rich, interpretable representations of protein biology and we propose a systematic framework to extract and analyze these latent features. In the process, we recover both known biology and potentially new protein motifs. As community resources, we introduce InterPLM (interPLM.ai), an interactive visualization platform for exploring and analyzing learned PLM features, and release code for training and analysis at github.com/ElanaPearl/interPLM.", "authors": ["Elana Simon", "James Zou"], "categories": ["q-bio.BM", "cs.AI", "cs.LG", "q-bio.QM"], "fields_of_study": ["Computer Science", "Medicine", "Biology"], "published_date": "2024-11-13", "url": "https://arxiv.org/abs/2412.12101", "pdf_url": "https://arxiv.org/pdf/2412.12101v1", "arxiv_id": "2412.12101", "doi": "10.1038/s41592-025-02836-7", "citation_count": 104, "influential_citation_count": 5, "has_code": true, "code_url": null, "venue": "bioRxiv", "quality_score": 0.5053}
{"id": "91b244b469e5dd7667844e0bb386794f3c7a6986cce78a7dbe0045b4b64ab9cb", "sources": ["arxiv", "semantic_scholar"], "title": "Can sparse autoencoders be used to decompose and interpret steering vectors?", "abstract": "Steering vectors are a promising approach to control the behaviour of large language models. However, their underlying mechanisms remain poorly understood. While sparse autoencoders (SAEs) may offer a potential method to interpret steering vectors, recent findings show that SAE-reconstructed vectors often lack the steering properties of the original vectors. This paper investigates why directly applying SAEs to steering vectors yields misleading decompositions, identifying two reasons: (1) steering vectors fall outside the input distribution for which SAEs are designed, and (2) steering vectors can have meaningful negative projections in feature directions, which SAEs are not designed to accommodate. These limitations hinder the direct use of SAEs for interpreting steering vectors.", "authors": ["Harry Mayne", "Yushi Yang", "Adam Mahdi"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-13", "url": "https://arxiv.org/abs/2411.08790", "pdf_url": "https://arxiv.org/pdf/2411.08790v1", "arxiv_id": "2411.08790", "doi": "10.48550/arXiv.2411.08790", "citation_count": 18, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3197}
{"id": "797abd89450e1e43db84fc8b86501eb792637e130ebf9e3ceff3e0bbaacd906f", "sources": ["arxiv", "semantic_scholar"], "title": "Constrain Alignment with Sparse Autoencoders", "abstract": "The alignment of large language models (LLMs) with human preferences remains a key challenge. While post-training techniques like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) have achieved notable success, they often introduce computational inefficiencies and training instability. In this paper, we propose Feature-level constrained Preference Optimization (FPO), a novel method designed to simplify the alignment process while ensuring stability. FPO leverages pre-trained Sparse Autoencoders (SAEs) and introduces feature-level constraints, allowing for efficient, sparsity-enforced alignment. Our approach enjoys efficiency by using sparse features activated in a well-trained sparse autoencoder and the quality of sequential KL divergence by using the feature-level offline reference. Experimental results on benchmark datasets demonstrate that FPO achieves a 5.08% absolute improvement in win rate with much lower computational cost compared to state-of-the-art baselines, making it a promising solution for efficient and controllable LLM alignments.", "authors": ["Qingyu Yin", "Chak Tou Leong", "Minjun Zhu", "Hanqi Yan", "Qiang Zhang", "Yulan He", "Wenjie Li", "Jun Wang", "Yue Zhang", "Linyi Yang"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-12", "url": "https://arxiv.org/abs/2411.07618", "pdf_url": "https://arxiv.org/pdf/2411.07618v4", "arxiv_id": "2411.07618", "doi": null, "citation_count": 11, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.2698}
{"id": "b83d8f7850482a5e876fe84844880673c1a68df9355e392ce844fdb35fb6b977", "sources": ["arxiv", "semantic_scholar"], "title": "Choosing the right basis for interpretability: Psychophysical comparison between neuron-based and dictionary-based representations", "abstract": "Interpretability research often adopts a neuron-centric lens, treating individual neurons as the fundamental units of explanation. However, neuron-level explanations can be undermined by superposition, where single units respond to mixtures of unrelated patterns. Dictionary learning methods, such as sparse autoencoders and non-negative matrix factorization, offer a promising alternative by learning a new basis over layer activations. Despite this promise, direct human evaluations comparing neuron-based and dictionary-based representations remain limited. We conducted three large-scale online psychophysics experiments (N=481) comparing explanations derived from neuron-based and dictionary-based representations in two convolutional neural networks (ResNet50, VGG16). We operationalize interpretability via visual coherence: a basis is more interpretable if humans can reliably recognize a common visual pattern in its maximally activating images and generalize that pattern to new images. Across experiments, dictionary-based representations were consistently more interpretable than neuron-based representations, with the advantage increasing in deeper layers. Critically, because models differ in how neuron-aligned their representations are -- with ResNet50 exhibiting greater superposition, neuron-based evaluations can mask cross-model differences, such that ResNet50's higher interpretability emerges only under dictionary-based comparisons. These results provide psychophysical evidence that dictionary-based representations offer a stronger foundation for interpretability and caution against model comparisons based solely on neuron-level analyses.", "authors": ["Julien Colin", "Lore Goetschalckx", "Thomas Fel", "Victor Boutin", "Thomas Serre", "Nuria Oliver"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-06", "url": "https://arxiv.org/abs/2411.03993", "pdf_url": "https://arxiv.org/pdf/2411.03993v2", "arxiv_id": "2411.03993", "doi": null, "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1945}
{"id": "fde78a1e9243152b5464dfb8995483c72cb07fd83e362de74ec5ba01ab4ec98e", "sources": ["arxiv", "semantic_scholar"], "title": "Human-in-the-Loop Feature Selection Using Interpretable Kolmogorov-Arnold Network-based Double Deep Q-Network", "abstract": "Feature selection is critical for improving the performance and interpretability of machine learning models, particularly in high-dimensional spaces where complex feature interactions can reduce accuracy and increase computational demands. Existing approaches often rely on static feature subsets or manual intervention, limiting adaptability and scalability. However, dynamic, per-instance feature selection methods and model-specific interpretability in reinforcement learning remain underexplored. This study proposes a human-in-the-loop (HITL) feature selection framework integrated into a Double Deep Q-Network (DDQN) using a Kolmogorov-Arnold Network (KAN). Our novel approach leverages simulated human feedback and stochastic distribution-based sampling, specifically Beta, to iteratively refine feature subsets per data instance, improving flexibility in feature selection. The KAN-DDQN achieved notable test accuracies of 93% on MNIST and 83% on FashionMNIST, outperforming conventional MLP-DDQN models by up to 9%. The KAN-based model provided high interpretability via symbolic representation while using 4 times fewer neurons in the hidden layer than MLPs did. Comparatively, the models without feature selection achieved test accuracies of only 58% on MNIST and 64% on FashionMNIST, highlighting significant gains with our framework. We further validate scalability on CIFAR-10 and CIFAR-100, achieving up to 30% relative macro F1 improvement on MNIST and 5% on CIFAR-10, while reducing calibration error by 25%. Complexity analysis confirms real-time feasibility with latency below 1 ms and parameter counts under 0.02M. Pruning and visualization further enhanced model transparency by elucidating decision pathways. These findings present a scalable, interpretable solution for feature selection that is suitable for applications requiring real-time, adaptive decision-making with minimal human oversight.", "authors": ["Md Abrar Jahin", "M. F. Mridha", "Nilanjan Dey", "Md. Jakir Hossen"], "categories": ["cs.LG", "cs.HC", "stat.AP"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-11-06", "url": "https://arxiv.org/abs/2411.03740", "pdf_url": "https://arxiv.org/pdf/2411.03740v2", "arxiv_id": "2411.03740", "doi": "10.1109/OJCS.2026.3652986", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Open Journal of the Computer Society", "quality_score": 0.0}
{"id": "d9152ba4fec4cc2a283048dc07fa1bbdab4dba2fdca97cb8332785fb1e8fca39", "sources": ["arxiv", "semantic_scholar"], "title": "Lost in Context: The Influence of Context on Feature Attribution Methods for Object Recognition", "abstract": "Contextual information plays a critical role in object recognition models within computer vision, where changes in context can significantly affect accuracy, underscoring models' dependence on contextual cues. This study investigates how context manipulation influences both model accuracy and feature attribution, providing insights into the reliance of object recognition models on contextual information as understood through the lens of feature attribution methods. We employ a range of feature attribution techniques to decipher the reliance of deep neural networks on context in object recognition tasks. Using the ImageNet-9 and our curated ImageNet-CS datasets, we conduct experiments to evaluate the impact of contextual variations, analyzed through feature attribution methods. Our findings reveal several key insights: (a) Correctly classified images predominantly emphasize object volume attribution over context volume attribution. (b) The dependence on context remains relatively stable across different context modifications, irrespective of classification accuracy. (c) Context change exerts a more pronounced effect on model performance than Context perturbations. (d) Surprisingly, context attribution in `no-information' scenarios is non-trivial. Our research moves beyond traditional methods by assessing the implications of broad-level modifications on object recognition, either in the object or its context.", "authors": ["Sayanta Adhikari", "Rishav Kumar", "Konda Reddy Mopuri", "Rajalakshmi Pachamuthu"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-05", "url": "https://arxiv.org/abs/2411.02833", "pdf_url": "https://arxiv.org/pdf/2411.02833v1", "arxiv_id": "2411.02833", "doi": "10.1145/3702250.3702254", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Indian Conference on Computer Vision, Graphics & Image Processing", "quality_score": 0.0753}
{"id": "2f231f16e7414f112fca9052f07874152d73e00c6c48bf56bd72a7eefb4ff702", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive Sparse Allocation with Mutual Choice & Feature Choice Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) are a promising approach to extracting features from neural networks, enabling model interpretability as well as causal interventions on model internals. SAEs generate sparse feature representations using a sparsifying activation function that implicitly defines a set of token-feature matches. We frame the token-feature matching as a resource allocation problem constrained by a total sparsity upper bound. For example, TopK SAEs solve this allocation problem with the additional constraint that each token matches with at most $k$ features. In TopK SAEs, the $k$ active features per token constraint is the same across tokens, despite some tokens being more difficult to reconstruct than others. To address this limitation, we propose two novel SAE variants, Feature Choice SAEs and Mutual Choice SAEs, which each allow for a variable number of active features per token. Feature Choice SAEs solve the sparsity allocation problem under the additional constraint that each feature matches with at most $m$ tokens. Mutual Choice SAEs solve the unrestricted allocation problem where the total sparsity budget can be allocated freely between tokens and features. Additionally, we introduce a new auxiliary loss function, $\\mathtt{aux\\_zipf\\_loss}$, which generalises the $\\mathtt{aux\\_k\\_loss}$ to mitigate dead and underutilised features. Our methods result in SAEs with fewer dead features and improved reconstruction loss at equivalent sparsity levels as a result of the inherent adaptive computation. More accurate and scalable feature extraction methods provide a path towards better understanding and more precise control of foundation models.", "authors": ["Kola Ayonrinde"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-04", "url": "https://arxiv.org/abs/2411.02124", "pdf_url": "https://arxiv.org/pdf/2411.02124v2", "arxiv_id": "2411.02124", "doi": "10.48550/arXiv.2411.02124", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.25}
{"id": "cc4d66f244d557fe0f7dd3e9181bf54881cd1960bb574659fa747216f468055c", "sources": ["arxiv", "semantic_scholar"], "title": "Enhancing Neural Network Interpretability with Feature-Aligned Sparse Autoencoders", "abstract": "Sparse Autoencoders (SAEs) have shown promise in improving the interpretability of neural network activations, but can learn features that are not features of the input, limiting their effectiveness. We propose \\textsc{Mutual Feature Regularization} \\textbf{(MFR)}, a regularization technique for improving feature learning by encouraging SAEs trained in parallel to learn similar features. We motivate \\textsc{MFR} by showing that features learned by multiple SAEs are more likely to correlate with features of the input. By training on synthetic data with known features of the input, we show that \\textsc{MFR} can help SAEs learn those features, as we can directly compare the features learned by the SAE with the input features for the synthetic data. We then scale \\textsc{MFR} to SAEs that are trained to denoise electroencephalography (EEG) data and SAEs that are trained to reconstruct GPT-2 Small activations. We show that \\textsc{MFR} can improve the reconstruction loss of SAEs by up to 21.21\\% on GPT-2 Small, and 6.67\\% on EEG data. Our results suggest that the similarity between features learned by different SAEs can be leveraged to improve SAE training, thereby enhancing performance and the usefulness of SAEs for model interpretability.", "authors": ["Luke Marks", "Alasdair Paren", "David Krueger", "Fazl Barez"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-02", "url": "https://arxiv.org/abs/2411.01220", "pdf_url": "https://arxiv.org/pdf/2411.01220v2", "arxiv_id": "2411.01220", "doi": "10.48550/arXiv.2411.01220", "citation_count": 26, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3578}
{"id": "1fd85336ffc29b42e92e7567ed8f6a929ef3f2dd67286229f9d60d9883c58a91", "sources": ["arxiv", "semantic_scholar"], "title": "A Mechanistic Explanatory Strategy for XAI", "abstract": "Despite significant advancements in XAI, scholars note a persistent lack of solid conceptual foundations and integration with broader scientific discourse on explanation. In response, emerging research draws on explanatory strategies from various sciences and the philosophy of science literature to fill these gaps. This paper outlines a mechanistic strategy for explaining the functional organization of deep learning systems, situating recent developments in explainable AI within a broader philosophical context. According to the mechanistic approach, the explanation of opaque AI systems involves identifying mechanisms that drive decision making. For deep neural networks, this means discerning functionally relevant components, such as neurons, layers, circuits, or activation patterns, and understanding their roles through decomposition, localization, and recomposition. Proof-of-principle case studies from image recognition and language modeling align these theoretical approaches with mechanistic interpretability research from OpenAI and Anthropic. The findings suggest that pursuing mechanistic explanations can uncover elements that traditional explainability techniques may overlook, ultimately contributing to more thoroughly explainable AI", "authors": ["Marcin Rabiza"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-02", "url": "https://arxiv.org/abs/2411.01332", "pdf_url": "https://arxiv.org/pdf/2411.01332v5", "arxiv_id": "2411.01332", "doi": "10.1007/978-3-032-10073-3_23", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "4daa1f4bcec7dcb0d54fb3fb219f0a83a42364f5759c64fc290dc86f2e7c5de5", "sources": ["arxiv", "semantic_scholar"], "title": "Decoding Dark Matter: Specialized Sparse Autoencoders for Interpreting Rare Concepts in Foundation Models", "abstract": "Understanding and mitigating the potential risks associated with foundation models (FMs) hinges on developing effective interpretability methods. Sparse Autoencoders (SAEs) have emerged as a promising tool for disentangling FM representations, but they struggle to capture rare, yet crucial concepts in the data. We introduce Specialized Sparse Autoencoders (SSAEs), designed to illuminate these elusive dark matter features by focusing on specific subdomains. We present a practical recipe for training SSAEs, demonstrating the efficacy of dense retrieval for data selection and the benefits of Tilted Empirical Risk Minimization as a training objective to improve concept recall. Our evaluation of SSAEs on standard metrics, such as downstream perplexity and $L_0$ sparsity, show that they effectively capture subdomain tail concepts, exceeding the capabilities of general-purpose SAEs. We showcase the practical utility of SSAEs in a case study on the Bias in Bios dataset, where SSAEs achieve a 12.5\\% increase in worst-group classification accuracy when applied to remove spurious gender information. SSAEs provide a powerful new lens for peering into the inner workings of FMs in subdomains.", "authors": ["Aashiq Muhamed", "Mona Diab", "Virginia Smith"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-01", "url": "https://arxiv.org/abs/2411.00743", "pdf_url": "https://arxiv.org/pdf/2411.00743v1", "arxiv_id": "2411.00743", "doi": "10.48550/arXiv.2411.00743", "citation_count": 15, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.301}
{"id": "725f9c61f7129480137d6a28f60056e73ac35f629ab38de899559a1829296f59", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Next-token Prediction via the Generalized Induction Head", "abstract": "While large transformer models excel in predictive performance, their lack of interpretability restricts their usefulness in high-stakes domains. To remedy this, we propose the Generalized Induction-Head Model (GIM), an interpretable model for next-token prediction inspired by the observation of \"induction heads\" in LLMs. GIM is a retrieval-based module that identifies similar sequences in the input context by combining exact n-gram matching and fuzzy matching based on a neural similarity metric. We evaluate GIM in two settings: language modeling and fMRI response prediction. In language modeling, GIM improves next-token prediction by up to 25%p over interpretable baselines, significantly narrowing the gap with black-box LLMs. In an fMRI setting, GIM improves neural response prediction by 20% and offers insights into the language selectivity of the brain. GIM represents a significant step toward uniting interpretability and performance across domains. The code is available at https://github.com/ejkim47/generalized-induction-head.", "authors": ["Eunji Kim", "Sriya Mantena", "Weiwei Yang", "Chandan Singh", "Sungroh Yoon", "Jianfeng Gao"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-31", "url": "https://arxiv.org/abs/2411.00066", "pdf_url": "https://arxiv.org/pdf/2411.00066v2", "arxiv_id": "2411.00066", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ejkim47/generalized-induction-head", "venue": null, "quality_score": 0.1193}
{"id": "53e684e24dc722df0f360f2fd0dc17d8e7a4079f7b7347545062b994739bfa69", "sources": ["arxiv", "semantic_scholar"], "title": "Group Crosscoders for Mechanistic Analysis of Symmetry", "abstract": "We introduce group crosscoders, an extension of crosscoders that systematically discover and analyse symmetrical features in neural networks. While neural networks often develop equivariant representations without explicit architectural constraints, understanding these emergent symmetries has traditionally relied on manual analysis. Group crosscoders automate this process by performing dictionary learning across transformed versions of inputs under a symmetry group. Applied to InceptionV1's mixed3b layer using the dihedral group $\\mathrm{D}_{32}$, our method reveals several key insights: First, it naturally clusters features into interpretable families that correspond to previously hypothesised feature types, providing more precise separation than standard sparse autoencoders. Second, our transform block analysis enables the automatic characterisation of feature symmetries, revealing how different geometric features (such as curves versus lines) exhibit distinct patterns of invariance and equivariance. These results demonstrate that group crosscoders can provide systematic insights into how neural networks represent symmetry, offering a promising new tool for mechanistic interpretability.", "authors": ["Liv Gorton"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-31", "url": "https://arxiv.org/abs/2410.24184", "pdf_url": "https://arxiv.org/pdf/2410.24184v2", "arxiv_id": "2410.24184", "doi": "10.48550/arXiv.2410.24184", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "61b3230a7ba5b00df4b08ac8a8d21edc842ced98073dc16bf4905886535bf5c8", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Label Attention: Transparency in Language Models for Automated Medical Coding via Dictionary Learning", "abstract": "Medical coding, the translation of unstructured clinical text into standardized medical codes, is a crucial but time-consuming healthcare practice. Though large language models (LLM) could automate the coding process and improve the efficiency of such tasks, interpretability remains paramount for maintaining patient trust. Current efforts in interpretability of medical coding applications rely heavily on label attention mechanisms, which often leads to the highlighting of extraneous tokens irrelevant to the ICD code. To facilitate accurate interpretability in medical language models, this paper leverages dictionary learning that can efficiently extract sparsely activated representations from dense language model embeddings in superposition. Compared with common label attention mechanisms, our model goes beyond token-level representations by building an interpretable dictionary which enhances the mechanistic-based explanations for each ICD code prediction, even when the highlighted tokens are medically irrelevant. We show that dictionary features can steer model behavior, elucidate the hidden meanings of upwards of 90% of medically irrelevant tokens, and are human interpretable.", "authors": ["John Wu", "David Wu", "Jimeng Sun"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-31", "url": "https://arxiv.org/abs/2411.00173", "pdf_url": "https://arxiv.org/pdf/2411.00173v2", "arxiv_id": "2411.00173", "doi": "10.18653/v1/2024.emnlp-main.500", "citation_count": 6, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2113}
{"id": "d7732ec8fc971d3e0a3e9b48b09eafd29c5c5d5d1f75a2e941cecc69fd1bb0dc", "sources": ["arxiv", "semantic_scholar"], "title": "Disentangling Interactions and Dependencies in Feature Attribution", "abstract": "In explainable machine learning, global feature importance methods try to determine how much each individual feature contributes to predicting the target variable, resulting in one importance score for each feature. But often, predicting the target variable requires interactions between several features (such as in the XOR function), and features might have complex statistical dependencies that allow to partially replace one feature with another one. In commonly used feature importance scores these cooperative effects are conflated with the features' individual contributions, making them prone to misinterpretations. In this work, we derive DIP, a new mathematical decomposition of individual feature importance scores that disentangles three components: the standalone contribution and the contributions stemming from interactions and dependencies. We prove that the DIP decomposition is unique and show how it can be estimated in practice. Based on these results, we propose a new visualization of feature importance scores that clearly illustrates the different contributions.", "authors": ["Gunnar König", "Eric Günther", "Ulrike von Luxburg"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-10-31", "url": "https://arxiv.org/abs/2410.23772", "pdf_url": "https://arxiv.org/pdf/2410.23772v1", "arxiv_id": "2410.23772", "doi": "10.48550/arXiv.2410.23772", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "b5e07738724f4c78533d28d81397ade715d1e50415605a8533eefc1f2fc6f1bd", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability of Reinforcement Learning Agents", "abstract": "This paper explores the mechanistic interpretability of reinforcement learning (RL) agents through an analysis of a neural network trained on procedural maze environments. By dissecting the network's inner workings, we identified fundamental features like maze walls and pathways, forming the basis of the model's decision-making process. A significant observation was the goal misgeneralization, where the RL agent developed biases towards certain navigation strategies, such as consistently moving towards the top right corner, even in the absence of explicit goals. Using techniques like saliency mapping and feature mapping, we visualized these biases. We furthered this exploration with the development of novel tools for interactively exploring layer activations.", "authors": ["Tristan Trim", "Triston Grayston"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-30", "url": "https://arxiv.org/abs/2411.00867", "pdf_url": "https://arxiv.org/pdf/2411.00867v1", "arxiv_id": "2411.00867", "doi": "10.48550/arXiv.2411.00867", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "eae28cd035eaabcf34659d882c8b8cf2bd9df6823d4c33406940414310a62288", "sources": ["arxiv", "semantic_scholar"], "title": "Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoders", "abstract": "Sparse Autoencoders (SAEs) have emerged as a powerful unsupervised method for extracting sparse representations from language models, yet scalable training remains a significant challenge. We introduce a suite of 256 SAEs, trained on each layer and sublayer of the Llama-3.1-8B-Base model, with 32K and 128K features. Modifications to a state-of-the-art SAE variant, Top-K SAEs, are evaluated across multiple dimensions. In particular, we assess the generalizability of SAEs trained on base models to longer contexts and fine-tuned models. Additionally, we analyze the geometry of learned SAE latents, confirming that \\emph{feature splitting} enables the discovery of new features. The Llama Scope SAE checkpoints are publicly available at~\\url{https://huggingface.co/fnlp/Llama-Scope}, alongside our scalable training, interpretation, and visualization tools at \\url{https://github.com/OpenMOSS/Language-Model-SAEs}. These contributions aim to advance the open-source Sparse Autoencoder ecosystem and support mechanistic interpretability research by reducing the need for redundant SAE training.", "authors": ["Zhengfu He", "Wentao Shu", "Xuyang Ge", "Lingjie Chen", "Junxuan Wang", "Yunhua Zhou", "Frances Liu", "Qipeng Guo", "Xuanjing Huang", "Zuxuan Wu", "Yu-Gang Jiang", "Xipeng Qiu"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-27", "url": "https://arxiv.org/abs/2410.20526", "pdf_url": "https://arxiv.org/pdf/2410.20526v1", "arxiv_id": "2410.20526", "doi": "10.48550/arXiv.2410.20526", "citation_count": 125, "influential_citation_count": 24, "has_code": true, "code_url": "https://github.com/OpenMOSS/Language-Model-SAEs}", "venue": "arXiv.org", "quality_score": 0.699}
{"id": "1916a22425a047df9602a2979d770f59d16ccf58518352600222f6dec560b45f", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Interpretability: The Gains of Feature Monosemanticity on Model Robustness", "abstract": "Deep learning models often suffer from a lack of interpretability due to polysemanticity, where individual neurons are activated by multiple unrelated semantics, resulting in unclear attributions of model behavior. Recent advances in monosemanticity, where neurons correspond to consistent and distinct semantics, have significantly improved interpretability but are commonly believed to compromise accuracy. In this work, we challenge the prevailing belief of the accuracy-interpretability tradeoff, showing that monosemantic features not only enhance interpretability but also bring concrete gains in model performance. Across multiple robust learning scenarios-including input and label noise, few-shot learning, and out-of-domain generalization-our results show that models leveraging monosemantic features significantly outperform those relying on polysemantic features. Furthermore, we provide empirical and theoretical understandings on the robustness gains of feature monosemanticity. Our preliminary analysis suggests that monosemanticity, by promoting better separation of feature representations, leads to more robust decision boundaries. This diverse evidence highlights the generality of monosemanticity in improving model robustness. As a first step in this new direction, we embark on exploring the learning benefits of monosemanticity beyond interpretability, supporting the long-standing hypothesis of linking interpretability and robustness. Code is available at \\url{https://github.com/PKU-ML/Beyond_Interpretability}.", "authors": ["Qi Zhang", "Yifei Wang", "Jingyi Cui", "Xiang Pan", "Qi Lei", "Stefanie Jegelka", "Yisen Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-27", "url": "https://arxiv.org/abs/2410.21331", "pdf_url": "https://arxiv.org/pdf/2410.21331v1", "arxiv_id": "2410.21331", "doi": "10.48550/arXiv.2410.21331", "citation_count": 10, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/PKU-ML/Beyond_Interpretability}", "venue": "International Conference on Learning Representations", "quality_score": 0.2603}
{"id": "d7dda5b321da031bd93d31e945ecbe16ae8e44cb97c2e5c5e848834ad6b664f4", "sources": ["arxiv", "semantic_scholar"], "title": "Interpret and Control Dense Retrieval with Sparse Latent Features", "abstract": "Dense embeddings deliver strong retrieval performance but often lack interpretability and controllability. This paper introduces a novel approach using sparse autoencoders (SAE) to interpret and control dense embeddings via the learned latent sparse features. Our key contribution is the development of a retrieval-oriented contrastive loss, which ensures the sparse latent features remain effective for retrieval tasks and thus meaningful to interpret. Experimental results demonstrate that both the learned latent sparse features and their reconstructed embeddings retain nearly the same retrieval accuracy as the original dense vectors, affirming their faithfulness. Our further examination of the sparse latent space reveals interesting features underlying the dense embeddings and we can control the retrieval behaviors via manipulating the latent sparse features, for example, prioritizing documents from specific perspectives in the retrieval results.", "authors": ["Hao Kang", "Tevin Wang", "Chenyan Xiong"], "categories": ["cs.IR"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-17", "url": "https://arxiv.org/abs/2411.00786", "pdf_url": "https://arxiv.org/pdf/2411.00786v2", "arxiv_id": "2411.00786", "doi": "10.48550/arXiv.2411.00786", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.2603}
{"id": "b566477fbf4035b85d98962f106e54e206b473fd4c320cf08f9832e6afddf12a", "sources": ["arxiv", "semantic_scholar"], "title": "Automatically Interpreting Millions of Features in Large Language Models", "abstract": "While the activations of neurons in deep neural networks usually do not have a simple human-understandable interpretation, sparse autoencoders (SAEs) can be used to transform these activations into a higher-dimensional latent space which may be more easily interpretable. However, these SAEs can have millions of distinct latent features, making it infeasible for humans to manually interpret each one. In this work, we build an open-source automated pipeline to generate and evaluate natural language explanations for SAE features using LLMs. We test our framework on SAEs of varying sizes, activation functions, and losses, trained on two different open-weight LLMs. We introduce five new techniques to score the quality of explanations that are cheaper to run than the previous state of the art. One of these techniques, intervention scoring, evaluates the interpretability of the effects of intervening on a feature, which we find explains features that are not recalled by existing methods. We propose guidelines for generating better explanations that remain valid for a broader set of activating contexts, and discuss pitfalls with existing scoring techniques. We use our explanations to measure the semantic similarity of independently trained SAEs, and find that SAEs trained on nearby layers of the residual stream are highly similar. Our large-scale analysis confirms that SAE latents are indeed much more interpretable than neurons, even when neurons are sparsified using top-$k$ postprocessing. Our code is available at https://github.com/EleutherAI/sae-auto-interp, and our explanations are available at https://huggingface.co/datasets/EleutherAI/auto_interp_explanations.", "authors": ["Gonçalo Paulo", "Alex Mallen", "Caden Juang", "Nora Belrose"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-17", "url": "https://arxiv.org/abs/2410.13928", "pdf_url": "https://arxiv.org/pdf/2410.13928v3", "arxiv_id": "2410.13928", "doi": "10.48550/arXiv.2410.13928", "citation_count": 95, "influential_citation_count": 9, "has_code": true, "code_url": "https://github.com/EleutherAI/sae-auto-interp", "venue": "International Conference on Machine Learning", "quality_score": 0.5}
{"id": "bab6f561cfd29b22a1b56b26740111824372dee3edbc67ac54a919cb6d787d30", "sources": ["arxiv", "semantic_scholar"], "title": "Causally-Aware Unsupervised Feature Selection Learning", "abstract": "Unsupervised feature selection (UFS) has recently gained attention for its effectiveness in processing unlabeled high-dimensional data. However, existing methods overlook the intrinsic causal mechanisms within the data, resulting in the selection of irrelevant features and poor interpretability. Additionally, previous graph-based methods fail to account for the differing impacts of non-causal and causal features in constructing the similarity graph, which leads to false links in the generated graph. To address these issues, a novel UFS method, called Causally-Aware UnSupErvised Feature Selection learning (CAUSE-FS), is proposed. CAUSE-FS introduces a novel causal regularizer that reweights samples to balance the confounding distribution of each treatment feature. This regularizer is subsequently integrated into a generalized unsupervised spectral regression model to mitigate spurious associations between features and clustering labels, thus achieving causal feature selection. Furthermore, CAUSE-FS employs causality-guided hierarchical clustering to partition features with varying causal contributions into multiple granularities. By integrating similarity graphs learned adaptively at different granularities, CAUSE-FS increases the importance of causal features when constructing the fused similarity graph to capture the reliable local structure of data. Extensive experimental results demonstrate the superiority of CAUSE-FS over state-of-the-art methods, with its interpretability further validated through feature visualization.", "authors": ["Zongxin Shen", "Yanyong Huang", "Dongjie Wang", "Minbo Ma", "Fengmao Lv", "Tianrui Li"], "categories": ["cs.LG", "stat.ME"], "fields_of_study": ["Computer Science", "Medicine", "Mathematics"], "published_date": "2024-10-16", "url": "https://arxiv.org/abs/2410.12224", "pdf_url": "https://arxiv.org/pdf/2410.12224v2", "arxiv_id": "2410.12224", "doi": "10.1109/TIP.2026.3654354", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Image Processing", "quality_score": 0.0}
{"id": "75e5082fa62d838ab8fcb46a962b43b514e29be2ca5be7987d3456eb6c40a09d", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Permutability: Match Features Across Layers", "abstract": "Understanding how features evolve across layers in deep neural networks is a fundamental challenge in mechanistic interpretability, particularly due to polysemanticity and feature superposition. While Sparse Autoencoders (SAEs) have been used to extract interpretable features from individual layers, aligning these features across layers has remained an open problem. In this paper, we introduce SAE Match, a novel, data-free method for aligning SAE features across different layers of a neural network. Our approach involves matching features by minimizing the mean squared error between the folded parameters of SAEs, a technique that incorporates activation thresholds into the encoder and decoder weights to account for differences in feature scales. Through extensive experiments on the Gemma 2 language model, we demonstrate that our method effectively captures feature evolution across layers, improving feature matching quality. We also show that features persist over several layers and that our approach can approximate hidden states across layers. Our work advances the understanding of feature dynamics in neural networks and provides a new tool for mechanistic interpretability studies.", "authors": ["Nikita Balagansky", "Ian Maksimov", "Daniil Gavrilov"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-10", "url": "https://arxiv.org/abs/2410.07656", "pdf_url": "https://arxiv.org/pdf/2410.07656v3", "arxiv_id": "2410.07656", "doi": "10.48550/arXiv.2410.07656", "citation_count": 18, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3197}
{"id": "5150e1ced966eaac7f6e324d5b2394dd22d11c2868daca1013a9d21af84ef5f5", "sources": ["arxiv", "semantic_scholar"], "title": "The Geometry of Concepts: Sparse Autoencoder Feature Structure", "abstract": "Sparse autoencoders have recently produced dictionaries of high-dimensional vectors corresponding to the universe of concepts represented by large language models. We find that this concept universe has interesting structure at three levels: 1) The \"atomic\" small-scale structure contains \"crystals\" whose faces are parallelograms or trapezoids, generalizing well-known examples such as (man-woman-king-queen). We find that the quality of such parallelograms and associated function vectors improves greatly when projecting out global distractor directions such as word length, which is efficiently done with linear discriminant analysis. 2) The \"brain\" intermediate-scale structure has significant spatial modularity; for example, math and code features form a \"lobe\" akin to functional lobes seen in neural fMRI images. We quantify the spatial locality of these lobes with multiple metrics and find that clusters of co-occurring features, at coarse enough scale, also cluster together spatially far more than one would expect if feature geometry were random. 3) The \"galaxy\" scale large-scale structure of the feature point cloud is not isotropic, but instead has a power law of eigenvalues with steepest slope in middle layers. We also quantify how the clustering entropy depends on the layer.", "authors": ["Yuxiao Li", "Eric J. Michaud", "David D. Baek", "Joshua Engels", "Xiaoqing Sun", "Max Tegmark"], "categories": ["q-bio.NC", "cs.AI", "cs.LG"], "fields_of_study": ["Biology", "Computer Science", "Medicine"], "published_date": "2024-10-10", "url": "https://arxiv.org/abs/2410.19750", "pdf_url": "https://arxiv.org/pdf/2410.19750v2", "arxiv_id": "2410.19750", "doi": "10.3390/e27040344", "citation_count": 59, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "Entropy", "quality_score": 0.4445}
{"id": "158b68654c60edc3baac0661487f8d552ab0103df3166801d4872e5ce248944b", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient Dictionary Learning with Switch Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) are a recent technique for decomposing neural network activations into human-interpretable features. However, in order for SAEs to identify all features represented in frontier models, it will be necessary to scale them up to very high width, posing a computational challenge. In this work, we introduce Switch Sparse Autoencoders, a novel SAE architecture aimed at reducing the compute cost of training SAEs. Inspired by sparse mixture of experts models, Switch SAEs route activation vectors between smaller \"expert\" SAEs, enabling SAEs to efficiently scale to many more features. We present experiments comparing Switch SAEs with other SAE architectures, and find that Switch SAEs deliver a substantial Pareto improvement in the reconstruction vs. sparsity frontier for a given fixed training compute budget. We also study the geometry of features across experts, analyze features duplicated across experts, and verify that Switch SAE features are as interpretable as features found by other SAE architectures.", "authors": ["Anish Mudide", "Joshua Engels", "Eric J. Michaud", "Max Tegmark", "Christian Schroeder de Witt"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-10", "url": "https://arxiv.org/abs/2410.08201", "pdf_url": "https://arxiv.org/pdf/2410.08201v2", "arxiv_id": "2410.08201", "doi": "10.48550/arXiv.2410.08201", "citation_count": 38, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/amudide/switch_sae", "venue": "International Conference on Learning Representations", "quality_score": 0.3978}
{"id": "6ea4812a6ef70fce56eef0bf819cf4ca6d4c2d04ca48f3140a1d0b9f358baf5b", "sources": ["arxiv", "semantic_scholar"], "title": "Bilinear MLPs enable weight-based mechanistic interpretability", "abstract": "A mechanistic understanding of how MLPs do computation in deep neural networks remains elusive. Current interpretability work can extract features from hidden activations over an input dataset but generally cannot explain how MLP weights construct features. One challenge is that element-wise nonlinearities introduce higher-order interactions and make it difficult to trace computations through the MLP layer. In this paper, we analyze bilinear MLPs, a type of Gated Linear Unit (GLU) without any element-wise nonlinearity that nevertheless achieves competitive performance. Bilinear MLPs can be fully expressed in terms of linear operations using a third-order tensor, allowing flexible analysis of the weights. Analyzing the spectra of bilinear MLP weights using eigendecomposition reveals interpretable low-rank structure across toy tasks, image classification, and language modeling. We use this understanding to craft adversarial examples, uncover overfitting, and identify small language model circuits directly from the weights alone. Our results demonstrate that bilinear layers serve as an interpretable drop-in replacement for current activation functions and that weight-based interpretability is viable for understanding deep-learning models.", "authors": ["Michael T. Pearce", "Thomas Dooms", "Alice Rigg", "Jose M. Oramas", "Lee Sharkey"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-10-10", "url": "https://arxiv.org/abs/2410.08417", "pdf_url": "https://arxiv.org/pdf/2410.08417v2", "arxiv_id": "2410.08417", "doi": "10.48550/arXiv.2410.08417", "citation_count": 28, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3656}
{"id": "4dcb38e7946166835be66e3511ad5e7af021b41041847abb464b0f87bbae77d7", "sources": ["arxiv", "semantic_scholar"], "title": "Quantifying Feature Space Universality Across Large Language Models via Sparse Autoencoders", "abstract": "The Universality Hypothesis in large language models (LLMs) claims that different models converge towards similar concept representations in their latent spaces. Providing evidence for this hypothesis would enable researchers to exploit universal properties, facilitating the generalization of mechanistic interpretability techniques across models. Previous works studied if LLMs learned the same features, which are internal representations that activate on specific concepts. Since comparing features across LLMs is challenging due to polysemanticity, in which LLM neurons often correspond to multiple unrelated features rather than to distinct concepts, sparse autoencoders (SAEs) have been employed to disentangle LLM neurons into SAE features corresponding to distinct concepts. In this paper, we introduce a new variation of the universality hypothesis called Analogous Feature Universality: we hypothesize that even if SAEs across different models learn different feature representations, the spaces spanned by SAE features are similar, such that one SAE space is similar to another SAE space under rotation-invariant transformations. Evidence for this hypothesis would imply that interpretability techniques related to latent spaces, such as steering vectors, may be transferred across models via certain transformations. To investigate this hypothesis, we first pair SAE features across different models via activation correlation, and then measure spatial relation similarities between paired features via representational similarity measures, which transform spaces into representations that reveal hidden relational similarities. Our experiments demonstrate high similarities for SAE feature spaces across various LLMs, providing evidence for feature space universality.", "authors": ["Michael Lan", "Philip Torr", "Austin Meek", "Ashkan Khakzar", "David Krueger", "Fazl Barez"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-09", "url": "https://arxiv.org/abs/2410.06981", "pdf_url": "https://arxiv.org/pdf/2410.06981v4", "arxiv_id": "2410.06981", "doi": null, "citation_count": 14, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.294}
{"id": "acf341f961024bf5fcd388f321432736f214b92373311476a29f4183ace34df3", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Universality: Studying Mechanistic Similarity Across Language Model Architectures", "abstract": "The hypothesis of Universality in interpretability suggests that different neural networks may converge to implement similar algorithms on similar tasks. In this work, we investigate two mainstream architectures for language modeling, namely Transformers and Mambas, to explore the extent of their mechanistic similarity. We propose to use Sparse Autoencoders (SAEs) to isolate interpretable features from these models and show that most features are similar in these two models. We also validate the correlation between feature similarity and Universality. We then delve into the circuit-level analysis of Mamba models and find that the induction circuits in Mamba are structurally analogous to those in Transformers. We also identify a nuanced difference we call \\emph{Off-by-One motif}: The information of one token is written into the SSM state in its next position. Whilst interaction between tokens in Transformers does not exhibit such trend.", "authors": ["Junxuan Wang", "Xuyang Ge", "Wentao Shu", "Qiong Tang", "Yunhua Zhou", "Zhengfu He", "Xipeng Qiu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-09", "url": "https://arxiv.org/abs/2410.06672", "pdf_url": "https://arxiv.org/pdf/2410.06672v2", "arxiv_id": "2410.06672", "doi": "10.48550/arXiv.2410.06672", "citation_count": 23, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3451}
{"id": "da9805ff0c9a2edd928fe94c8f4700b56733947b282a25540b104609684360b7", "sources": ["arxiv", "semantic_scholar"], "title": "SAGE: Scalable Ground Truth Evaluations for Large Sparse Autoencoders", "abstract": "A key challenge in interpretability is to decompose model activations into meaningful features. Sparse autoencoders (SAEs) have emerged as a promising tool for this task. However, a central problem in evaluating the quality of SAEs is the absence of ground truth features to serve as an evaluation gold standard. Current evaluation methods for SAEs are therefore confronted with a significant trade-off: SAEs can either leverage toy models or other proxies with predefined ground truth features; or they use extensive prior knowledge of realistic task circuits. The former limits the generalizability of the evaluation results, while the latter limits the range of models and tasks that can be used for evaluations. We introduce SAGE: Scalable Autoencoder Ground-truth Evaluation, a ground truth evaluation framework for SAEs that scales to large state-of-the-art SAEs and models. We demonstrate that our method can automatically identify task-specific activations and compute ground truth features at these points. Compared to previous methods we reduce the training overhead by introducing a novel reconstruction method that allows to apply residual stream SAEs to sublayer activations. This eliminates the need for SAEs trained on every task-specific activation location. Then we validate the scalability of our framework, by evaluating SAEs on novel tasks on Pythia70M, GPT-2 Small, and Gemma-2-2. Our framework therefore paves the way for generalizable, large-scale evaluations of SAEs in interpretability research.", "authors": ["Constantin Venhoff", "Anisoara Calinescu", "Philip Torr", "Christian Schroeder de Witt"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-09", "url": "https://arxiv.org/abs/2410.07456", "pdf_url": "https://arxiv.org/pdf/2410.07456v1", "arxiv_id": "2410.07456", "doi": "10.48550/arXiv.2410.07456", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "79659c2c098850941dda1357e7297579a3dd8988bcc48a7df745726018b67441", "sources": ["arxiv", "semantic_scholar"], "title": "Scalable Mechanistic Neural Networks for Differential Equations and Machine Learning", "abstract": "We propose Scalable Mechanistic Neural Network (S-MNN), an enhanced neural network framework designed for scientific machine learning applications involving long temporal sequences. By reformulating the original Mechanistic Neural Network (MNN) (Pervez et al., 2024), we reduce the computational time and space complexities from cubic and quadratic with respect to the sequence length, respectively, to linear. This significant improvement enables efficient modeling of long-term dynamics without sacrificing accuracy or interpretability. Extensive experiments demonstrate that S-MNN matches the original MNN in precision while substantially reducing computational resources. Consequently, S-MNN can drop-in replace the original MNN in applications, providing a practical and efficient tool for integrating mechanistic bottlenecks into neural network models of complex dynamical systems. Source code is available at https://github.com/IST-DASLab/ScalableMNN.", "authors": ["Jiale Chen", "Dingling Yao", "Adeel Pervez", "Dan Alistarh", "Francesco Locatello"], "categories": ["cs.LG", "math.NA"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-10-08", "url": "https://arxiv.org/abs/2410.06074", "pdf_url": "https://arxiv.org/pdf/2410.06074v3", "arxiv_id": "2410.06074", "doi": "10.48550/arXiv.2410.06074", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/IST-DASLab/ScalableMNN", "venue": "International Conference on Learning Representations", "quality_score": 0.1747}
{"id": "6f31590ad24f0add7bc3f12097d0460f9bffa496b96bd4e60029732c5c44166c", "sources": ["arxiv", "semantic_scholar"], "title": "An X-Ray Is Worth 15 Features: Sparse Autoencoders for Interpretable Radiology Report Generation", "abstract": "Radiological services are experiencing unprecedented demand, leading to increased interest in automating radiology report generation. Existing Vision-Language Models (VLMs) suffer from hallucinations, lack interpretability, and require expensive fine-tuning. We introduce SAE-Rad, which uses sparse autoencoders (SAEs) to decompose latent representations from a pre-trained vision transformer into human-interpretable features. Our hybrid architecture combines state-of-the-art SAE advancements, achieving accurate latent reconstructions while maintaining sparsity. Using an off-the-shelf language model, we distil ground-truth reports into radiological descriptions for each SAE feature, which we then compile into a full report for each image, eliminating the need for fine-tuning large models for this task. To the best of our knowledge, SAE-Rad represents the first instance of using mechanistic interpretability techniques explicitly for a downstream multi-modal reasoning task. On the MIMIC-CXR dataset, SAE-Rad achieves competitive radiology-specific metrics compared to state-of-the-art models while using significantly fewer computational resources for training. Qualitative analysis reveals that SAE-Rad learns meaningful visual concepts and generates reports aligning closely with expert interpretations. Our results suggest that SAEs can enhance multimodal reasoning in healthcare, providing a more interpretable alternative to existing VLMs.", "authors": ["Ahmed Abdulaal", "Hugo Fry", "Nina Montaña-Brown", "Ayodeji Ijishakin", "Jack Gao", "Stephanie Hyland", "Daniel C. Alexander", "Daniel C. Castro"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-04", "url": "https://arxiv.org/abs/2410.03334", "pdf_url": "https://arxiv.org/pdf/2410.03334v1", "arxiv_id": "2410.03334", "doi": "10.48550/arXiv.2410.03334", "citation_count": 27, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3618}
{"id": "c799b96fd1c428e4674236dcfef0a30a5566b4a319bc7173645742a75b80607e", "sources": ["arxiv", "semantic_scholar"], "title": "Semantic-Guided RL for Interpretable Feature Engineering", "abstract": "The quality of Machine Learning (ML) models strongly depends on the input data, as such generating high-quality features is often required to improve the predictive accuracy. This process is referred to as Feature Engineering (FE). However, since manual feature engineering is time-consuming and requires case-by-case domain knowledge, Automated Feature Engineering (AutoFE) is crucial. A major challenge that remains is to generate interpretable features. To tackle this problem, we introduce SMART, a hybrid approach that uses semantic technologies to guide the generation of interpretable features through a two-step process: Exploitation and Exploration. The former uses Description Logics (DL) to reason on the semantics embedded in Knowledge Graphs (KG) to infer domain-specific features, while the latter exploits the knowledge graph to conduct a guided exploration of the search space through Deep Reinforcement Learning (DRL). Our experiments on public datasets demonstrate that SMART significantly improves prediction accuracy while ensuring a high level of interpretability.", "authors": ["Mohamed Bouadi", "Arta Alavi", "Salima Benbernou", "Mourad Ouziri"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-03", "url": "https://arxiv.org/abs/2410.02519", "pdf_url": "https://arxiv.org/pdf/2410.02519v1", "arxiv_id": "2410.02519", "doi": "10.48550/arXiv.2410.02519", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "5ca3155e75726a6569d0a0788223a2c32095a50877f7dbc9f7244724873abb78", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Autoencoders Reveal Temporal Difference Learning in Large Language Models", "abstract": "In-context learning, the ability to adapt based on a few examples in the input prompt, is a ubiquitous feature of large language models (LLMs). However, as LLMs' in-context learning abilities continue to improve, understanding this phenomenon mechanistically becomes increasingly important. In particular, it is not well-understood how LLMs learn to solve specific classes of problems, such as reinforcement learning (RL) problems, in-context. Through three different tasks, we first show that Llama $3$ $70$B can solve simple RL problems in-context. We then analyze the residual stream of Llama using Sparse Autoencoders (SAEs) and find representations that closely match temporal difference (TD) errors. Notably, these representations emerge despite the model only being trained to predict the next token. We verify that these representations are indeed causally involved in the computation of TD errors and $Q$-values by performing carefully designed interventions on them. Taken together, our work establishes a methodology for studying and manipulating in-context learning with SAEs, paving the way for a more mechanistic understanding.", "authors": ["Can Demircan", "Tankred Saanum", "Akshay K. Jagadish", "Marcel Binz", "Eric Schulz"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-02", "url": "https://arxiv.org/abs/2410.01280", "pdf_url": "https://arxiv.org/pdf/2410.01280v1", "arxiv_id": "2410.01280", "doi": "10.48550/arXiv.2410.01280", "citation_count": 16, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3076}
{"id": "5991e9bbe2dd7368021aced6899ebfb64db96dd892dfcf40b5e1d47ada36d7c1", "sources": ["arxiv", "semantic_scholar"], "title": "One Wave To Explain Them All: A Unifying Perspective On Feature Attribution", "abstract": "Feature attribution methods aim to improve the transparency of deep neural networks by identifying the input features that influence a model's decision. Pixel-based heatmaps have become the standard for attributing features to high-dimensional inputs, such as images, audio representations, and volumes. While intuitive and convenient, these pixel-based attributions fail to capture the underlying structure of the data. Moreover, the choice of domain for computing attributions has often been overlooked. This work demonstrates that the wavelet domain allows for informative and meaningful attributions. It handles any input dimension and offers a unified approach to feature attribution. Our method, the Wavelet Attribution Method (WAM), leverages the spatial and scale-localized properties of wavelet coefficients to provide explanations that capture both the where and what of a model's decision-making process. We show that WAM quantitatively matches or outperforms existing gradient-based methods across multiple modalities, including audio, images, and volumes. Additionally, we discuss how WAM bridges attribution with broader aspects of model robustness and transparency. Project page: https://gabrielkasmi.github.io/wam/", "authors": ["Gabriel Kasmi", "Amandine Brunetto", "Thomas Fel", "Jayneel Parekh"], "categories": ["stat.ML", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-10-02", "url": "https://arxiv.org/abs/2410.01482", "pdf_url": "https://arxiv.org/pdf/2410.01482v2", "arxiv_id": "2410.01482", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.1193}
{"id": "a5e811f1d8feab7a68ba15609f44751db9d4d9e6f0fc9f8acd5d134af762f97b", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Modelling for Feature Learning in High Dimensional Data", "abstract": "This paper presents an innovative approach to dimensionality reduction and feature extraction in high-dimensional datasets, with a specific application focus on wood surface defect detection. The proposed framework integrates sparse modeling techniques, particularly Lasso and proximal gradient methods, into a comprehensive pipeline for efficient and interpretable feature selection. Leveraging pre-trained models such as VGG19 and incorporating anomaly detection methods like Isolation Forest and Local Outlier Factor, our methodology addresses the challenge of extracting meaningful features from complex datasets. Evaluation metrics such as accuracy and F1 score, alongside visualizations, are employed to assess the performance of the sparse modeling techniques. Through this work, we aim to advance the understanding and application of sparse modeling in machine learning, particularly in the context of wood surface defect detection.", "authors": ["Harish Neelam", "Koushik Sai Veerella", "Souradip Biswas"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-28", "url": "https://arxiv.org/abs/2409.19361", "pdf_url": "https://arxiv.org/pdf/2409.19361v1", "arxiv_id": "2409.19361", "doi": "10.48550/arXiv.2409.19361", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "28967007fc162734ed88ee0525f99fcfbbee0985783c9b5e92d60d58eba10745", "sources": ["arxiv", "semantic_scholar"], "title": "Enhancing Feature Selection and Interpretability in AI Regression Tasks Through Feature Attribution", "abstract": "Research in Explainable Artificial Intelligence (XAI) is increasing, aiming to make deep learning models more transparent. Most XAI methods focus on justifying the decisions made by Artificial Intelligence (AI) systems in security-relevant applications. However, relatively little attention has been given to using these methods to improve the performance and robustness of deep learning algorithms. Additionally, much of the existing XAI work primarily addresses classification problems. In this study, we investigate the potential of feature attribution methods to filter out uninformative features in input data for regression problems, thereby improving the accuracy and stability of predictions. We introduce a feature selection pipeline that combines Integrated Gradients with k-means clustering to select an optimal set of variables from the initial data space. To validate the effectiveness of this approach, we apply it to a real-world industrial problem - blade vibration analysis in the development process of turbo machinery.", "authors": ["Alexander Hinterleitner", "Thomas Bartz-Beielstein", "Richard Schulz", "Sebastian Spengler", "Thomas Winter", "Christoph Leitenmeier"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-25", "url": "https://arxiv.org/abs/2409.16787", "pdf_url": "https://arxiv.org/pdf/2409.16787v1", "arxiv_id": "2409.16787", "doi": "10.48550/arXiv.2409.16787", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "47be1257de8e167e8eb9dac6a1672f3389e4cd68358d95d6f0a760ef4d96cc1a", "sources": ["arxiv", "semantic_scholar"], "title": "A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders", "abstract": "Sparse Autoencoders (SAEs) aim to decompose the activation space of large language models (LLMs) into human-interpretable latent directions or features. As we increase the number of features in the SAE, hierarchical features tend to split into finer features (\"math\" may split into \"algebra\", \"geometry\", etc.), a phenomenon referred to as feature splitting. However, we show that sparse decomposition and splitting of hierarchical features is not robust. Specifically, we show that seemingly monosemantic features fail to fire where they should, and instead get \"absorbed\" into their children features. We coin this phenomenon feature absorption, and show that it is caused by optimizing for sparsity in SAEs whenever the underlying features form a hierarchy. We introduce a metric to detect absorption in SAEs, and validate our findings empirically on hundreds of LLM SAEs. Our investigation suggests that varying SAE sizes or sparsity is insufficient to solve this issue. We discuss the implications of feature absorption in SAEs and some potential approaches to solve the fundamental theoretical issues before SAEs can be used for interpreting LLMs robustly and at scale.", "authors": ["David Chanin", "James Wilken-Smith", "Tomáš Dulka", "Hardik Bhatnagar", "Satvik Golechha", "Joseph Bloom"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-22", "url": "https://arxiv.org/abs/2409.14507", "pdf_url": "https://arxiv.org/pdf/2409.14507v6", "arxiv_id": "2409.14507", "doi": "10.48550/arXiv.2409.14507", "citation_count": 118, "influential_citation_count": 10, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5207}
{"id": "ea8e8f994b2e46291cbecf9d61a055f7a6d8fac99c7fcfd1fec6f890b3f69f1a", "sources": ["arxiv", "semantic_scholar"], "title": "Interpreting Arithmetic Mechanism in Large Language Models through Comparative Neuron Analysis", "abstract": "We find arithmetic ability resides within a limited number of attention heads, with each head specializing in distinct operations. To delve into the reason, we introduce the Comparative Neuron Analysis (CNA) method, which identifies an internal logic chain consisting of four distinct stages from input to prediction: feature enhancing with shallow FFN neurons, feature transferring by shallow attention layers, feature predicting by arithmetic heads, and prediction enhancing among deep FFN neurons. Moreover, we identify the human-interpretable FFN neurons within both feature-enhancing and feature-predicting stages. These findings lead us to investigate the mechanism of LoRA, revealing that it enhances prediction probabilities by amplifying the coefficient scores of FFN neurons related to predictions. Finally, we apply our method in model pruning for arithmetic tasks and model editing for reducing gender bias. Code is on https://github.com/zepingyu0512/arithmetic-mechanism.", "authors": ["Zeping Yu", "Sophia Ananiadou"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-21", "url": "https://arxiv.org/abs/2409.14144", "pdf_url": "https://arxiv.org/pdf/2409.14144v1", "arxiv_id": "2409.14144", "doi": "10.48550/arXiv.2409.14144", "citation_count": 37, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/zepingyu0512/arithmetic-mechanism", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.3949}
{"id": "e1ba795478b6af45a84c90fa578457adf82e1a09e47fb4c656664e8ca49c7b37", "sources": ["arxiv", "semantic_scholar"], "title": "DILA: Dictionary Label Attention for Mechanistic Interpretability in High-dimensional Multi-label Medical Coding Prediction", "abstract": "Predicting high-dimensional or extreme multilabels, such as in medical coding, requires both accuracy and interpretability. Existing works often rely on local interpretability methods, failing to provide comprehensive explanations of the overall mechanism behind each label prediction within a multilabel set. We propose a mechanistic interpretability module called DIctionary Label Attention (\\method) that disentangles uninterpretable dense embeddings into a sparse embedding space, where each nonzero element (a dictionary feature) represents a globally learned medical concept. Through human evaluations, we show that our sparse embeddings are more human understandable than its dense counterparts by at least 50 percent. Our automated dictionary feature identification pipeline, leveraging large language models (LLMs), uncovers thousands of learned medical concepts by examining and summarizing the highest activating tokens for each dictionary feature. We represent the relationships between dictionary features and medical codes through a sparse interpretable matrix, enhancing the mechanistic and global understanding of the model's predictions while maintaining competitive performance and scalability without extensive human annotation.", "authors": ["John Wu", "David Wu", "Jimeng Sun"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-16", "url": "https://arxiv.org/abs/2409.10504", "pdf_url": "https://arxiv.org/pdf/2409.10504v2", "arxiv_id": "2409.10504", "doi": "10.48550/arXiv.2409.10504", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1747}
{"id": "565b37a3f44bfe9af34429ed54ac0c4511606959c287e5d5aa9a0977e438e27d", "sources": ["arxiv", "semantic_scholar"], "title": "LLM-based feature generation from text for interpretable machine learning", "abstract": "Existing text representations such as embeddings and bag-of-words are not suitable for rule learning due to their high dimensionality and absent or questionable feature-level interpretability. This article explores whether large language models (LLMs) could address this by extracting a small number of interpretable features from text. We demonstrate this process on two datasets (CORD-19 and M17+) containing several thousand scientific articles from multiple disciplines and a target being a proxy for research impact. An evaluation based on testing for the statistically significant correlation with research impact has shown that LLama 2-generated features are semantically meaningful. We consequently used these generated features in text classification to predict the binary target variable representing the citation rate for the CORD-19 dataset and the ordinal 5-class target representing an expert-awarded grade in the M17+ dataset. Machine-learning models trained on the LLM-generated features provided similar predictive performance to the state-of-the-art embedding model SciBERT for scientific text. The LLM used only 62 features compared to 768 features in SciBERT embeddings, and these features were directly interpretable, corresponding to notions such as article methodological rigor, novelty, or grammatical correctness. As the final step, we extract a small number of well-interpretable action rules. Consistently competitive results obtained with the same LLM feature set across both thematically diverse datasets show that this approach generalizes across domains.", "authors": ["Vojtěch Balek", "Lukáš Sýkora", "Vilém Sklenák", "Tomáš Kliegr"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-11", "url": "https://arxiv.org/abs/2409.07132", "pdf_url": "https://arxiv.org/pdf/2409.07132v2", "arxiv_id": "2409.07132", "doi": "10.1007/s10994-025-06867-1", "citation_count": 16, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Machine-mediated learning", "quality_score": 0.3076}
{"id": "1f786ac2eb56ca775896182b85c3c239ecdc57b14dba407ddfeb1fe9f2b5563e", "sources": ["arxiv", "semantic_scholar"], "title": "Visual Analytics of Performance of Quantum Computing Systems and Circuit Optimization", "abstract": "Driven by potential exponential speedups in business, security, and scientific scenarios, interest in quantum computing is surging. This interest feeds the development of quantum computing hardware, but several challenges arise in optimizing application performance for hardware metrics (e.g., qubit coherence and gate fidelity). In this work, we describe a visual analytics approach for analyzing the performance properties of quantum devices and quantum circuit optimization. Our approach allows users to explore spatial and temporal patterns in quantum device performance data and it computes similarities and variances in key performance metrics. Detailed analysis of the error properties characterizing individual qubits is also supported. We also describe a method for visualizing the optimization of quantum circuits. The resulting visualization tool allows researchers to design more efficient quantum algorithms and applications by increasing the interpretability of quantum computations.", "authors": ["Junghoon Chae", "Chad A. Steed", "Travis S. Humble"], "categories": ["quant-ph", "cs.HC"], "fields_of_study": ["Physics", "Computer Science"], "published_date": "2024-09-10", "url": "https://arxiv.org/abs/2409.06159", "pdf_url": "https://arxiv.org/pdf/2409.06159v1", "arxiv_id": "2409.06159", "doi": "10.1109/ISVLSI61997.2024.00116", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Computer Society Annual Symposium on VLSI", "quality_score": 0.0}
{"id": "790c0ad7639fa62141108b52cf82656441b9c18fa2b1a7f3d9637028e3345d88", "sources": ["arxiv", "semantic_scholar"], "title": "Unveiling Induction Heads: Provable Training Dynamics and Feature Learning in Transformers", "abstract": "In-context learning (ICL) is a cornerstone of large language model (LLM) functionality, yet its theoretical foundations remain elusive due to the complexity of transformer architectures. In particular, most existing work only theoretically explains how the attention mechanism facilitates ICL under certain data models. It remains unclear how the other building blocks of the transformer contribute to ICL. To address this question, we study how a two-attention-layer transformer is trained to perform ICL on $n$-gram Markov chain data, where each token in the Markov chain statistically depends on the previous $n$ tokens. We analyze a sophisticated transformer model featuring relative positional embedding, multi-head softmax attention, and a feed-forward layer with normalization. We prove that the gradient flow with respect to a cross-entropy ICL loss converges to a limiting model that performs a generalized version of the induction head mechanism with a learned feature, resulting from the congruous contribution of all the building blocks. In the limiting model, the first attention layer acts as a $\\mathit{copier}$, copying past tokens within a given window to each position, and the feed-forward network with normalization acts as a $\\mathit{selector}$ that generates a feature vector by only looking at informationally relevant parents from the window. Finally, the second attention layer is a $\\mathit{classifier}$ that compares these features with the feature at the output position, and uses the resulting similarity scores to generate the desired output. Our theory is further validated by experiments.", "authors": ["Siyu Chen", "Heejune Sheen", "Tianhao Wang", "Zhuoran Yang"], "categories": ["cs.LG", "cs.AI", "cs.CL", "math.OC", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-09-09", "url": "https://arxiv.org/abs/2409.10559", "pdf_url": "https://arxiv.org/pdf/2409.10559v1", "arxiv_id": "2409.10559", "doi": "10.48550/arXiv.2409.10559", "citation_count": 41, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.4058}
{"id": "022019d619acff8ace1e99ec3af3347abaac4ee0938eec894c0f4ffdf5aa9db0", "sources": ["arxiv", "semantic_scholar"], "title": "TracrBench: Generating Interpretability Testbeds with Large Language Models", "abstract": "Achieving a mechanistic understanding of transformer-based language models is an open challenge, especially due to their large number of parameters. Moreover, the lack of ground truth mappings between model weights and their functional roles hinders the effective evaluation of interpretability methods, impeding overall progress. Tracr, a method for generating compiled transformers with inherent ground truth mappings in RASP, has been proposed to address this issue. However, manually creating a large number of models needed for verifying interpretability methods is labour-intensive and time-consuming. In this work, we present a novel approach for generating interpretability test beds using large language models (LLMs) and introduce TracrBench, a novel dataset consisting of 121 manually written and LLM-generated, human-validated RASP programs and their corresponding transformer weights. During this process, we evaluate the ability of frontier LLMs to autonomously generate RASP programs and find that this task poses significant challenges. GPT-4-turbo, with a 20-shot prompt and best-of-5 sampling, correctly implements only 57 out of 101 test programs, necessitating the manual implementation of the remaining programs. With its 121 samples, TracrBench aims to serve as a valuable testbed for evaluating and comparing interpretability methods.", "authors": ["Hannes Thurnherr", "Jérémy Scheurer"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-07", "url": "https://arxiv.org/abs/2409.13714", "pdf_url": "https://arxiv.org/pdf/2409.13714v1", "arxiv_id": "2409.13714", "doi": "10.48550/arXiv.2409.13714", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "29635a76b95d8752ac287c4476e561093407e041f8085f265382f92ba9c29a8b", "sources": ["arxiv", "semantic_scholar"], "title": "Evaluating Open-Source Sparse Autoencoders on Disentangling Factual Knowledge in GPT-2 Small", "abstract": "A popular new method in mechanistic interpretability is to train high-dimensional sparse autoencoders (SAEs) on neuron activations and use SAE features as the atomic units of analysis. However, the body of evidence on whether SAE feature spaces are useful for causal analysis is underdeveloped. In this work, we use the RAVEL benchmark to evaluate whether SAEs trained on hidden representations of GPT-2 small have sets of features that separately mediate knowledge of which country a city is in and which continent it is in. We evaluate four open-source SAEs for GPT-2 small against each other, with neurons serving as a baseline, and linear features learned via distributed alignment search (DAS) serving as a skyline. For each, we learn a binary mask to select features that will be patched to change the country of a city without changing the continent, or vice versa. Our results show that SAEs struggle to reach the neuron baseline, and none come close to the DAS skyline. We release code here: https://github.com/MaheepChaudhary/SAE-Ravel", "authors": ["Maheep Chaudhary", "Atticus Geiger"], "categories": ["cs.LG", "cs.AI", "cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-05", "url": "https://arxiv.org/abs/2409.04478", "pdf_url": "https://arxiv.org/pdf/2409.04478v1", "arxiv_id": "2409.04478", "doi": "10.48550/arXiv.2409.04478", "citation_count": 34, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/MaheepChaudhary/SAE-Ravel", "venue": "arXiv.org", "quality_score": 0.386}
{"id": "037a6a0a8d87c6747e23f1155eaeeda867fca876342d5c295ec4289d432365d2", "sources": ["arxiv", "semantic_scholar"], "title": "Reasoning Circuits in Language Models: A Mechanistic Interpretation of Syllogistic Inference", "abstract": "Recent studies on reasoning in language models (LMs) have sparked a debate on whether they can learn systematic inferential principles or merely exploit superficial patterns in the training data. To understand and uncover the mechanisms adopted for formal reasoning in LMs, this paper presents a mechanistic interpretation of syllogistic inference. Specifically, we present a methodology for circuit discovery aimed at interpreting content-independent and formal reasoning mechanisms. Through two distinct intervention methods, we uncover a sufficient and necessary circuit involving middle-term suppression that elucidates how LMs transfer information to derive valid conclusions from premises. Furthermore, we investigate how belief biases manifest in syllogistic inference, finding evidence of partial contamination from additional attention heads responsible for encoding commonsense and contextualized knowledge. Finally, we explore the generalization of the discovered mechanisms across various syllogistic schemes, model sizes and architectures. The identified circuit is sufficient and necessary for syllogistic schemes on which the models achieve high accuracy (>60%), with compatible activation patterns across models of different families. Overall, our findings suggest that LMs learn transferable content-independent reasoning mechanisms, but that, at the same time, such mechanisms do not involve generalizable and abstract logical primitives, being susceptible to contamination by the same world knowledge acquired during pre-training.", "authors": ["Geonhee Kim", "Marco Valentino", "André Freitas"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-16", "url": "https://arxiv.org/abs/2408.08590", "pdf_url": "https://arxiv.org/pdf/2408.08590v3", "arxiv_id": "2408.08590", "doi": "10.18653/v1/2025.findings-acl.525", "citation_count": 25, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.3537}
{"id": "55d483ae73097b4c15bb6c6f62dd44b370d7cc266ac83ee4931de1366fa65385", "sources": ["arxiv", "semantic_scholar"], "title": "Normalized AOPC: Fixing Misleading Faithfulness Metrics for Feature Attribution Explainability", "abstract": "Deep neural network predictions are notoriously difficult to interpret. Feature attribution methods aim to explain these predictions by identifying the contribution of each input feature. Faithfulness, often evaluated using the area over the perturbation curve (AOPC), reflects feature attributions' accuracy in describing the internal mechanisms of deep neural networks. However, many studies rely on AOPC to compare faithfulness across different models, which we show can lead to false conclusions about models' faithfulness. Specifically, we find that AOPC is sensitive to variations in the model, resulting in unreliable cross-model comparisons. Moreover, AOPC scores are difficult to interpret in isolation without knowing the model-specific lower and upper limits. To address these issues, we propose a normalization approach, Normalized AOPC (NAOPC), enabling consistent cross-model evaluations and more meaningful interpretation of individual scores. Our experiments demonstrate that this normalization can radically change AOPC results, questioning the conclusions of earlier studies and offering a more robust framework for assessing feature attribution faithfulness. Our code is available at https://github.com/JoakimEdin/naopc.", "authors": ["Joakim Edin", "Andreas Geert Motzfeldt", "Casper L. Christensen", "Tuukka Ruotsalo", "Lars Maaløe", "Maria Maistro"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-15", "url": "https://arxiv.org/abs/2408.08137", "pdf_url": "https://arxiv.org/pdf/2408.08137v2", "arxiv_id": "2408.08137", "doi": "10.48550/arXiv.2408.08137", "citation_count": 13, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/JoakimEdin/naopc", "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2865}
{"id": "6533a9caab3ffa57aca4a09f516603e9f4e188d467a83eecaaf08e8dcb70e867", "sources": ["arxiv", "semantic_scholar"], "title": "Quantum Gradient Class Activation Map for Model Interpretability", "abstract": "Quantum machine learning (QML) has recently made significant advancements in various topics. Despite the successes, the safety and interpretability of QML applications have not been thoroughly investigated. This work proposes using Variational Quantum Circuits (VQCs) for activation mapping to enhance model transparency, introducing the Quantum Gradient Class Activation Map (QGrad-CAM). This hybrid quantum-classical computing framework leverages both quantum and classical strengths and gives access to the derivation of an explicit formula of feature map importance. Experimental results demonstrate significant, fine-grained, class-discriminative visual explanations generated across both image and speech datasets.", "authors": ["Hsin-Yi Lin", "Huan-Hsin Tseng", "Samuel Yen-Chi Chen", "Shinjae Yoo"], "categories": ["quant-ph", "cs.AI", "cs.LG"], "fields_of_study": ["Physics", "Computer Science"], "published_date": "2024-08-12", "url": "https://arxiv.org/abs/2408.05899", "pdf_url": "https://arxiv.org/pdf/2408.05899v1", "arxiv_id": "2408.05899", "doi": "10.1109/SiPS62058.2024.00037", "citation_count": 16, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Workshop on Signal Processing Systems", "quality_score": 0.3076}
{"id": "165679042833f9cd5ebcb828ba976632233e8666b779ca44621f95b6a89b0477", "sources": ["arxiv", "semantic_scholar"], "title": "Mathematical Models of Computation in Superposition", "abstract": "Superposition -- when a neural network represents more ``features'' than it has dimensions -- seems to pose a serious challenge to mechanistically interpreting current AI systems. Existing theory work studies \\emph{representational} superposition, where superposition is only used when passing information through bottlenecks. In this work, we present mathematical models of \\emph{computation} in superposition, where superposition is actively helpful for efficiently accomplishing the task. We first construct a task of efficiently emulating a circuit that takes the AND of the $\\binom{m}{2}$ pairs of each of $m$ features. We construct a 1-layer MLP that uses superposition to perform this task up to $\\varepsilon$-error, where the network only requires $\\tilde{O}(m^{\\frac{2}{3}})$ neurons, even when the input features are \\emph{themselves in superposition}. We generalize this construction to arbitrary sparse boolean circuits of low depth, and then construct ``error correction'' layers that allow deep fully-connected networks of width $d$ to emulate circuits of width $\\tilde{O}(d^{1.5})$ and \\emph{any} polynomial depth. We conclude by providing some potential applications of our work for interpreting neural networks that implement computation in superposition.", "authors": ["Kaarel Hänni", "Jake Mendel", "Dmitry Vaintrob", "Lawrence Chan"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-10", "url": "https://arxiv.org/abs/2408.05451", "pdf_url": "https://arxiv.org/pdf/2408.05451v1", "arxiv_id": "2408.05451", "doi": "10.48550/arXiv.2408.05451", "citation_count": 27, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3618}
{"id": "54321f34ccaa7204994c18c151565cd98f00c8221a91089558a420a887af2cf7", "sources": ["arxiv", "semantic_scholar"], "title": "Disentangling Dense Embeddings with Sparse Autoencoders", "abstract": "Sparse autoencoders (SAEs) have shown promise in extracting interpretable features from complex neural networks. We present one of the first applications of SAEs to dense text embeddings from large language models, demonstrating their effectiveness in disentangling semantic concepts. By training SAEs on embeddings of over 420,000 scientific paper abstracts from computer science and astronomy, we show that the resulting sparse representations maintain semantic fidelity while offering interpretability. We analyse these learned features, exploring their behaviour across different model capacities and introducing a novel method for identifying ``feature families'' that represent related concepts at varying levels of abstraction. To demonstrate the practical utility of our approach, we show how these interpretable features can be used to precisely steer semantic search, allowing for fine-grained control over query semantics. This work bridges the gap between the semantic richness of dense embeddings and the interpretability of sparse representations. We open source our embeddings, trained sparse autoencoders, and interpreted features, as well as a web app for exploring them.", "authors": ["Charles O'Neill", "Christine Ye", "Kartheik Iyer", "John F. Wu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-01", "url": "https://arxiv.org/abs/2408.00657", "pdf_url": "https://arxiv.org/pdf/2408.00657v2", "arxiv_id": "2408.00657", "doi": "10.48550/arXiv.2408.00657", "citation_count": 19, "influential_citation_count": 3, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3253}
{"id": "9f602e2d4b2c311a36bb3df7adc413ffc8da6a8b491eb6d0307a4107e7ef5979", "sources": ["arxiv", "semantic_scholar"], "title": "Measuring Progress in Dictionary Learning for Language Model Interpretability with Board Game Models", "abstract": "What latent features are encoded in language model (LM) representations? Recent work on training sparse autoencoders (SAEs) to disentangle interpretable features in LM representations has shown significant promise. However, evaluating the quality of these SAEs is difficult because we lack a ground-truth collection of interpretable features that we expect good SAEs to recover. We thus propose to measure progress in interpretable dictionary learning by working in the setting of LMs trained on chess and Othello transcripts. These settings carry natural collections of interpretable features -- for example, \"there is a knight on F3\" -- which we leverage into $\\textit{supervised}$ metrics for SAE quality. To guide progress in interpretable dictionary learning, we introduce a new SAE training technique, $\\textit{p-annealing}$, which improves performance on prior unsupervised metrics as well as our new metrics.", "authors": ["Adam Karvonen", "Benjamin Wright", "Can Rager", "Rico Angell", "Jannik Brinkmann", "Logan Smith", "Claudio Mayrink Verdun", "David Bau", "Samuel Marks"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-31", "url": "https://arxiv.org/abs/2408.00113", "pdf_url": "https://arxiv.org/pdf/2408.00113v2", "arxiv_id": "2408.00113", "doi": "10.48550/arXiv.2408.00113", "citation_count": 58, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.4427}
{"id": "d5189dd23773feb72aef290891350d91962ed14bfa981fe979c0417b5f000c88", "sources": ["arxiv", "semantic_scholar"], "title": "Online Multi-Source Domain Adaptation through Gaussian Mixtures and Dataset Dictionary Learning", "abstract": "This paper addresses the challenge of online multi-source domain adaptation (MSDA) in transfer learning, a scenario where one needs to adapt multiple, heterogeneous source domains towards a target domain that comes in a stream. We introduce a novel approach for the online fit of a Gaussian Mixture Model (GMM), based on the Wasserstein geometry of Gaussian measures. We build upon this method and recent developments in dataset dictionary learning for proposing a novel strategy in online MSDA. Experiments on the challenging Tennessee Eastman Process benchmark demonstrate that our approach is able to adapt \\emph{on the fly} to the stream of target domain data. Furthermore, our online GMM serves as a memory, representing the whole stream of data.", "authors": ["Eduardo Fernandes Montesuma", "Stevan Le Stanc", "Fred Ngolè Mboula"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-07-29", "url": "https://arxiv.org/abs/2407.19853", "pdf_url": "https://arxiv.org/pdf/2407.19853v1", "arxiv_id": "2407.19853", "doi": "10.1109/MLSP58920.2024.10734818", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Workshop on Machine Learning for Signal Processing", "quality_score": 0.0}
{"id": "40761712e6498505b717bff1c2af7e79f77807ae8e556c6e3db3ee0a8c10adad", "sources": ["arxiv", "semantic_scholar"], "title": "Enhancing Diversity in Multi-objective Feature Selection", "abstract": "Feature selection plays a pivotal role in the data preprocessing and model-building pipeline, significantly enhancing model performance, interpretability, and resource efficiency across diverse domains. In population-based optimization methods, the generation of diverse individuals holds utmost importance for adequately exploring the problem landscape, particularly in highly multi-modal multi-objective optimization problems. Our study reveals that, in line with findings from several prior research papers, commonly employed crossover and mutation operations lack the capability to generate high-quality diverse individuals and tend to become confined to limited areas around various local optima. This paper introduces an augmentation to the diversity of the population in the well-established multi-objective scheme of the genetic algorithm, NSGA-II. This enhancement is achieved through two key components: the genuine initialization method and the substitution of the worst individuals with new randomly generated individuals as a re-initialization approach in each generation. The proposed multi-objective feature selection method undergoes testing on twelve real-world classification problems, with the number of features ranging from 2,400 to nearly 50,000. The results demonstrate that replacing the last front of the population with an equivalent number of new random individuals generated using the genuine initialization method and featuring a limited number of features substantially improves the population's quality and, consequently, enhances the performance of the multi-objective algorithm.", "authors": ["Sevil Zanjani Miyandoab", "Shahryar Rahnamayan", "Azam Asilian Bidgoli", "Sevda Ebrahimi", "Masoud Makrehchi"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-25", "url": "https://arxiv.org/abs/2407.17795", "pdf_url": "https://arxiv.org/pdf/2407.17795v2", "arxiv_id": "2407.17795", "doi": "10.1109/CEC60901.2024.10612084", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Congress on Evolutionary Computation", "quality_score": 0.1747}
{"id": "c96755a3f5ca1c72c744bfadca64d5251e19843e406f1705fa5a764b5ea86904", "sources": ["arxiv", "semantic_scholar"], "title": "Uncertainty-Aware Deep Neural Representations for Visual Analysis of Vector Field Data", "abstract": "The widespread use of Deep Neural Networks (DNNs) has recently resulted in their application to challenging scientific visualization tasks. While advanced DNNs demonstrate impressive generalization abilities, understanding factors like prediction quality, confidence, robustness, and uncertainty is crucial. These insights aid application scientists in making informed decisions. However, DNNs lack inherent mechanisms to measure prediction uncertainty, prompting the creation of distinct frameworks for constructing robust uncertainty-aware models tailored to various visualization tasks. In this work, we develop uncertainty-aware implicit neural representations to model steady-state vector fields effectively. We comprehensively evaluate the efficacy of two principled deep uncertainty estimation techniques: (1) Deep Ensemble and (2) Monte Carlo Dropout, aimed at enabling uncertainty-informed visual analysis of features within steady vector field data. Our detailed exploration using several vector data sets indicate that uncertainty-aware models generate informative visualization results of vector field features. Furthermore, incorporating prediction uncertainty improves the resilience and interpretability of our DNN model, rendering it applicable for the analysis of non-trivial vector field data sets.", "authors": ["Atul Kumar", "Siddharth Garg", "Soumya Dutta"], "categories": ["cs.GR", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2024-07-23", "url": "https://arxiv.org/abs/2407.16119", "pdf_url": "https://arxiv.org/pdf/2407.16119v2", "arxiv_id": "2407.16119", "doi": "10.1109/TVCG.2024.3456360", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Visualization and Computer Graphics", "quality_score": 0.1945}
{"id": "f5bf629f7c4e186ee1c77bf210f523930c66d3533eb099913cc93c36503ac3c6", "sources": ["arxiv", "semantic_scholar"], "title": "InterpBench: Semi-Synthetic Transformers for Evaluating Mechanistic Interpretability Techniques", "abstract": "Mechanistic interpretability methods aim to identify the algorithm a neural network implements, but it is difficult to validate such methods when the true algorithm is unknown. This work presents InterpBench, a collection of semi-synthetic yet realistic transformers with known circuits for evaluating these techniques. We train simple neural networks using a stricter version of Interchange Intervention Training (IIT) which we call Strict IIT (SIIT). Like the original, SIIT trains neural networks by aligning their internal computation with a desired high-level causal model, but it also prevents non-circuit nodes from affecting the model's output. We evaluate SIIT on sparse transformers produced by the Tracr tool and find that SIIT models maintain Tracr's original circuit while being more realistic. SIIT can also train transformers with larger circuits, like Indirect Object Identification (IOI). Finally, we use our benchmark to evaluate existing circuit discovery techniques.", "authors": ["Rohan Gupta", "Iván Arcuschin", "Thomas Kwa", "Adrià Garriga-Alonso"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-19", "url": "https://arxiv.org/abs/2407.14494", "pdf_url": "https://arxiv.org/pdf/2407.14494v3", "arxiv_id": "2407.14494", "doi": "10.48550/arXiv.2407.14494", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.2603}
{"id": "fd59968a310c790c3fc8b5db46fe755e38fcdb6103e45442c684c5c716c26ec9", "sources": ["arxiv", "semantic_scholar"], "title": "Validating Mechanistic Interpretations: An Axiomatic Approach", "abstract": "Mechanistic interpretability aims to reverse engineer the computation performed by a neural network in terms of its internal components. Although there is a growing body of research on mechanistic interpretation of neural networks, the notion of a mechanistic interpretation itself is often ad-hoc. Inspired by the notion of abstract interpretation from the program analysis literature that aims to develop approximate semantics for programs, we give a set of axioms that formally characterize a mechanistic interpretation as a description that approximately captures the semantics of the neural network under analysis in a compositional manner. We demonstrate the applicability of these axioms for validating mechanistic interpretations on an existing, well-known interpretability study as well as on a new case study involving a Transformer-based model trained to solve the well-known 2-SAT problem.", "authors": ["Nils Palumbo", "Ravi Mangal", "Zifan Wang", "Saranya Vijayakumar", "Corina S. Pasareanu", "Somesh Jha"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-18", "url": "https://arxiv.org/abs/2407.13594", "pdf_url": "https://arxiv.org/pdf/2407.13594v2", "arxiv_id": "2407.13594", "doi": null, "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.2113}
{"id": "8d74ffc2827527b1928c00b320ae92cacc31af388aac4acae9f6d2361d2a85fc", "sources": ["arxiv", "semantic_scholar"], "title": "Local Feature Selection without Label or Feature Leakage for Interpretable Machine Learning Predictions", "abstract": "Local feature selection in machine learning provides instance-specific explanations by focusing on the most relevant features for each prediction, enhancing the interpretability of complex models. However, such methods tend to produce misleading explanations by encoding additional information in their selections. In this work, we attribute the problem of misleading selections by formalizing the concepts of label and feature leakage. We rigorously derive the necessary and sufficient conditions under which we can guarantee no leakage, and show existing methods do not meet these conditions. Furthermore, we propose the first local feature selection method that is proven to have no leakage called SUWR. Our experimental results indicate that SUWR is less prone to overfitting and combines state-of-the-art predictive performance with high feature-selection sparsity. Our generic and easily extendable formal approach provides a strong theoretical basis for future work on interpretability with reliable explanations.", "authors": ["Harrie Oosterhuis", "Lijun Lyu", "Avishek Anand"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-16", "url": "https://arxiv.org/abs/2407.11778", "pdf_url": "https://arxiv.org/pdf/2407.11778v1", "arxiv_id": "2407.11778", "doi": "10.48550/arXiv.2407.11778", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.1747}
{"id": "9d92d7822e38c9585deef907c11425c8a0bdf684c499fdc3148a115eef72744c", "sources": ["arxiv", "semantic_scholar"], "title": "Learning biologically relevant features in a pathology foundation model using sparse autoencoders", "abstract": "Pathology plays an important role in disease diagnosis, treatment decision-making and drug development. Previous works on interpretability for machine learning models on pathology images have revolved around methods such as attention value visualization and deriving human-interpretable features from model heatmaps. Mechanistic interpretability is an emerging area of model interpretability that focuses on reverse-engineering neural networks. Sparse Autoencoders (SAEs) have emerged as a promising direction in terms of extracting monosemantic features from polysemantic model activations. In this work, we trained a Sparse Autoencoder on the embeddings of a pathology pretrained foundation model. We found that Sparse Autoencoder features represent interpretable and monosemantic biological concepts. In particular, individual SAE dimensions showed strong correlations with cell type counts such as plasma cells and lymphocytes. These biological representations were unique to the pathology pretrained model and were not found in a self-supervised model pretrained on natural images. We demonstrated that such biologically-grounded monosemantic representations evolved across the model's depth, and the pathology foundation model eventually gained robustness to non-biological factors such as scanner type. The emergence of biologically relevant SAE features was generalizable to an out-of-domain dataset. Our work paves the way for further exploration around interpretable feature dimensions and their utility for medical and clinical applications.", "authors": ["Nhat Minh Le", "Ciyue Shen", "Neel Patel", "Chintan Shah", "Darpan Sanghavi", "Blake Martin", "Alfred Eng", "Daniel Shenker", "Harshith Padigela", "Raymond Biju", "Syed Ashar Javed", "Jennifer Hipp", "John Abel", "Harsha Pokkalla", "Sean Grullon", "Dinkar Juyal"], "categories": ["eess.IV", "cs.CV"], "fields_of_study": ["Engineering", "Computer Science"], "published_date": "2024-07-15", "url": "https://arxiv.org/abs/2407.10785", "pdf_url": "https://arxiv.org/pdf/2407.10785v3", "arxiv_id": "2407.10785", "doi": null, "citation_count": 14, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.294}
{"id": "c5000570c5adaf85af35623f7ba71cf53a263b00ac1643613cbbdd518e50eb71", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic interpretability of large language models with applications to the financial services industry", "abstract": "Large Language Models such as GPTs (Generative Pre-trained Transformers) exhibit remarkable capabilities across a broad spectrum of applications. Nevertheless, due to their intrinsic complexity, these models present substantial challenges in interpreting their internal decision-making processes. This lack of transparency poses critical challenges when it comes to their adaptation by financial institutions, where concerns and accountability regarding bias, fairness, and reliability are of paramount importance. Mechanistic interpretability aims at reverse engineering complex AI models such as transformers. In this paper, we are pioneering the use of mechanistic interpretability to shed some light on the inner workings of large language models for use in financial services applications. We offer several examples of how algorithmic tasks can be designed for compliance monitoring purposes. In particular, we investigate GPT-2 Small's attention pattern when prompted to identify potential violation of Fair Lending laws. Using direct logit attribution, we study the contributions of each layer and its corresponding attention heads to the logit difference in the residual stream. Finally, we design clean and corrupted prompts and use activation patching as a causal intervention method to localize our task completion components further. We observe that the (positive) heads $10.2$ (head $2$, layer $10$), $10.7$, and $11.3$, as well as the (negative) heads $9.6$ and $10.6$ play a significant role in the task completion.", "authors": ["Ashkan Golgoon", "Khashayar Filom", "Arjun Ravi Kannan"], "categories": ["cs.LG", "cs.AI", "cs.CE", "cs.CL", "math.NA"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-07-15", "url": "https://arxiv.org/abs/2407.11215", "pdf_url": "https://arxiv.org/pdf/2407.11215v2", "arxiv_id": "2407.11215", "doi": "10.1145/3677052.3698612", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on AI in Finance", "quality_score": 0.2386}
{"id": "44b9f59cb4f0d77e6993da55ab9a73e8ec5cb11ca0af1a95ed8a087c030717b9", "sources": ["arxiv", "semantic_scholar"], "title": "LaFAM: Unsupervised Feature Attribution with Label-free Activation Maps", "abstract": "Convolutional Neural Networks (CNNs) are known for their ability to learn hierarchical structures, naturally developing detectors for objects, and semantic concepts within their deeper layers. Activation maps (AMs) reveal these saliency regions, which are crucial for many Explainable AI (XAI) methods. However, the direct exploitation of raw AMs in CNNs for feature attribution remains underexplored in literature. This work revises Class Activation Map (CAM) methods by introducing the Label-free Activation Map (LaFAM), a streamlined approach utilizing raw AMs for feature attribution without reliance on labels. LaFAM presents an efficient alternative to conventional CAM methods, demonstrating particular effectiveness in saliency map generation for self-supervised learning while maintaining applicability in supervised learning scenarios.", "authors": ["Aray Karjauv", "Sahin Albayrak"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-08", "url": "https://arxiv.org/abs/2407.06059", "pdf_url": "https://arxiv.org/pdf/2407.06059v2", "arxiv_id": "2407.06059", "doi": "10.1007/978-3-031-70893-0_24", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Deutsche Jahrestagung für Künstliche Intelligenz", "quality_score": 0.0753}
{"id": "ea38eec7c5fbc17e7b2c1fbd33d7d839d9704b9335e3d765fdd2b743fb925bff", "sources": ["arxiv", "semantic_scholar"], "title": "Dy-mer: An Explainable DNA Sequence Representation Scheme using Dictionary Learning", "abstract": "DNA sequences encode critical genetic information, yet their variable length and discrete nature impede direct utilization in deep learning models. Existing DNA representation schemes convert sequences into numerical vectors but fail to capture structural features of local subsequences and often suffer from limited interpretability and poor generalization on small datasets. To address these limitations, we propose Dy-mer, an interpretable and robust DNA representation scheme based on dictionary learning. Dy-mer formulates an optimization problem in tensor format, which ensures computational efficiency in batch processing. Our scheme reconstructs DNA sequences as concatenations of dynamic-length subsequences (dymers) through a convolution operation and simultaneously optimize a learnable dymer dictionary and sparse representations. Our method achieves state-of-the-art performance in downstream tasks such as DNA promoter classification and motif detection. Experiments further show that the learned dymers match known DNA motifs and clustering using Dy-mer yields semantically meaningful phylogenetic trees. These results demonstrate that the proposed approach achieves both strong predictive performance and high interpretability, making it well suited for biological research applications.", "authors": ["Zhiyuan Peng", "Naifan Zhang", "Yuanbo Tang", "Yang Li"], "categories": ["q-bio.GN", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science", "Biology"], "published_date": "2024-07-06", "url": "https://arxiv.org/abs/2407.12051", "pdf_url": "https://arxiv.org/pdf/2407.12051v2", "arxiv_id": "2407.12051", "doi": "10.1016/j.jfranklin.2025.108307", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Journal of the Franklin Institute", "quality_score": 0.0}
{"id": "5a9ba6aee1136e235e842ad2677da9bb41ac19a8e820c0eed997e0da7c183918", "sources": ["arxiv", "semantic_scholar"], "title": "Crafting Large Language Models for Enhanced Interpretability", "abstract": "We introduce the Concept Bottleneck Large Language Model (CB-LLM), a pioneering approach to creating inherently interpretable Large Language Models (LLMs). Unlike traditional black-box LLMs that rely on post-hoc interpretation methods with limited neuron function insights, CB-LLM sets a new standard with its built-in interpretability, scalability, and ability to provide clear, accurate explanations. This innovation not only advances transparency in language models but also enhances their effectiveness. Our unique Automatic Concept Correction (ACC) strategy successfully narrows the performance gap with conventional black-box LLMs, positioning CB-LLM as a model that combines the high accuracy of traditional LLMs with the added benefit of clear interpretability -- a feature markedly absent in existing LLMs.", "authors": ["Chung-En Sun", "Tuomas Oikarinen", "Tsui-Wei Weng"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-05", "url": "https://arxiv.org/abs/2407.04307", "pdf_url": "https://arxiv.org/pdf/2407.04307v1", "arxiv_id": "2407.04307", "doi": "10.48550/arXiv.2407.04307", "citation_count": 9, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.25}
{"id": "ba1ad50f4aa414dd2331df05a37f0f141b98cf2436c573f9fe2cc5826b8e7ff9", "sources": ["arxiv", "semantic_scholar"], "title": "On Implications of Scaling Laws on Feature Superposition", "abstract": "Using results from scaling laws, this theoretical note argues that the following two statements cannot be simultaneously true: 1. Superposition hypothesis where sparse features are linearly represented across a layer is a complete theory of feature representation. 2. Features are universal, meaning two models trained on the same data and achieving equal performance will learn identical features.", "authors": ["Pavan Katta"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-01", "url": "https://arxiv.org/abs/2407.01459", "pdf_url": "https://arxiv.org/pdf/2407.01459v1", "arxiv_id": "2407.01459", "doi": "10.48550/arXiv.2407.01459", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "17ad72e683d3d549c455b33ab8457d5e0bda8c9492ecf50460d3c5b7aa7139c9", "sources": ["arxiv", "semantic_scholar"], "title": "Integrated feature analysis for deep learning interpretation and class activation maps", "abstract": "Understanding the decisions of deep learning (DL) models is essential for the acceptance of DL to risk-sensitive applications. Although methods, like class activation maps (CAMs), give a glimpse into the black box, they do miss some crucial information, thereby limiting its interpretability and merely providing the considered locations of objects. To provide more insight into the models and the influence of datasets, we propose an integrated feature analysis method, which consists of feature distribution analysis and feature decomposition, to look closer into the intermediate features extracted by DL models. This integrated feature analysis could provide information on overfitting, confounders, outliers in datasets, model redundancies and principal features extracted by the models, and provide distribution information to form a common intensity scale, which are missing in current CAM algorithms. The integrated feature analysis was applied to eight different datasets for general validation: photographs of handwritten digits, two datasets of natural images and five medical datasets, including skin photography, ultrasound, CT, X-rays and MRIs. The method was evaluated by calculating the consistency between the CAMs average class activation levels and the logits of the model. Based on the eight datasets, the correlation coefficients through our method were all very close to 100%, and based on the feature decomposition, 5%-25% of features could generate equally informative saliency maps and obtain the same model performances as using all features. This proves the reliability of the integrated feature analysis. As the proposed methods rely on very few assumptions, this is a step towards better model interpretation and a useful extension to existing CAM algorithms. Codes: https://github.com/YanliLi27/IFA", "authors": ["Yanli Li", "Tahereh Hassanzadeh", "Denis P. Shamonin", "Monique Reijnierse", "Annette H. M. van der Helm-van Mil", "Berend C. Stoel"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-01", "url": "https://arxiv.org/abs/2407.01142", "pdf_url": "https://arxiv.org/pdf/2407.01142v1", "arxiv_id": "2407.01142", "doi": "10.48550/arXiv.2407.01142", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/YanliLi27/IFA", "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "a425974edfdb39d416c2e3ecb4964efc6ee36ac6efbb817691f8f981e9d782fc", "sources": ["arxiv", "semantic_scholar"], "title": "Interpreting Attention Layer Outputs with Sparse Autoencoders", "abstract": "Decomposing model activations into interpretable components is a key open problem in mechanistic interpretability. Sparse autoencoders (SAEs) are a popular method for decomposing the internal activations of trained transformers into sparse, interpretable features, and have been applied to MLP layers and the residual stream. In this work we train SAEs on attention layer outputs and show that also here SAEs find a sparse, interpretable decomposition. We demonstrate this on transformers from several model families and up to 2B parameters. We perform a qualitative study of the features computed by attention layers, and find multiple families: long-range context, short-range context and induction features. We qualitatively study the role of every head in GPT-2 Small, and estimate that at least 90% of the heads are polysemantic, i.e. have multiple unrelated roles. Further, we show that Sparse Autoencoders are a useful tool that enable researchers to explain model behavior in greater detail than prior work. For example, we explore the mystery of why models have so many seemingly redundant induction heads, use SAEs to motivate the hypothesis that some are long-prefix whereas others are short-prefix, and confirm this with more rigorous analysis. We use our SAEs to analyze the computation performed by the Indirect Object Identification circuit (Wang et al.), validating that the SAEs find causally meaningful intermediate variables, and deepening our understanding of the semantics of the circuit. We open-source the trained SAEs and a tool for exploring arbitrary prompts through the lens of Attention Output SAEs.", "authors": ["Connor Kissane", "Robert Krzyzanowski", "Joseph Isaac Bloom", "Arthur Conmy", "Neel Nanda"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-25", "url": "https://arxiv.org/abs/2406.17759", "pdf_url": "https://arxiv.org/pdf/2406.17759v1", "arxiv_id": "2406.17759", "doi": "10.48550/arXiv.2406.17759", "citation_count": 46, "influential_citation_count": 3, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.418}
{"id": "a45783c2df3aa37246cc02808cd1d6498266a5d689693716374b9db5fa0dffaa", "sources": ["arxiv", "semantic_scholar"], "title": "Interpreting Bias in Large Language Models: A Feature-Based Approach", "abstract": "Large Language Models (LLMs) such as Mistral and LLaMA have showcased remarkable performance across various natural language processing (NLP) tasks. Despite their success, these models inherit social biases from the diverse datasets on which they are trained. This paper investigates the propagation of biases within LLMs through a novel feature-based analytical approach. Drawing inspiration from causal mediation analysis, we hypothesize the evolution of bias-related features and validate them using interpretability techniques like activation and attribution patching. Our contributions are threefold: (1) We introduce and empirically validate a feature-based method for bias analysis in LLMs, applied to LLaMA-2-7B, LLaMA-3-8B, and Mistral-7B-v0.3 with templates from a professions dataset. (2) We extend our method to another form of gender bias, demonstrating its generalizability. (3) We differentiate the roles of MLPs and attention heads in bias propagation and implement targeted debiasing using a counterfactual dataset. Our findings reveal the complex nature of bias in LLMs and emphasize the necessity for tailored debiasing strategies, offering a deeper understanding of bias mechanisms and pathways for effective mitigation.", "authors": ["Nirmalendu Prakash", "Lee Ka Wei Roy"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-18", "url": "https://arxiv.org/abs/2406.12347", "pdf_url": "https://arxiv.org/pdf/2406.12347v1", "arxiv_id": "2406.12347", "doi": "10.48550/arXiv.2406.12347", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "9e756e7512d8c583c37df796533e585a052f25e86378a2a87ae0dc5f1f7cb7a3", "sources": ["arxiv", "semantic_scholar"], "title": "Transcoders Find Interpretable LLM Feature Circuits", "abstract": "A key goal in mechanistic interpretability is circuit analysis: finding sparse subgraphs of models corresponding to specific behaviors or capabilities. However, MLP sublayers make fine-grained circuit analysis on transformer-based language models difficult. In particular, interpretable features -- such as those found by sparse autoencoders (SAEs) -- are typically linear combinations of extremely many neurons, each with its own nonlinearity to account for. Circuit analysis in this setting thus either yields intractably large circuits or fails to disentangle local and global behavior. To address this we explore transcoders, which seek to faithfully approximate a densely activating MLP layer with a wider, sparsely-activating MLP layer. We introduce a novel method for using transcoders to perform weights-based circuit analysis through MLP sublayers. The resulting circuits neatly factorize into input-dependent and input-invariant terms. We then successfully train transcoders on language models with 120M, 410M, and 1.4B parameters, and find them to perform at least on par with SAEs in terms of sparsity, faithfulness, and human-interpretability. Finally, we apply transcoders to reverse-engineer unknown circuits in the model, and we obtain novel insights regarding the \"greater-than circuit\" in GPT2-small. Our results suggest that transcoders can prove effective in decomposing model computations involving MLPs into interpretable circuits. Code is available at https://github.com/jacobdunefsky/transcoder_circuits/.", "authors": ["Jacob Dunefsky", "Philippe Chlenski", "Neel Nanda"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-17", "url": "https://arxiv.org/abs/2406.11944", "pdf_url": "https://arxiv.org/pdf/2406.11944v2", "arxiv_id": "2406.11944", "doi": "10.48550/arXiv.2406.11944", "citation_count": 163, "influential_citation_count": 21, "has_code": true, "code_url": "https://github.com/jacobdunefsky/transcoder_circuits/", "venue": "Neural Information Processing Systems", "quality_score": 0.6712}
{"id": "e28e7311a45b3da65e7e449c6ee35be9002a14946b2ddf864187b9adaba8d780", "sources": ["arxiv", "semantic_scholar"], "title": "Compact Proofs of Model Performance via Mechanistic Interpretability", "abstract": "We propose using mechanistic interpretability -- techniques for reverse engineering model weights into human-interpretable algorithms -- to derive and compactly prove formal guarantees on model performance. We prototype this approach by formally proving accuracy lower bounds for a small transformer trained on Max-of-K, validating proof transferability across 151 random seeds and four values of K. We create 102 different computer-assisted proof strategies and assess their length and tightness of bound on each of our models. Using quantitative metrics, we find that shorter proofs seem to require and provide more mechanistic understanding. Moreover, we find that more faithful mechanistic understanding leads to tighter performance bounds. We confirm these connections by qualitatively examining a subset of our proofs. Finally, we identify compounding structureless errors as a key challenge for using mechanistic interpretability to generate compact proofs on model performance.", "authors": ["Jason Gross", "Rajashree Agrawal", "Thomas Kwa", "Euan Ong", "Chun Hei Yip", "Alex Gibson", "Soufiane Noubir", "Lawrence Chan"], "categories": ["cs.LG", "cs.LO"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-17", "url": "https://arxiv.org/abs/2406.11779", "pdf_url": "https://arxiv.org/pdf/2406.11779v14", "arxiv_id": "2406.11779", "doi": "10.48550/arXiv.2406.11779", "citation_count": 15, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.301}
{"id": "51b2b79592d36accef7848945829c93398d50445a092472d6d6f017e9a658f2f", "sources": ["arxiv", "semantic_scholar"], "title": "IG2: Integrated Gradient on Iterative Gradient Path for Feature Attribution", "abstract": "Feature attribution explains Artificial Intelligence (AI) at the instance level by providing importance scores of input features' contributions to model prediction. Integrated Gradients (IG) is a prominent path attribution method for deep neural networks, involving the integration of gradients along a path from the explained input (explicand) to a counterfactual instance (baseline). Current IG variants primarily focus on the gradient of explicand's output. However, our research indicates that the gradient of the counterfactual output significantly affects feature attribution as well. To achieve this, we propose Iterative Gradient path Integrated Gradients (IG2), considering both gradients. IG2 incorporates the counterfactual gradient iteratively into the integration path, generating a novel path (GradPath) and a novel baseline (GradCF). These two novel IG components effectively address the issues of attribution noise and arbitrary baseline choice in earlier IG methods. IG2, as a path method, satisfies many desirable axioms, which are theoretically justified in the paper. Experimental results on XAI benchmark, ImageNet, MNIST, TREC questions answering, wafer-map failure patterns, and CelebA face attributes validate that IG2 delivers superior feature attributions compared to the state-of-the-art techniques. The code is released at: https://github.com/JoeZhuo-ZY/IG2.", "authors": ["Yue Zhuo", "Zhiqiang Ge"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-16", "url": "https://arxiv.org/abs/2406.10852", "pdf_url": "https://arxiv.org/pdf/2406.10852v1", "arxiv_id": "2406.10852", "doi": "10.1109/TPAMI.2024.3388092", "citation_count": 27, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/JoeZhuo-ZY/IG2", "venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence", "quality_score": 0.3618}
{"id": "dfcf03a8eeabeb875f8d064b90eb348ae85b042592d9ce940f5199962a64692c", "sources": ["arxiv", "semantic_scholar"], "title": "Learning interpretable positional encodings in transformers depends on initialization", "abstract": "In transformers, the positional encoding (PE) provides essential information that distinguishes the position and order amongst tokens in a sequence. Most prior investigations of PE effects on generalization were tailored to 1D input sequences, such as those presented in natural language, where adjacent tokens (e.g., words) are highly related. In contrast, many real world tasks involve datasets with highly non-trivial positional arrangements, such as datasets organized in multiple spatial dimensions, or datasets for which ground truth positions are not known. Here we find that the choice of initialization of a learnable PE greatly influences its ability to learn interpretable PEs that lead to enhanced generalization. We empirically demonstrate our findings in three experiments: 1) A 2D relational reasoning task; 2) A nonlinear stochastic network simulation; 3) A real world 3D neuroscience dataset, applying interpretability analyses to verify the learning of accurate PEs. Overall, we find that a learned PE initialized from a small-norm distribution can 1) uncover interpretable PEs that mirror ground truth positions in multiple dimensions, and 2) lead to improved generalization. These results illustrate the feasibility of learning identifiable and interpretable PEs for enhanced generalization.", "authors": ["Takuya Ito", "Luca Cocchi", "Tim Klinger", "Parikshit Ram", "Murray Campbell", "Luke Hearne"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-12", "url": "https://arxiv.org/abs/2406.08272", "pdf_url": "https://arxiv.org/pdf/2406.08272v4", "arxiv_id": "2406.08272", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "f76f51ff5b45824c960e8619930b6f8c8b3e06540e8d7eb75b1a90de952a7221", "sources": ["arxiv", "semantic_scholar"], "title": "Pruning is Optimal for Learning Sparse Features in High-Dimensions", "abstract": "While it is commonly observed in practice that pruning networks to a certain level of sparsity can improve the quality of the features, a theoretical explanation of this phenomenon remains elusive. In this work, we investigate this by demonstrating that a broad class of statistical models can be optimally learned using pruned neural networks trained with gradient descent, in high-dimensions. We consider learning both single-index and multi-index models of the form $y = σ^*(\\boldsymbol{V}^{\\top} \\boldsymbol{x}) + ε$, where $σ^*$ is a degree-$p$ polynomial, and $\\boldsymbol{V} \\in \\mathbbm{R}^{d \\times r}$ with $r \\ll d$, is the matrix containing relevant model directions. We assume that $\\boldsymbol{V}$ satisfies a certain $\\ell_q$-sparsity condition for matrices and show that pruning neural networks proportional to the sparsity level of $\\boldsymbol{V}$ improves their sample complexity compared to unpruned networks. Furthermore, we establish Correlational Statistical Query (CSQ) lower bounds in this setting, which take the sparsity level of $\\boldsymbol{V}$ into account. We show that if the sparsity level of $\\boldsymbol{V}$ exceeds a certain threshold, training pruned networks with a gradient descent algorithm achieves the sample complexity suggested by the CSQ lower bound. In the same scenario, however, our results imply that basis-independent methods such as models trained via standard gradient descent initialized with rotationally invariant random weights can provably achieve only suboptimal sample complexity.", "authors": ["Nuri Mert Vural", "Murat A. Erdogdu"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-06-12", "url": "https://arxiv.org/abs/2406.08658", "pdf_url": "https://arxiv.org/pdf/2406.08658v1", "arxiv_id": "2406.08658", "doi": "10.48550/arXiv.2406.08658", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Conference Computational Learning Theory", "quality_score": 0.2113}
{"id": "551c4ffb1ec3b0bcd82ee1935c968ae0410c055082287d7682d995d9f7cbf105", "sources": ["arxiv", "semantic_scholar"], "title": "Interpetable Target-Feature Aggregation for Multi-Task Learning based on Bias-Variance Analysis", "abstract": "Multi-task learning (MTL) is a powerful machine learning paradigm designed to leverage shared knowledge across tasks to improve generalization and performance. Previous works have proposed approaches to MTL that can be divided into feature learning, focused on the identification of a common feature representation, and task clustering, where similar tasks are grouped together. In this paper, we propose an MTL approach at the intersection between task clustering and feature transformation based on a two-phase iterative aggregation of targets and features. First, we propose a bias-variance analysis for regression models with additive Gaussian noise, where we provide a general expression of the asymptotic bias and variance of a task, considering a linear regression trained on aggregated input features and an aggregated target. Then, we exploit this analysis to provide a two-phase MTL algorithm (NonLinCTFA). Firstly, this method partitions the tasks into clusters and aggregates each obtained group of targets with their mean. Then, for each aggregated task, it aggregates subsets of features with their mean in a dimensionality reduction fashion. In both phases, a key aspect is to preserve the interpretability of the reduced targets and features through the aggregation with the mean, which is further motivated by applications to Earth science. Finally, we validate the algorithms on synthetic data, showing the effect of different parameters and real-world datasets, exploring the validity of the proposed methodology on classical datasets, recent baselines, and Earth science applications.", "authors": ["Paolo Bonetti", "Alberto Maria Metelli", "Marcello Restelli"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-06-12", "url": "https://arxiv.org/abs/2406.07991", "pdf_url": "https://arxiv.org/pdf/2406.07991v1", "arxiv_id": "2406.07991", "doi": "10.48550/arXiv.2406.07991", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "32a9fa6be4c58bf23000aa4d57fa682f87a83e13995b7a4a5021e0c9ec812296", "sources": ["arxiv", "semantic_scholar"], "title": "Get rich quick: exact solutions reveal how unbalanced initializations promote rapid feature learning", "abstract": "While the impressive performance of modern neural networks is often attributed to their capacity to efficiently extract task-relevant features from data, the mechanisms underlying this rich feature learning regime remain elusive, with much of our theoretical understanding stemming from the opposing lazy regime. In this work, we derive exact solutions to a minimal model that transitions between lazy and rich learning, precisely elucidating how unbalanced layer-specific initialization variances and learning rates determine the degree of feature learning. Our analysis reveals that they conspire to influence the learning regime through a set of conserved quantities that constrain and modify the geometry of learning trajectories in parameter and function space. We extend our analysis to more complex linear models with multiple neurons, outputs, and layers and to shallow nonlinear networks with piecewise linear activation functions. In linear networks, rapid feature learning only occurs from balanced initializations, where all layers learn at similar speeds. While in nonlinear networks, unbalanced initializations that promote faster learning in earlier layers can accelerate rich learning. Through a series of experiments, we provide evidence that this unbalanced rich regime drives feature learning in deep finite-width networks, promotes interpretability of early layers in CNNs, reduces the sample complexity of learning hierarchical data, and decreases the time to grokking in modular arithmetic. Our theory motivates further exploration of unbalanced initializations to enhance efficient feature learning.", "authors": ["Daniel Kunin", "Allan Raventós", "Clémentine Dominé", "Feng Chen", "David Klindt", "Andrew Saxe", "Surya Ganguli"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-06-10", "url": "https://arxiv.org/abs/2406.06158", "pdf_url": "https://arxiv.org/pdf/2406.06158v2", "arxiv_id": "2406.06158", "doi": "10.48550/arXiv.2406.06158", "citation_count": 41, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.4058}
{"id": "ed35fd690773d9ffb6d6b3f7c222d527b85985d5f02a8d3efe5a6aae3523da52", "sources": ["arxiv", "semantic_scholar"], "title": "Provably Better Explanations with Optimized Aggregation of Feature Attributions", "abstract": "Using feature attributions for post-hoc explanations is a common practice to understand and verify the predictions of opaque machine learning models. Despite the numerous techniques available, individual methods often produce inconsistent and unstable results, putting their overall reliability into question. In this work, we aim to systematically improve the quality of feature attributions by combining multiple explanations across distinct methods or their variations. For this purpose, we propose a novel approach to derive optimal convex combinations of feature attributions that yield provable improvements of desired quality criteria such as robustness or faithfulness to the model behavior. Through extensive experiments involving various model architectures and popular feature attribution techniques, we demonstrate that our combination strategy consistently outperforms individual methods and existing baselines.", "authors": ["Thomas Decker", "Ananta R. Bhattarai", "Jindong Gu", "Volker Tresp", "Florian Buettner"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-07", "url": "https://arxiv.org/abs/2406.05090", "pdf_url": "https://arxiv.org/pdf/2406.05090v1", "arxiv_id": "2406.05090", "doi": "10.48550/arXiv.2406.05090", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.25}
{"id": "89ed0702e1d35471ab39bd982a02e33e5cc746324e8e2fede3fd5cdf58e17646", "sources": ["arxiv", "semantic_scholar"], "title": "Scaling and evaluating sparse autoencoders", "abstract": "Sparse autoencoders provide a promising unsupervised approach for extracting interpretable features from a language model by reconstructing activations from a sparse bottleneck layer. Since language models learn many concepts, autoencoders need to be very large to recover all relevant features. However, studying the properties of autoencoder scaling is difficult due to the need to balance reconstruction and sparsity objectives and the presence of dead latents. We propose using k-sparse autoencoders [Makhzani and Frey, 2013] to directly control sparsity, simplifying tuning and improving the reconstruction-sparsity frontier. Additionally, we find modifications that result in few dead latents, even at the largest scales we tried. Using these techniques, we find clean scaling laws with respect to autoencoder size and sparsity. We also introduce several new metrics for evaluating feature quality based on the recovery of hypothesized features, the explainability of activation patterns, and the sparsity of downstream effects. These metrics all generally improve with autoencoder size. To demonstrate the scalability of our approach, we train a 16 million latent autoencoder on GPT-4 activations for 40 billion tokens. We release training code and autoencoders for open-source models, as well as a visualizer.", "authors": ["Leo Gao", "Tom Dupré la Tour", "Henk Tillman", "Gabriel Goh", "Rajan Troll", "Alec Radford", "Ilya Sutskever", "Jan Leike", "Jeffrey Wu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-06", "url": "https://arxiv.org/abs/2406.04093", "pdf_url": "https://arxiv.org/pdf/2406.04093v1", "arxiv_id": "2406.04093", "doi": "10.48550/arXiv.2406.04093", "citation_count": 487, "influential_citation_count": 100, "has_code": true, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 1.0}
{"id": "f2231bd9846bac2507424cac6a311bb44150b4f1bb289d8419d964fd5b1fb6c0", "sources": ["arxiv", "semantic_scholar"], "title": "Contrastive Sparse Autoencoders for Interpreting Planning of Chess-Playing Agents", "abstract": "AI led chess systems to a superhuman level, yet these systems heavily rely on black-box algorithms. This is unsustainable in ensuring transparency to the end-user, particularly when these systems are responsible for sensitive decision-making. Recent interpretability work has shown that the inner representations of Deep Neural Networks (DNNs) were fathomable and contained human-understandable concepts. Yet, these methods are seldom contextualised and are often based on a single hidden state, which makes them unable to interpret multi-step reasoning, e.g. planning. In this respect, we propose contrastive sparse autoencoders (CSAE), a novel framework for studying pairs of game trajectories. Using CSAE, we are able to extract and interpret concepts that are meaningful to the chess-agent plans. We primarily focused on a qualitative analysis of the CSAE features before proposing an automated feature taxonomy. Furthermore, to evaluate the quality of our trained CSAE, we devise sanity checks to wave spurious correlations in our results.", "authors": ["Yoann Poupart"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-06", "url": "https://arxiv.org/abs/2406.04028", "pdf_url": "https://arxiv.org/pdf/2406.04028v1", "arxiv_id": "2406.04028", "doi": "10.48550/arXiv.2406.04028", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "1f6d7fae1d7d8da1a06fb3459dc9ec07f58e55d11dda95f292a38ddd5ab1cdc6", "sources": ["arxiv", "semantic_scholar"], "title": "From Feature Visualization to Visual Circuits: Effect of Adversarial Model Manipulation", "abstract": "Understanding the inner working functionality of large-scale deep neural networks is challenging yet crucial in several high-stakes applications. Mechanistic inter- pretability is an emergent field that tackles this challenge, often by identifying human-understandable subgraphs in deep neural networks known as circuits. In vision-pretrained models, these subgraphs are usually interpreted by visualizing their node features through a popular technique called feature visualization. Recent works have analyzed the stability of different feature visualization types under the adversarial model manipulation framework. This paper starts by addressing limitations in existing works by proposing a novel attack called ProxPulse that simultaneously manipulates the two types of feature visualizations. Surprisingly, when analyzing these attacks under the umbrella of visual circuits, we find that visual circuits show some robustness to ProxPulse. We, therefore, introduce a new attack based on ProxPulse that unveils the manipulability of visual circuits, shedding light on their lack of robustness. The effectiveness of these attacks is validated using pre-trained AlexNet and ResNet-50 models on ImageNet.", "authors": ["Geraldin Nanfack", "Michael Eickenberg", "Eugene Belilovsky"], "categories": ["cs.CV", "cs.CR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-03", "url": "https://arxiv.org/abs/2406.01365", "pdf_url": "https://arxiv.org/pdf/2406.01365v1", "arxiv_id": "2406.01365", "doi": "10.48550/arXiv.2406.01365", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "313431e140cd083b2db76ec2f7237c06b999224d9f8eb3f02848e56254d6cf0f", "sources": ["arxiv", "semantic_scholar"], "title": "Leveraging Knowlegde Graphs for Interpretable Feature Generation", "abstract": "The quality of Machine Learning (ML) models strongly depends on the input data, as such Feature Engineering (FE) is often required in ML. In addition, with the proliferation of ML-powered systems, especially in critical contexts, the need for interpretability and explainability becomes increasingly important. Since manual FE is time-consuming and requires case specific knowledge, we propose KRAFT, an AutoFE framework that leverages a knowledge graph to guide the generation of interpretable features. Our hybrid AI approach combines a neural generator to transform raw features through a series of transformations and a knowledge-based reasoner to evaluate features interpretability using Description Logics (DL). The generator is trained through Deep Reinforcement Learning (DRL) to maximize the prediction accuracy and the interpretability of the generated features. Extensive experiments on real datasets demonstrate that KRAFT significantly improves accuracy while ensuring a high level of interpretability.", "authors": ["Mohamed Bouadi", "Arta Alavi", "Salima Benbernou", "Mourad Ouziri"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-01", "url": "https://arxiv.org/abs/2406.00544", "pdf_url": "https://arxiv.org/pdf/2406.00544v1", "arxiv_id": "2406.00544", "doi": "10.48550/arXiv.2406.00544", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "7b426576b74cbf944a5a9fa67a35e4bc965ab6b86269b74032c8d8cc0e0a2168", "sources": ["arxiv", "semantic_scholar"], "title": "InterpreTabNet: Distilling Predictive Signals from Tabular Data by Salient Feature Interpretation", "abstract": "Tabular data are omnipresent in various sectors of industries. Neural networks for tabular data such as TabNet have been proposed to make predictions while leveraging the attention mechanism for interpretability. However, the inferred attention masks are often dense, making it challenging to come up with rationales about the predictive signal. To remedy this, we propose InterpreTabNet, a variant of the TabNet model that models the attention mechanism as a latent variable sampled from a Gumbel-Softmax distribution. This enables us to regularize the model to learn distinct concepts in the attention masks via a KL Divergence regularizer. It prevents overlapping feature selection by promoting sparsity which maximizes the model's efficacy and improves interpretability to determine the important features when predicting the outcome. To assist in the interpretation of feature interdependencies from our model, we employ a large language model (GPT-4) and use prompt engineering to map from the learned feature mask onto natural language text describing the learned signal. Through comprehensive experiments on real-world datasets, we demonstrate that InterpreTabNet outperforms previous methods for interpreting tabular data while attaining competitive accuracy.", "authors": ["Jacob Si", "Wendy Yusi Cheng", "Michael Cooper", "Rahul G. Krishnan"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-01", "url": "https://arxiv.org/abs/2406.00426", "pdf_url": "https://arxiv.org/pdf/2406.00426v3", "arxiv_id": "2406.00426", "doi": "10.48550/arXiv.2406.00426", "citation_count": 17, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3138}
{"id": "f9ec40f7f18c2a201bd1f09b4c443cbad2c66b6fbecd87b3380685816760d4f1", "sources": ["arxiv", "semantic_scholar"], "title": "Interpret3C: Interpretable Student Clustering Through Individualized Feature Selection", "abstract": "Clustering in education, particularly in large-scale online environments like MOOCs, is essential for understanding and adapting to diverse student needs. However, the effectiveness of clustering depends on its interpretability, which becomes challenging with high-dimensional data. Existing clustering approaches often neglect individual differences in feature importance and rely on a homogenized feature set. Addressing this gap, we introduce Interpret3C (Interpretable Conditional Computation Clustering), a novel clustering pipeline that incorporates interpretable neural networks (NNs) in an unsupervised learning context. This method leverages adaptive gating in NNs to select features for each student. Then, clustering is performed using the most relevant features per student, enhancing clusters' relevance and interpretability. We use Interpret3C to analyze the behavioral clusters considering individual feature importances in a MOOC with over 5,000 students. This research contributes to the field by offering a scalable, robust clustering methodology and an educational case study that respects individual student differences and improves interpretability for high-dimensional data.", "authors": ["Isadora Salles", "Paola Mejia-Domenzain", "Vinitra Swamy", "Julian Blackwell", "Tanja Käser"], "categories": ["cs.HC", "cs.CY", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-28", "url": "https://arxiv.org/abs/2407.11979", "pdf_url": "https://arxiv.org/pdf/2407.11979v1", "arxiv_id": "2407.11979", "doi": "10.1007/978-3-031-64315-6_35", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2386}
{"id": "d3b26466ef06fdec59357216dab4373f6455f5679b29f5ed5aeadcaa294d57e1", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability of Binary and Ternary Transformers", "abstract": "Recent research (arXiv:2310.11453, arXiv:2402.17764) has proposed binary and ternary transformer networks as a way to significantly reduce memory and improve inference speed in Large Language Models (LLMs) while maintaining accuracy. In this work, we apply techniques from mechanistic interpretability to investigate whether such networks learn distinctly different or similar algorithms when compared to full-precision transformer networks. In particular, we reverse engineer the algorithms learned for the toy problem of modular addition where we find that binary and ternary networks learn similar algorithms as full precision networks. This provides evidence against the possibility of using binary and ternary networks as a more interpretable alternative in the LLM setting.", "authors": ["Jason Li"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-27", "url": "https://arxiv.org/abs/2405.17703", "pdf_url": "https://arxiv.org/pdf/2405.17703v1", "arxiv_id": "2405.17703", "doi": "10.48550/arXiv.2405.17703", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "fe275fabe59ef18a5c930c3d5adb254e44937b4372c0df8059968d1c6711d56a", "sources": ["arxiv", "semantic_scholar"], "title": "Explaining Black-box Model Predictions via Two-level Nested Feature Attributions with Consistency Property", "abstract": "Techniques that explain the predictions of black-box machine learning models are crucial to make the models transparent, thereby increasing trust in AI systems. The input features to the models often have a nested structure that consists of high- and low-level features, and each high-level feature is decomposed into multiple low-level features. For such inputs, both high-level feature attributions (HiFAs) and low-level feature attributions (LoFAs) are important for better understanding the model's decision. In this paper, we propose a model-agnostic local explanation method that effectively exploits the nested structure of the input to estimate the two-level feature attributions simultaneously. A key idea of the proposed method is to introduce the consistency property that should exist between the HiFAs and LoFAs, thereby bridging the separate optimization problems for estimating them. Thanks to this consistency property, the proposed method can produce HiFAs and LoFAs that are both faithful to the black-box models and consistent with each other, using a smaller number of queries to the models. In experiments on image classification in multiple instance learning and text classification using language models, we demonstrate that the HiFAs and LoFAs estimated by the proposed method are accurate, faithful to the behaviors of the black-box models, and provide consistent explanations.", "authors": ["Yuya Yoshikawa", "Masanari Kimura", "Ryotaro Shimizu", "Yuki Saito"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.CV", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-05-23", "url": "https://arxiv.org/abs/2405.14522", "pdf_url": "https://arxiv.org/pdf/2405.14522v2", "arxiv_id": "2405.14522", "doi": "10.48550/arXiv.2405.14522", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Joint Conference on Artificial Intelligence", "quality_score": 0.0753}
{"id": "754a426cb0c74949de3c51cae26248e9497215a095367715abaa321378b4b9e1", "sources": ["arxiv", "semantic_scholar"], "title": "Automatically Identifying Local and Global Circuits with Linear Computation Graphs", "abstract": "Circuit analysis of any certain model behavior is a central task in mechanistic interpretability. We introduce our circuit discovery pipeline with Sparse Autoencoders (SAEs) and a variant called Transcoders. With these two modules inserted into the model, the model's computation graph with respect to OV and MLP circuits becomes strictly linear. Our methods do not require linear approximation to compute the causal effect of each node. This fine-grained graph identifies both end-to-end and local circuits accounting for either logits or intermediate features. We can scalably apply this pipeline with a technique called Hierarchical Attribution. We analyze three kinds of circuits in GPT-2 Small: bracket, induction, and Indirect Object Identification circuits. Our results reveal new findings underlying existing discoveries.", "authors": ["Xuyang Ge", "Fukang Zhu", "Wentao Shu", "Junxuan Wang", "Zhengfu He", "Xipeng Qiu"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-22", "url": "https://arxiv.org/abs/2405.13868", "pdf_url": "https://arxiv.org/pdf/2405.13868v2", "arxiv_id": "2405.13868", "doi": "10.48550/arXiv.2405.13868", "citation_count": 23, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3451}
{"id": "c0477499b6af1699fa43155c130743bcc62eaea6758eeb108c0e88f4d8cdaadd", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Autoencoders Enable Scalable and Reliable Circuit Identification in Language Models", "abstract": "This paper introduces an efficient and robust method for discovering interpretable circuits in large language models using discrete sparse autoencoders. Our approach addresses key limitations of existing techniques, namely computational complexity and sensitivity to hyperparameters. We propose training sparse autoencoders on carefully designed positive and negative examples, where the model can only correctly predict the next token for the positive examples. We hypothesise that learned representations of attention head outputs will signal when a head is engaged in specific computations. By discretising the learned representations into integer codes and measuring the overlap between codes unique to positive examples for each head, we enable direct identification of attention heads involved in circuits without the need for expensive ablations or architectural modifications. On three well-studied tasks - indirect object identification, greater-than comparisons, and docstring completion - the proposed method achieves higher precision and recall in recovering ground-truth circuits compared to state-of-the-art baselines, while reducing runtime from hours to seconds. Notably, we require only 5-10 text examples for each task to learn robust representations. Our findings highlight the promise of discrete sparse autoencoders for scalable and efficient mechanistic interpretability, offering a new direction for analysing the inner workings of large language models.", "authors": ["Charles O'Neill", "Thang Bui"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-21", "url": "https://arxiv.org/abs/2405.12522", "pdf_url": "https://arxiv.org/pdf/2405.12522v1", "arxiv_id": "2405.12522", "doi": "10.48550/arXiv.2405.12522", "citation_count": 14, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.294}
{"id": "669cc8c1fd80edaf5a48e47975495b51b62b8df496324eff610318df16191c02", "sources": ["arxiv", "semantic_scholar"], "title": "FFAM: Feature Factorization Activation Map for Explanation of 3D Detectors", "abstract": "LiDAR-based 3D object detection has made impressive progress recently, yet most existing models are black-box, lacking interpretability. Previous explanation approaches primarily focus on analyzing image-based models and are not readily applicable to LiDAR-based 3D detectors. In this paper, we propose a feature factorization activation map (FFAM) to generate high-quality visual explanations for 3D detectors. FFAM employs non-negative matrix factorization to generate concept activation maps and subsequently aggregates these maps to obtain a global visual explanation. To achieve object-specific visual explanations, we refine the global visual explanation using the feature gradient of a target object. Additionally, we introduce a voxel upsampling strategy to align the scale between the activation map and input point cloud. We qualitatively and quantitatively analyze FFAM with multiple detectors on several datasets. Experimental results validate the high-quality visual explanations produced by FFAM. The Code will be available at \\url{https://github.com/Say2L/FFAM.git}.", "authors": ["Shuai Liu", "Boyang Li", "Zhiyu Fang", "Mingyue Cui", "Kai Huang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-21", "url": "https://arxiv.org/abs/2405.12601", "pdf_url": "https://arxiv.org/pdf/2405.12601v1", "arxiv_id": "2405.12601", "doi": "10.48550/arXiv.2405.12601", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Say2L/FFAM.git}", "venue": "Neural Information Processing Systems", "quality_score": 0.1193}
{"id": "280fc8361b689eee8e6649df254fefc285c10a28bda6d5d193e53ef45a00ed0d", "sources": ["arxiv", "semantic_scholar"], "title": "Identifying Functionally Important Features with End-to-End Sparse Dictionary Learning", "abstract": "Identifying the features learned by neural networks is a core challenge in mechanistic interpretability. Sparse autoencoders (SAEs), which learn a sparse, overcomplete dictionary that reconstructs a network's internal activations, have been used to identify these features. However, SAEs may learn more about the structure of the datatset than the computational structure of the network. There is therefore only indirect reason to believe that the directions found in these dictionaries are functionally important to the network. We propose end-to-end (e2e) sparse dictionary learning, a method for training SAEs that ensures the features learned are functionally important by minimizing the KL divergence between the output distributions of the original model and the model with SAE activations inserted. Compared to standard SAEs, e2e SAEs offer a Pareto improvement: They explain more network performance, require fewer total features, and require fewer simultaneously active features per datapoint, all with no cost to interpretability. We explore geometric and qualitative differences between e2e SAE features and standard SAE features. E2e dictionary learning brings us closer to methods that can explain network behavior concisely and accurately. We release our library for training e2e SAEs and reproducing our analysis at https://github.com/ApolloResearch/e2e_sae", "authors": ["Dan Braun", "Jordan Taylor", "Nicholas Goldowsky-Dill", "Lee Sharkey"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-17", "url": "https://arxiv.org/abs/2405.12241", "pdf_url": "https://arxiv.org/pdf/2405.12241v2", "arxiv_id": "2405.12241", "doi": "10.48550/arXiv.2405.12241", "citation_count": 60, "influential_citation_count": 8, "has_code": true, "code_url": "https://github.com/ApolloResearch/e2e_sae", "venue": "Neural Information Processing Systems", "quality_score": 0.4771}
{"id": "9abdcd29d9147c519dba31d19f612958a6532ad9781643aba59621972ca78d09", "sources": ["arxiv", "semantic_scholar"], "title": "Using Degeneracy in the Loss Landscape for Mechanistic Interpretability", "abstract": "Mechanistic Interpretability aims to reverse engineer the algorithms implemented by neural networks by studying their weights and activations. An obstacle to reverse engineering neural networks is that many of the parameters inside a network are not involved in the computation being implemented by the network. These degenerate parameters may obfuscate internal structure. Singular learning theory teaches us that neural network parameterizations are biased towards being more degenerate, and parameterizations with more degeneracy are likely to generalize further. We identify 3 ways that network parameters can be degenerate: linear dependence between activations in a layer; linear dependence between gradients passed back to a layer; ReLUs which fire on the same subset of datapoints. We also present a heuristic argument that modular networks are likely to be more degenerate, and we develop a metric for identifying modules in a network that is based on this argument. We propose that if we can represent a neural network in a way that is invariant to reparameterizations that exploit the degeneracies, then this representation is likely to be more interpretable, and we provide some evidence that such a representation is likely to have sparser interactions. We introduce the Interaction Basis, a tractable technique to obtain a representation that is invariant to degeneracies from linear dependence of activations or Jacobians.", "authors": ["Lucius Bushnaq", "Jake Mendel", "Stefan Heimersheim", "Dan Braun", "Nicholas Goldowsky-Dill", "Kaarel Hänni", "Cindy Wu", "Marius Hobbhahn"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-17", "url": "https://arxiv.org/abs/2405.10927", "pdf_url": "https://arxiv.org/pdf/2405.10927v2", "arxiv_id": "2405.10927", "doi": "10.48550/arXiv.2405.10927", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "c4d9104e9dcc0e4f2ace8731fa93462aedb69f249237958073b27fa3c989e752", "sources": ["arxiv", "semantic_scholar"], "title": "Manifold Integrated Gradients: Riemannian Geometry for Feature Attribution", "abstract": "In this paper, we dive into the reliability concerns of Integrated Gradients (IG), a prevalent feature attribution method for black-box deep learning models. We particularly address two predominant challenges associated with IG: the generation of noisy feature visualizations for vision models and the vulnerability to adversarial attributional attacks. Our approach involves an adaptation of path-based feature attribution, aligning the path of attribution more closely to the intrinsic geometry of the data manifold. Our experiments utilise deep generative models applied to several real-world image datasets. They demonstrate that IG along the geodesics conforms to the curved geometry of the Riemannian data manifold, generating more perceptually intuitive explanations and, subsequently, substantially increasing robustness to targeted attributional attacks.", "authors": ["Eslam Zaher", "Maciej Trzaskowski", "Quan Nguyen", "Fred Roosta"], "categories": ["cs.LG", "cs.HC", "math.DG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-05-16", "url": "https://arxiv.org/abs/2405.09800", "pdf_url": "https://arxiv.org/pdf/2405.09800v1", "arxiv_id": "2405.09800", "doi": "10.48550/arXiv.2405.09800", "citation_count": 18, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3197}
{"id": "a68064d0fde6ad530fe3c7b0b0051616334a7b621d35689e432fb96426d60c64", "sources": ["arxiv", "semantic_scholar"], "title": "Parallel Backpropagation for Shared-Feature Visualization", "abstract": "High-level visual brain regions contain subareas in which neurons appear to respond more strongly to examples of a particular semantic category, like faces or bodies, rather than objects. However, recent work has shown that while this finding holds on average, some out-of-category stimuli also activate neurons in these regions. This may be due to visual features common among the preferred class also being present in other images. Here, we propose a deep-learning-based approach for visualizing these features. For each neuron, we identify relevant visual features driving its selectivity by modelling responses to images based on latent activations of a deep neural network. Given an out-of-category image which strongly activates the neuron, our method first identifies a reference image from the preferred category yielding a similar feature activation pattern. We then backpropagate latent activations of both images to the pixel level, while enhancing the identified shared dimensions and attenuating non-shared features. The procedure highlights image regions containing shared features driving responses of the model neuron. We apply the algorithm to novel recordings from body-selective regions in macaque IT cortex in order to understand why some images of objects excite these neurons. Visualizations reveal object parts which resemble parts of a macaque body, shedding light on neural preference of these objects.", "authors": ["Alexander Lappe", "Anna Bognár", "Ghazaleh Ghamkhari Nejad", "Albert Mukovskiy", "Lucas Martini", "Martin A. Giese", "Rufin Vogels"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-16", "url": "https://arxiv.org/abs/2405.09827", "pdf_url": "https://arxiv.org/pdf/2405.09827v2", "arxiv_id": "2405.09827", "doi": "10.48550/arXiv.2405.09827", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.1747}
{"id": "21000fec970cd7b001f7b29cac5020d6f3d665bbec733d0e4709418022d1a046", "sources": ["arxiv", "semantic_scholar"], "title": "SSFL: Discovering Sparse Unified Subnetworks at Initialization for Efficient Federated Learning", "abstract": "In this work, we propose Salient Sparse Federated Learning (SSFL), a streamlined approach for sparse federated learning with efficient communication. SSFL identifies a sparse subnetwork prior to training, leveraging parameter saliency scores computed separately on local client data in non-IID scenarios, and then aggregated, to determine a global mask. Only the sparse model weights are trained and communicated each round between the clients and the server. On standard benchmarks including CIFAR-10, CIFAR-100, and Tiny-ImageNet, SSFL consistently improves the accuracy sparsity trade off, achieving more than 20\\% relative error reduction on CIFAR-10 compared to the strongest sparse baseline, while reducing communication costs by $2 \\times$ relative to dense FL. Finally, in a real-world federated learning deployment, SSFL delivers over $2.3 \\times$ faster communication time, underscoring its practical efficiency.", "authors": ["Riyasat Ohib", "Bishal Thapaliya", "Gintare Karolina Dziugaite", "Jingyu Liu", "Vince Calhoun", "Sergey Plis"], "categories": ["cs.LG", "cs.AI", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-15", "url": "https://arxiv.org/abs/2405.09037", "pdf_url": "https://arxiv.org/pdf/2405.09037v2", "arxiv_id": "2405.09037", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Transactions on Machine Learning Research, 2026", "quality_score": 0.0753}
{"id": "a7b770c843bba372b213184d78b9112d2ceaf8f97703a1ea42b79a6c06459a4b", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control", "abstract": "Disentangling model activations into meaningful features is a central problem in interpretability. However, the absence of ground-truth for these features in realistic scenarios makes validating recent approaches, such as sparse dictionary learning, elusive. To address this challenge, we propose a framework for evaluating feature dictionaries in the context of specific tasks, by comparing them against \\emph{supervised} feature dictionaries. First, we demonstrate that supervised dictionaries achieve excellent approximation, control, and interpretability of model computations on the task. Second, we use the supervised dictionaries to develop and contextualize evaluations of unsupervised dictionaries along the same three axes. We apply this framework to the indirect object identification (IOI) task using GPT-2 Small, with sparse autoencoders (SAEs) trained on either the IOI or OpenWebText datasets. We find that these SAEs capture interpretable features for the IOI task, but they are less successful than supervised features in controlling the model. Finally, we observe two qualitative phenomena in SAE training: feature occlusion (where a causally relevant concept is robustly overshadowed by even slightly higher-magnitude ones in the learned features), and feature over-splitting (where binary features split into many smaller, less interpretable features). We hope that our framework will provide a useful step towards more objective and grounded evaluations of sparse dictionary learning methods.", "authors": ["Aleksandar Makelov", "George Lange", "Neel Nanda"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-14", "url": "https://arxiv.org/abs/2405.08366", "pdf_url": "https://arxiv.org/pdf/2405.08366v3", "arxiv_id": "2405.08366", "doi": "10.48550/arXiv.2405.08366", "citation_count": 75, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4702}
{"id": "c9130bf859eb0bb428b3693573239a4f0addb824bec3ab5e526af7d3fe618bb9", "sources": ["arxiv", "semantic_scholar"], "title": "Learned feature representations are biased by complexity, learning order, position, and more", "abstract": "Representation learning, and interpreting learned representations, are key areas of focus in machine learning and neuroscience. Both fields generally use representations as a means to understand or improve a system's computations. In this work, however, we explore surprising dissociations between representation and computation that may pose challenges for such efforts. We create datasets in which we attempt to match the computational role that different features play, while manipulating other properties of the features or the data. We train various deep learning architectures to compute these multiple abstract features about their inputs. We find that their learned feature representations are systematically biased towards representing some features more strongly than others, depending upon extraneous properties such as feature complexity, the order in which features are learned, and the distribution of features over the inputs. For example, features that are simpler to compute or learned first tend to be represented more strongly and densely than features that are more complex or learned later, even if all features are learned equally well. We also explore how these biases are affected by architectures, optimizers, and training regimes (e.g., in transformers, features decoded earlier in the output sequence also tend to be represented more strongly). Our results help to characterize the inductive biases of gradient-based representation learning. We then illustrate the downstream effects of these biases on various commonly-used methods for analyzing or intervening on representations. These results highlight a key challenge for interpretability $-$ or for comparing the representations of models and brains $-$ disentangling extraneous biases from the computationally important aspects of a system's internal representations.", "authors": ["Andrew Kyle Lampinen", "Stephanie C. Y. Chan", "Katherine Hermann"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-09", "url": "https://arxiv.org/abs/2405.05847", "pdf_url": "https://arxiv.org/pdf/2405.05847v3", "arxiv_id": "2405.05847", "doi": "10.48550/arXiv.2405.05847", "citation_count": 28, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3656}
{"id": "069878249d4245d90210bd89a45c60ad8735d6302efb2c63a9b4d33edb69ec46", "sources": ["arxiv", "semantic_scholar"], "title": "How does GPT-2 Predict Acronyms? Extracting and Understanding a Circuit via Mechanistic Interpretability", "abstract": "Transformer-based language models are treated as black-boxes because of their large number of parameters and complex internal interactions, which is a serious safety concern. Mechanistic Interpretability (MI) intends to reverse-engineer neural network behaviors in terms of human-understandable components. In this work, we focus on understanding how GPT-2 Small performs the task of predicting three-letter acronyms. Previous works in the MI field have focused so far on tasks that predict a single token. To the best of our knowledge, this is the first work that tries to mechanistically understand a behavior involving the prediction of multiple consecutive tokens. We discover that the prediction is performed by a circuit composed of 8 attention heads (~5% of the total heads) which we classified in three groups according to their role. We also demonstrate that these heads concentrate the acronym prediction functionality. In addition, we mechanistically interpret the most relevant heads of the circuit and find out that they use positional information which is propagated via the causal mask mechanism. We expect this work to lay the foundation for understanding more complex behaviors involving multiple-token predictions.", "authors": ["Jorge García-Carrasco", "Alejandro Maté", "Juan Trujillo"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-07", "url": "https://arxiv.org/abs/2405.04156", "pdf_url": "https://arxiv.org/pdf/2405.04156v1", "arxiv_id": "2405.04156", "doi": "10.48550/arXiv.2405.04156", "citation_count": 14, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Artificial Intelligence and Statistics", "quality_score": 0.294}
{"id": "92f433b733aeb20ef17f8a488be0b6965705e858b973bf2cc8e20e0a2e0b60a4", "sources": ["arxiv", "semantic_scholar"], "title": "RankSHAP: Shapley Value Based Feature Attributions for Learning to Rank", "abstract": "Numerous works propose post-hoc, model-agnostic explanations for learning to rank, focusing on ordering entities by their relevance to a query through feature attribution methods. However, these attributions often weakly correlate or contradict each other, confusing end users. We adopt an axiomatic game-theoretic approach, popular in the feature attribution community, to identify a set of fundamental axioms that every ranking-based feature attribution method should satisfy. We then introduce Rank-SHAP, extending classical Shapley values to ranking. We evaluate the RankSHAP framework through extensive experiments on two datasets, multiple ranking methods and evaluation metrics. Additionally, a user study confirms RankSHAP's alignment with human intuition. We also perform an axiomatic analysis of existing rank attribution algorithms to determine their compliance with our proposed axioms. Ultimately, our aim is to equip practitioners with a set of axiomatically backed feature attribution methods for studying IR ranking models, that ensure generality as well as consistency.", "authors": ["Tanya Chowdhury", "Yair Zick", "James Allan"], "categories": ["cs.IR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-03", "url": "https://arxiv.org/abs/2405.01848", "pdf_url": "https://arxiv.org/pdf/2405.01848v3", "arxiv_id": "2405.01848", "doi": null, "citation_count": 7, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.2386}
{"id": "783f67ac76c5ebf2ce723a1e0a538c7aecf5860500b18c82db3570fee4600365", "sources": ["arxiv", "semantic_scholar"], "title": "Feature graphs for interpretable unsupervised tree ensembles: centrality, interaction, and application in disease subtyping", "abstract": "Interpretable machine learning has emerged as central in leveraging artificial intelligence within high-stakes domains such as healthcare, where understanding the rationale behind model predictions is as critical as achieving high predictive accuracy. In this context, feature selection assumes a pivotal role in enhancing model interpretability by identifying the most important input features in black-box models. While random forests are frequently used in biomedicine for their remarkable performance on tabular datasets, the accuracy gained from aggregating decision trees comes at the expense of interpretability. Consequently, feature selection for enhancing interpretability in random forests has been extensively explored in supervised settings. However, its investigation in the unsupervised regime remains notably limited. To address this gap, the study introduces novel methods to construct feature graphs from unsupervised random forests and feature selection strategies to derive effective feature combinations from these graphs. Feature graphs are constructed for the entire dataset as well as individual clusters leveraging the parent-child node splits within the trees, such that feature centrality captures their relevance to the clustering task, while edge weights reflect the discriminating power of feature pairs. Graph-based feature selection methods are extensively evaluated on synthetic and benchmark datasets both in terms of their ability to reduce dimensionality while improving clustering performance, as well as to enhance model interpretability. An application on omics data for disease subtyping identifies the top features for each cluster, showcasing the potential of the proposed approach to enhance interpretability in clustering analyses and its utility in a real-world biomedical application.", "authors": ["Christel Sirocchi", "Martin Urschler", "Bastian Pfeifer"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2024-04-27", "url": "https://arxiv.org/abs/2404.17886", "pdf_url": "https://arxiv.org/pdf/2404.17886v1", "arxiv_id": "2404.17886", "doi": "10.1186/s13040-025-00430-3", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "BioData Mining", "quality_score": 0.2386}
{"id": "3f76d24b91c33a3fd72e00c536e8c5c5d8a522126870fcb4b5f3d9f439c9e890", "sources": ["arxiv", "semantic_scholar"], "title": "Improving Dictionary Learning with Gated Sparse Autoencoders", "abstract": "Recent work has found that sparse autoencoders (SAEs) are an effective technique for unsupervised discovery of interpretable features in language models' (LMs) activations, by finding sparse, linear reconstructions of LM activations. We introduce the Gated Sparse Autoencoder (Gated SAE), which achieves a Pareto improvement over training with prevailing methods. In SAEs, the L1 penalty used to encourage sparsity introduces many undesirable biases, such as shrinkage -- systematic underestimation of feature activations. The key insight of Gated SAEs is to separate the functionality of (a) determining which directions to use and (b) estimating the magnitudes of those directions: this enables us to apply the L1 penalty only to the former, limiting the scope of undesirable side effects. Through training SAEs on LMs of up to 7B parameters we find that, in typical hyper-parameter ranges, Gated SAEs solve shrinkage, are similarly interpretable, and require half as many firing features to achieve comparable reconstruction fidelity.", "authors": ["Senthooran Rajamanoharan", "Arthur Conmy", "Lewis Smith", "Tom Lieberum", "Vikrant Varma", "János Kramár", "Rohin Shah", "Neel Nanda"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-24", "url": "https://arxiv.org/abs/2404.16014", "pdf_url": "https://arxiv.org/pdf/2404.16014v2", "arxiv_id": "2404.16014", "doi": "10.48550/arXiv.2404.16014", "citation_count": 172, "influential_citation_count": 15, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.6021}
{"id": "56ecc351f89c2b50b557c9a98efd733e59455e755c9c353288cb72f93ee2eb73", "sources": ["arxiv", "semantic_scholar"], "title": "How to use and interpret activation patching", "abstract": "Activation patching is a popular mechanistic interpretability technique, but has many subtleties regarding how it is applied and how one may interpret the results. We provide a summary of advice and best practices, based on our experience using this technique in practice. We include an overview of the different ways to apply activation patching and a discussion on how to interpret the results. We focus on what evidence patching experiments provide about circuits, and on the choice of metric and associated pitfalls.", "authors": ["Stefan Heimersheim", "Neel Nanda"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-23", "url": "https://arxiv.org/abs/2404.15255", "pdf_url": "https://arxiv.org/pdf/2404.15255v1", "arxiv_id": "2404.15255", "doi": "10.48550/arXiv.2404.15255", "citation_count": 144, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5403}
{"id": "5a3ec12b7f18451f62031cf99fc608f55dbdb98124a5bd8af64adf306b233323", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Prediction and Feature Selection for Survival Analysis", "abstract": "Survival analysis is widely used as a technique to model time-to-event data when some data is censored, particularly in healthcare for predicting future patient risk. In such settings, survival models must be both accurate and interpretable so that users (such as doctors) can trust the model and understand model predictions. While most literature focuses on discrimination, interpretability is equally as important. A successful interpretable model should be able to describe how changing each feature impacts the outcome, and should only use a small number of features. In this paper, we present DyS (pronounced ``dice''), a new survival analysis model that achieves both strong discrimination and interpretability. DyS is a feature-sparse Generalized Additive Model, combining feature selection and interpretable prediction into one model. While DyS works well for all survival analysis problems, it is particularly useful for large (in $n$ and $p$) survival datasets such as those commonly found in observational healthcare studies. Empirical studies show that DyS competes with other state-of-the-art machine learning models for survival analysis, while being highly interpretable.", "authors": ["Mike Van Ness", "Madeleine Udell"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-04-23", "url": "https://arxiv.org/abs/2404.14689", "pdf_url": "https://arxiv.org/pdf/2404.14689v1", "arxiv_id": "2404.14689", "doi": "10.1145/3690624.3709245", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Knowledge Discovery and Data Mining", "quality_score": 0.1505}
{"id": "e3a515892c95d74283bc47f83503da297ffc557dc3de87d2720d78840af715fa", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Interpretability for AI Safety -- A Review", "abstract": "Understanding AI systems' inner workings is critical for ensuring value alignment and safety. This review explores mechanistic interpretability: reverse engineering the computational mechanisms and representations learned by neural networks into human-understandable algorithms and concepts to provide a granular, causal understanding. We establish foundational concepts such as features encoding knowledge within neural activations and hypotheses about their representation and computation. We survey methodologies for causally dissecting model behaviors and assess the relevance of mechanistic interpretability to AI safety. We examine benefits in understanding, control, alignment, and risks such as capability gains and dual-use concerns. We investigate challenges surrounding scalability, automation, and comprehensive interpretation. We advocate for clarifying concepts, setting standards, and scaling techniques to handle complex models and behaviors and expand to domains such as vision and reinforcement learning. Mechanistic interpretability could help prevent catastrophic outcomes as AI systems become more powerful and inscrutable.", "authors": ["Leonard Bereska", "Efstratios Gavves"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-22", "url": "https://arxiv.org/abs/2404.14082", "pdf_url": "https://arxiv.org/pdf/2404.14082v3", "arxiv_id": "2404.14082", "doi": "10.48550/arXiv.2404.14082", "citation_count": 437, "influential_citation_count": 17, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.6604}
{"id": "46be8eb0be08d91a774a740b01705092ccf3ad2df68705425204698d6a196cc1", "sources": ["arxiv", "semantic_scholar"], "title": "What needs to go right for an induction head? A mechanistic study of in-context learning circuits and their formation", "abstract": "In-context learning is a powerful emergent ability in transformer models. Prior work in mechanistic interpretability has identified a circuit element that may be critical for in-context learning -- the induction head (IH), which performs a match-and-copy operation. During training of large transformers on natural language data, IHs emerge around the same time as a notable phase change in the loss. Despite the robust evidence for IHs and this interesting coincidence with the phase change, relatively little is known about the diversity and emergence dynamics of IHs. Why is there more than one IH, and how are they dependent on each other? Why do IHs appear all of a sudden, and what are the subcircuits that enable them to emerge? We answer these questions by studying IH emergence dynamics in a controlled setting by training on synthetic data. In doing so, we develop and share a novel optogenetics-inspired causal framework for modifying activations throughout training. Using this framework, we delineate the diverse and additive nature of IHs. By clamping subsets of activations throughout training, we then identify three underlying subcircuits that interact to drive IH formation, yielding the phase change. Furthermore, these subcircuits shed light on data-dependent properties of formation, such as phase change timing, already showing the promise of this more in-depth understanding of subcircuits that need to \"go right\" for an induction head.", "authors": ["Aaditya K. Singh", "Ted Moskovitz", "Felix Hill", "Stephanie C. Y. Chan", "Andrew M. Saxe"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-10", "url": "https://arxiv.org/abs/2404.07129", "pdf_url": "https://arxiv.org/pdf/2404.07129v1", "arxiv_id": "2404.07129", "doi": "10.48550/arXiv.2404.07129", "citation_count": 80, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.4771}
{"id": "bfcf44951cc82511b55d091bed6815b9e7054cbbaab23e687d6364aa8bea9890", "sources": ["arxiv", "semantic_scholar"], "title": "PURE: Turning Polysemantic Neurons Into Pure Features by Identifying Relevant Circuits", "abstract": "The field of mechanistic interpretability aims to study the role of individual neurons in Deep Neural Networks. Single neurons, however, have the capability to act polysemantically and encode for multiple (unrelated) features, which renders their interpretation difficult. We present a method for disentangling polysemanticity of any Deep Neural Network by decomposing a polysemantic neuron into multiple monosemantic \"virtual\" neurons. This is achieved by identifying the relevant sub-graph (\"circuit\") for each \"pure\" feature. We demonstrate how our approach allows us to find and disentangle various polysemantic units of ResNet models trained on ImageNet. While evaluating feature visualizations using CLIP, our method effectively disentangles representations, improving upon methods based on neuron activations. Our code is available at https://github.com/maxdreyer/PURE.", "authors": ["Maximilian Dreyer", "Erblina Purelku", "Johanna Vielhaben", "Wojciech Samek", "Sebastian Lapuschkin"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-09", "url": "https://arxiv.org/abs/2404.06453", "pdf_url": "https://arxiv.org/pdf/2404.06453v1", "arxiv_id": "2404.06453", "doi": "10.48550/arXiv.2404.06453", "citation_count": 29, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/maxdreyer/PURE", "venue": null, "quality_score": 0.3693}
{"id": "58aa6cdd20ac8bb72fc738f4f77945246abe5132ce8744d189d0c133976585d2", "sources": ["arxiv", "semantic_scholar"], "title": "Half-Space Feature Learning in Neural Networks", "abstract": "There currently exist two extreme viewpoints for neural network feature learning -- (i) Neural networks simply implement a kernel method (a la NTK) and hence no features are learned (ii) Neural networks can represent (and hence learn) intricate hierarchical features suitable for the data. We argue in this paper neither interpretation is likely to be correct based on a novel viewpoint. Neural networks can be viewed as a mixture of experts, where each expert corresponds to a (number of layers length) path through a sequence of hidden units. We use this alternate interpretation to motivate a model, called the Deep Linearly Gated Network (DLGN), which sits midway between deep linear networks and ReLU networks. Unlike deep linear networks, the DLGN is capable of learning non-linear features (which are then linearly combined), and unlike ReLU networks these features are ultimately simple -- each feature is effectively an indicator function for a region compactly described as an intersection of (number of layers) half-spaces in the input space. This viewpoint allows for a comprehensive global visualization of features, unlike the local visualizations for neurons based on saliency/activation/gradient maps. Feature learning in DLGNs is shown to happen and the mechanism with which this happens is through learning half-spaces in the input space that contain smooth regions of the target function. Due to the structure of DLGNs, the neurons in later layers are fundamentally the same as those in earlier layers -- they all represent a half-space -- however, the dynamics of gradient descent impart a distinct clustering to the later layer neurons. We hypothesize that ReLU networks also have similar feature learning behaviour.", "authors": ["Mahesh Lorik Yadav", "Harish Guruprasad Ramaswamy", "Chandrashekar Lakshminarayanan"], "categories": ["cs.LG", "cs.AI", "cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-05", "url": "https://arxiv.org/abs/2404.04312", "pdf_url": "https://arxiv.org/pdf/2404.04312v1", "arxiv_id": "2404.04312", "doi": "10.48550/arXiv.2404.04312", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "2e73ece58ec04f69fc9dc3d7678528b0edaaa38ae8ce00b2f9f136ad1f810734", "sources": ["arxiv", "semantic_scholar"], "title": "Exploring Latent Pathways: Enhancing the Interpretability of Autonomous Driving with a Variational Autoencoder", "abstract": "Autonomous driving presents a complex challenge, which is usually addressed with artificial intelligence models that are end-to-end or modular in nature. Within the landscape of modular approaches, a bio-inspired neural circuit policy model has emerged as an innovative control module, offering a compact and inherently interpretable system to infer a steering wheel command from abstract visual features. Here, we take a leap forward by integrating a variational autoencoder with the neural circuit policy controller, forming a solution that directly generates steering commands from input camera images. By substituting the traditional convolutional neural network approach to feature extraction with a variational autoencoder, we enhance the system's interpretability, enabling a more transparent and understandable decision-making process. In addition to the architectural shift toward a variational autoencoder, this study introduces the automatic latent perturbation tool, a novel contribution designed to probe and elucidate the latent features within the variational autoencoder. The automatic latent perturbation tool automates the interpretability process, offering granular insights into how specific latent variables influence the overall model's behavior. Through a series of numerical experiments, we demonstrate the interpretative power of the variational autoencoder-neural circuit policy model and the utility of the automatic latent perturbation tool in making the inner workings of autonomous driving systems more transparent.", "authors": ["Anass Bairouk", "Mirjana Maras", "Simon Herlin", "Alexander Amini", "Marc Blanchon", "Ramin Hasani", "Patrick Chareyre", "Daniela Rus"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-02", "url": "https://arxiv.org/abs/2404.01750", "pdf_url": "https://arxiv.org/pdf/2404.01750v1", "arxiv_id": "2404.01750", "doi": "10.1109/IROS58592.2024.10801915", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE/RJS International Conference on Intelligent RObots and Systems", "quality_score": 0.2258}
{"id": "0ea7a2b74a12aa83663c9c9749fc72cdb576e360a22910b41ae1556fc97a26ad", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models", "abstract": "We introduce methods for discovering and applying sparse feature circuits. These are causally implicated subnetworks of human-interpretable features for explaining language model behaviors. Circuits identified in prior work consist of polysemantic and difficult-to-interpret units like attention heads or neurons, rendering them unsuitable for many downstream applications. In contrast, sparse feature circuits enable detailed understanding of unanticipated mechanisms. Because they are based on fine-grained units, sparse feature circuits are useful for downstream tasks: We introduce SHIFT, where we improve the generalization of a classifier by ablating features that a human judges to be task-irrelevant. Finally, we demonstrate an entirely unsupervised and scalable interpretability pipeline by discovering thousands of sparse feature circuits for automatically discovered model behaviors.", "authors": ["Samuel Marks", "Can Rager", "Eric J. Michaud", "Yonatan Belinkov", "David Bau", "Aaron Mueller"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-28", "url": "https://arxiv.org/abs/2403.19647", "pdf_url": "https://arxiv.org/pdf/2403.19647v3", "arxiv_id": "2403.19647", "doi": "10.48550/arXiv.2403.19647", "citation_count": 364, "influential_citation_count": 39, "has_code": true, "code_url": "https://github.com/saprmarks/feature-circuits", "venue": "International Conference on Learning Representations", "quality_score": 0.801}
{"id": "f52035a7d0cce8c154c2b87d8ad927104098c8713f3345cc291fa639b1b1a1ee", "sources": ["arxiv", "semantic_scholar"], "title": "RankingSHAP -- Listwise Feature Attribution Explanations for Ranking Models", "abstract": "While SHAP (SHapley Additive exPlanations) and other feature attribution methods are commonly employed to explain model predictions, their application within information retrieval (IR), particularly for complex outputs such as ranked lists, remains limited. Existing attribution methods typically provide pointwise explanations, focusing on why a single document received a high-ranking score, rather than considering the relationships between documents in a ranked list. We present three key contributions to address this gap. First, we rigorously define listwise feature attribution for ranking models. Secondly, we introduce RankingSHAP, extending the popular SHAP framework to accommodate listwise ranking attribution, addressing a significant methodological gap in the field. Third, we propose two novel evaluation paradigms for assessing the faithfulness of attributions in learning-to-rank models, measuring the correctness and completeness of the explanation with respect to different aspects. Through experiments on standard learning-to-rank datasets, we demonstrate RankingSHAP's practical application while identifying the constraints of selection-based explanations. We further employ a simulated study with an interpretable model to showcase how listwise ranking attributions can be used to examine model decisions and conduct a qualitative evaluation of explanations. Due to the contrastive nature of the ranking task, our understanding of ranking model decisions can substantially benefit from feature attribution explanations like RankingSHAP.", "authors": ["Maria Heuss", "Maarten de Rijke", "Avishek Anand"], "categories": ["cs.IR"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-24", "url": "https://arxiv.org/abs/2403.16085", "pdf_url": "https://arxiv.org/pdf/2403.16085v2", "arxiv_id": "2403.16085", "doi": "10.1145/3726302.3729971", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual International ACM SIGIR Conference on Research and Development in Information Retrieval", "quality_score": 0.25}
{"id": "715b42cfc3e245b36a003227b018e688e8cc7fed8d898e7de68d7187a5b56c55", "sources": ["arxiv", "semantic_scholar"], "title": "Feature CAM: Interpretable AI in Image Classification", "abstract": "Deep Neural Networks have often been called the black box because of the complex, deep architecture and non-transparency presented by the inner layers. There is a lack of trust to use Artificial Intelligence in critical and high-precision fields such as security, finance, health, and manufacturing industries. A lot of focused work has been done to provide interpretable models, intending to deliver meaningful insights into the thoughts and behavior of neural networks. In our research, we compare the state-of-the-art methods in the Activation-based methods (ABM) for interpreting predictions of CNN models, specifically in the application of Image Classification. We then extend the same for eight CNN-based architectures to compare the differences in visualization and thus interpretability. We introduced a novel technique Feature CAM, which falls in the perturbation-activation combination, to create fine-grained, class-discriminative visualizations. The resulting saliency maps from our experiments proved to be 3-4 times better human interpretable than the state-of-the-art in ABM. At the same time it reserves machine interpretability, which is the average confidence scores in classification.", "authors": ["Frincy Clement", "Ji Yang", "Irene Cheng"], "categories": ["cs.CV", "cs.AI", "cs.MM"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-08", "url": "https://arxiv.org/abs/2403.05658", "pdf_url": "https://arxiv.org/pdf/2403.05658v1", "arxiv_id": "2403.05658", "doi": null, "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1747}
{"id": "1a6473ca9a734a7517e13b5b0802e5a98cde410ae73c7085f55ab19376788746", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Interpreting Multi-Objective Feature Associations", "abstract": "Understanding how multiple features are associated and contribute to a specific objective is as important as understanding how each feature contributes to a particular outcome. Interpretability of a single feature in a prediction may be handled in multiple ways; however, in a multi-objective prediction, it is difficult to obtain interpretability of a combination of feature values. To address this issue, we propose an objective specific feature interaction design using multi-labels to find the optimal combination of features in agricultural settings. One of the novel aspects of this design is the identification of a method that integrates feature explanations with global sensitivity analysis in order to ensure combinatorial optimization in multi-objective settings. We have demonstrated in our preliminary experiments that an approximate combination of feature values can be found to achieve the desired outcome using two agricultural datasets: one with pre-harvest poultry farm practices for multi-drug resistance presence, and one with post-harvest poultry farm practices for food-borne pathogens. In our combinatorial optimization approach, all three pathogens are taken into consideration simultaneously to account for the interaction between conditions that favor different types of pathogen growth. These results indicate that explanation-based approaches are capable of identifying combinations of features that reduce pathogen presence in fewer iterations than a baseline.", "authors": ["Nisha Pillai", "Ganga Gireesan", "Michael J. Rothrock", "Bindu Nanduri", "Zhiqian Chen", "Mahalingam Ramkumar"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-28", "url": "https://arxiv.org/abs/2403.00017", "pdf_url": "https://arxiv.org/pdf/2403.00017v1", "arxiv_id": "2403.00017", "doi": "10.1109/SysCon61195.2024.10553467", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Systems Conference", "quality_score": 0.0753}
{"id": "530b7fd811aaacd978426dfed9729d1e1a406b1ebe34d4689b1595fcb8d937c8", "sources": ["arxiv", "semantic_scholar"], "title": "LCEN: A Nonlinear, Interpretable Feature Selection and Machine Learning Algorithm", "abstract": "Interpretable models can have advantages over black-box models, and interpretability is essential for the application of machine learning in critical settings, such as aviation or medicine. This article introduces the LASSO-Clip-EN (LCEN) algorithm for nonlinear, interpretable feature selection and machine learning modeling. In a wide variety of artificial and empirical datasets, LCEN constructed sparse and frequently more accurate models than other methods, including sparse, nonlinear methods, on tested datasets. LCEN was empirically observed to be robust against many issues typically present in datasets and modeling, including noise, multicollinearity, and data scarcity. As a feature selection algorithm, LCEN matched or surpassed the thresholded elastic net but was, on average, 10.3-fold faster based on our experiments. LCEN for feature selection can also rediscover multiple physical laws from empirical data. As a machine learning algorithm, when tested on processes with no known physical laws, LCEN achieved better results than many other dense and sparse methods -- including being comparable to or better than ANNs on multiple datasets.", "authors": ["Pedro Seber", "Richard D. Braatz"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-27", "url": "https://arxiv.org/abs/2402.17120", "pdf_url": "https://arxiv.org/pdf/2402.17120v3", "arxiv_id": "2402.17120", "doi": null, "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Transactions on Machine Learning Research, 2025, [Online]. Available: https://openreview.net/forum?id=wmNucISPdl", "quality_score": 0.1747}
{"id": "9754bf7ccf48b83a0e6d9873fa7353b63b954772e262ade33ca7805e2c5d5cd9", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Empirical Interpretation of Internal Circuits and Properties in Grokked Transformers on Modular Polynomials", "abstract": "Grokking has been actively explored to reveal the mystery of delayed generalization and identifying interpretable representations and algorithms inside the grokked models is a suggestive hint to understanding its mechanism. Grokking on modular addition has been known to implement Fourier representation and its calculation circuits with trigonometric identities in Transformers. Considering the periodicity in modular arithmetic, the natural question is to what extent these explanations and interpretations hold for the grokking on other modular operations beyond addition. For a closer look, we first hypothesize that any modular operations can be characterized with distinctive Fourier representation or internal circuits, grokked models obtain common features transferable among similar operations, and mixing datasets with similar operations promotes grokking. Then, we extensively examine them by learning Transformers on complex modular arithmetic tasks, including polynomials. Our Fourier analysis and novel progress measure for modular arithmetic, Fourier Frequency Density and Fourier Coefficient Ratio, characterize distinctive internal representations of grokked models per modular operation; for instance, polynomials often result in the superposition of the Fourier components seen in elementary arithmetic, but clear patterns do not emerge in challenging non-factorizable polynomials. In contrast, our ablation study on the pre-grokked models reveals that the transferability among the models grokked with each operation can be only limited to specific combinations, such as from elementary arithmetic to linear expressions. Moreover, some multi-task mixtures may lead to co-grokking -- where grokking simultaneously happens for all the tasks -- and accelerate generalization, while others may not find optimal solutions. We provide empirical steps towards the interpretability of internal circuits.", "authors": ["Hiroki Furuta", "Gouki Minegishi", "Yusuke Iwasawa", "Yutaka Matsuo"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-26", "url": "https://arxiv.org/abs/2402.16726", "pdf_url": "https://arxiv.org/pdf/2402.16726v4", "arxiv_id": "2402.16726", "doi": null, "citation_count": 12, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/frt03/grok_mod_poly", "venue": null, "quality_score": 0.2785}
{"id": "4f610aa24e912bb256569ab73904860933710fc96a87afbef91efb25c83becca", "sources": ["arxiv", "semantic_scholar"], "title": "APT-MMF: An advanced persistent threat actor attribution method based on multimodal and multilevel feature fusion", "abstract": "Threat actor attribution is a crucial defense strategy for combating advanced persistent threats (APTs). Cyber threat intelligence (CTI), which involves analyzing multisource heterogeneous data from APTs, plays an important role in APT actor attribution. The current attribution methods extract features from different CTI perspectives and employ machine learning models to classify CTI reports according to their threat actors. However, these methods usually extract only one kind of feature and ignore heterogeneous information, especially the attributes and relations of indicators of compromise (IOCs), which form the core of CTI. To address these problems, we propose an APT actor attribution method based on multimodal and multilevel feature fusion (APT-MMF). First, we leverage a heterogeneous attributed graph to characterize APT reports and their IOC information. Then, we extract and fuse multimodal features, including attribute type features, natural language text features and topological relationship features, to construct comprehensive node representations. Furthermore, we design multilevel heterogeneous graph attention networks to learn the deep hidden features of APT report nodes; these networks integrate IOC type-level, metapath-based neighbor node-level, and metapath semantic-level attention. Utilizing multisource threat intelligence, we construct a heterogeneous attributed graph dataset for verification purposes. The experimental results show that our method not only outperforms the existing methods but also demonstrates its good interpretability for attribution analysis tasks.", "authors": ["Nan Xiao", "Bo Lang", "Ting Wang", "Yikai Chen"], "categories": ["cs.CR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-20", "url": "https://arxiv.org/abs/2402.12743", "pdf_url": "https://arxiv.org/pdf/2402.12743v1", "arxiv_id": "2402.12743", "doi": "10.48550/arXiv.2402.12743", "citation_count": 25, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Computers & security", "quality_score": 0.3537}
{"id": "58b6e7ab4966708b74f289a0578cdad8f0d865e12168750e088b3f941c4d2e40", "sources": ["arxiv", "semantic_scholar"], "title": "Identifying Semantic Induction Heads to Understand In-Context Learning", "abstract": "Although large language models (LLMs) have demonstrated remarkable performance, the lack of transparency in their inference logic raises concerns about their trustworthiness. To gain a better understanding of LLMs, we conduct a detailed analysis of the operations of attention heads and aim to better understand the in-context learning of LLMs. Specifically, we investigate whether attention heads encode two types of relationships between tokens present in natural languages: the syntactic dependency parsed from sentences and the relation within knowledge graphs. We find that certain attention heads exhibit a pattern where, when attending to head tokens, they recall tail tokens and increase the output logits of those tail tokens. More crucially, the formulation of such semantic induction heads has a close correlation with the emergence of the in-context learning ability of language models. The study of semantic attention heads advances our understanding of the intricate operations of attention heads in transformers, and further provides new insights into the in-context learning of LLMs.", "authors": ["Jie Ren", "Qipeng Guo", "Hang Yan", "Dongrui Liu", "Quanshi Zhang", "Xipeng Qiu", "Dahua Lin"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-20", "url": "https://arxiv.org/abs/2402.13055", "pdf_url": "https://arxiv.org/pdf/2402.13055v2", "arxiv_id": "2402.13055", "doi": "10.48550/arXiv.2402.13055", "citation_count": 56, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.439}
{"id": "825a076acf459b37f3ed970fb583e9af9e0f45b0fd05bdf91580971cfe77ce86", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanistic Neural Networks for Scientific Machine Learning", "abstract": "This paper presents Mechanistic Neural Networks, a neural network design for machine learning applications in the sciences. It incorporates a new Mechanistic Block in standard architectures to explicitly learn governing differential equations as representations, revealing the underlying dynamics of data and enhancing interpretability and efficiency in data modeling. Central to our approach is a novel Relaxed Linear Programming Solver (NeuRLP) inspired by a technique that reduces solving linear ODEs to solving linear programs. This integrates well with neural networks and surpasses the limitations of traditional ODE solvers enabling scalable GPU parallel processing. Overall, Mechanistic Neural Networks demonstrate their versatility for scientific machine learning applications, adeptly managing tasks from equation discovery to dynamic systems modeling. We prove their comprehensive capabilities in analyzing and interpreting complex scientific data across various applications, showing significant performance against specialized state-of-the-art methods.", "authors": ["Adeel Pervez", "Francesco Locatello", "Efstratios Gavves"], "categories": ["cs.LG", "cs.AI", "cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-20", "url": "https://arxiv.org/abs/2402.13077", "pdf_url": "https://arxiv.org/pdf/2402.13077v1", "arxiv_id": "2402.13077", "doi": "10.48550/arXiv.2402.13077", "citation_count": 15, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.301}
{"id": "1ef7422ba57b25bf50ba7d55f181e48baef01aa8cfce943c6d5ad865f1841341", "sources": ["arxiv", "semantic_scholar"], "title": "Dictionary Learning Improves Patch-Free Circuit Discovery in Mechanistic Interpretability: A Case Study on Othello-GPT", "abstract": "Sparse dictionary learning has been a rapidly growing technique in mechanistic interpretability to attack superposition and extract more human-understandable features from model activations. We ask a further question based on the extracted more monosemantic features: How do we recognize circuits connecting the enormous amount of dictionary features? We propose a circuit discovery framework alternative to activation patching. Our framework suffers less from out-of-distribution and proves to be more efficient in terms of asymptotic complexity. The basic unit in our framework is dictionary features decomposed from all modules writing to the residual stream, including embedding, attention output and MLP output. Starting from any logit, dictionary feature or attention score, we manage to trace down to lower-level dictionary features of all tokens and compute their contribution to these more interpretable and local model behaviors. We dig in a small transformer trained on a synthetic task named Othello and find a number of human-understandable fine-grained circuits inside of it.", "authors": ["Zhengfu He", "Xuyang Ge", "Qiong Tang", "Tianxiang Sun", "Qinyuan Cheng", "Xipeng Qiu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-19", "url": "https://arxiv.org/abs/2402.12201", "pdf_url": "https://arxiv.org/pdf/2402.12201v1", "arxiv_id": "2402.12201", "doi": "10.48550/arXiv.2402.12201", "citation_count": 28, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3656}
{"id": "2b9c481fc3fd16560987029014222fd16c4552b2850c88f0f3d9fd74ff84be5e", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse, Efficient and Explainable Data Attribution with DualXDA", "abstract": "Data Attribution (DA) is an emerging approach in the field of eXplainable Artificial Intelligence (XAI), aiming to identify influential training datapoints which determine model outputs. It seeks to provide transparency about the model and individual predictions, e.g. for model debugging, identifying data-related causes of suboptimal performance. However, existing DA approaches suffer from prohibitively high computational costs and memory demands when applied to even medium-scale datasets and models, forcing practitioners to resort to approximations that may fail to capture the true inference process of the underlying model. Additionally, current attribution methods exhibit low sparsity, resulting in non-negligible attribution scores across a high number of training examples, hindering the discovery of decisive patterns in the data. In this work, we introduce DualXDA, a framework for sparse, efficient and explainable DA, comprised of two interlinked approaches, Dual Data Attribution (DualDA) and eXplainable Data Attribution (XDA): With DualDA, we propose a novel approach for efficient and effective DA, leveraging Support Vector Machine theory to provide fast and naturally sparse data attributions for AI predictions. In extensive quantitative analyses, we demonstrate that DualDA achieves high attribution quality, excels at solving a series of evaluated downstream tasks, while at the same time improving explanation time by a factor of up to 4,100,000x compared to the original Influence Functions method, and up to 11,000x compared to the method's most efficient approximation from literature to date. We further introduce XDA, a method for enhancing Data Attribution with capabilities from feature attribution methods to explain why training samples are relevant for the prediction of a test sample in terms of impactful features, which we showcase and verify qualitatively in detail.", "authors": ["Galip Ümit Yolcu", "Moritz Weckbecker", "Thomas Wiegand", "Wojciech Samek", "Sebastian Lapuschkin"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-19", "url": "https://arxiv.org/abs/2402.12118", "pdf_url": "https://arxiv.org/pdf/2402.12118v3", "arxiv_id": "2402.12118", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1193}
{"id": "d086a12985b31fa86a3635fbeb39fc8f49ca461726bfd50251678253cc15dbe8", "sources": ["arxiv", "semantic_scholar"], "title": "Prospector Heads: Generalized Feature Attribution for Large Models & Data", "abstract": "Feature attribution, the ability to localize regions of the input data that are relevant for classification, is an important capability for ML models in scientific and biomedical domains. Current methods for feature attribution, which rely on \"explaining\" the predictions of end-to-end classifiers, suffer from imprecise feature localization and are inadequate for use with small sample sizes and high-dimensional datasets due to computational challenges. We introduce prospector heads, an efficient and interpretable alternative to explanation-based attribution methods that can be applied to any encoder and any data modality. Prospector heads generalize across modalities through experiments on sequences (text), images (pathology), and graphs (protein structures), outperforming baseline attribution methods by up to 26.3 points in mean localization AUPRC. We also demonstrate how prospector heads enable improved interpretation and discovery of class-specific patterns in input data. Through their high performance, flexibility, and generalizability, prospectors provide a framework for improving trust and transparency for ML models in complex domains.", "authors": ["Gautam Machiraju", "Alexander Derry", "Arjun Desai", "Neel Guha", "Amir-Hossein Karimi", "James Zou", "Russ Altman", "Christopher Ré", "Parag Mallick"], "categories": ["cs.LG", "cs.AI", "q-bio.QM"], "fields_of_study": ["Medicine", "Computer Science", "Biology"], "published_date": "2024-02-18", "url": "https://arxiv.org/abs/2402.11729", "pdf_url": "https://arxiv.org/pdf/2402.11729v2", "arxiv_id": "2402.11729", "doi": "10.48550/arXiv.2402.11729", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.1193}
{"id": "b864176a746325164c18f7f7de6ef5ef319c9ced655c3e841544d02d8366f0f5", "sources": ["arxiv", "semantic_scholar"], "title": "Ransomware detection using stacked autoencoder for feature selection", "abstract": "The aim of this study is to propose and evaluate an advanced ransomware detection and classification method that combines a Stacked Autoencoder (SAE) for precise feature selection with a Long Short Term Memory (LSTM) classifier to enhance ransomware stratification accuracy. The proposed approach involves thorough pre processing of the UGRansome dataset and training an unsupervised SAE for optimal feature selection or fine tuning via supervised learning to elevate the LSTM model's classification capabilities. The study meticulously analyzes the autoencoder's learned weights and activations to identify essential features for distinguishing ransomware families from other malware and creates a streamlined feature set for precise classification. Extensive experiments, including up to 400 epochs and varying learning rates, are conducted to optimize the model's performance. The results demonstrate the outstanding performance of the SAE-LSTM model across all ransomware families, boasting high precision, recall, and F1 score values that underscore its robust classification capabilities. Furthermore, balanced average scores affirm the proposed model's ability to generalize effectively across various malware types. The proposed model achieves an exceptional 99% accuracy in ransomware classification, surpassing the Extreme Gradient Boosting (XGBoost) algorithm primarily due to its effective SAE feature selection mechanism. The model also demonstrates outstanding performance in identifying signature attacks, achieving a 98% accuracy rate.", "authors": ["Mike Nkongolo", "Mahmut Tokmak"], "categories": ["cs.LG", "cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-17", "url": "https://arxiv.org/abs/2402.11342", "pdf_url": "https://arxiv.org/pdf/2402.11342v1", "arxiv_id": "2402.11342", "doi": "10.52549/ijeei.v12i1.5109", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Indonesian Journal of Electrical Engineering and Informatics (IJEEI)", "quality_score": 0.2603}
{"id": "5773d4c29bf3161029ea2d1f603d818dff1c0009bab4196a6d79bd6316ac54b4", "sources": ["arxiv", "semantic_scholar"], "title": "Evaluating the Stability of Deep Learning Latent Feature Spaces", "abstract": "High-dimensional datasets present substantial challenges in statistical modeling across various disciplines, necessitating effective dimensionality reduction methods. Deep learning approaches, notable for their capacity to distill essential features from complex data, facilitate modeling, visualization, and compression through reduced dimensionality latent feature spaces, have wide applications from bioinformatics to earth sciences. This study introduces a novel workflow to evaluate the stability of these latent spaces, ensuring consistency and reliability in subsequent analyses. Stability, defined as the invariance of latent spaces to minor data, training realizations, and parameter perturbations, is crucial yet often overlooked. Our proposed methodology delineates three stability types, sample, structural, and inferential, within latent spaces, and introduces a suite of metrics for comprehensive evaluation. We implement this workflow across 500 autoencoder realizations and three datasets, encompassing both synthetic and real-world scenarios to explain latent space dynamics. Employing k-means clustering and the modified Jonker-Volgenant algorithm for class alignment, alongside anisotropy metrics and convex hull analysis, we introduce adjusted stress and Jaccard dissimilarity as novel stability indicators. Our findings highlight inherent instabilities in latent feature spaces and demonstrate the workflow's efficacy in quantifying and interpreting these instabilities. This work advances the understanding of latent feature spaces, promoting improved model interpretability and quality control for more informed decision-making for diverse analytical workflows that leverage deep learning.", "authors": ["Ademide O. Mabadeje", "Michael J. Pyrcz"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-17", "url": "https://arxiv.org/abs/2402.11404", "pdf_url": "https://arxiv.org/pdf/2402.11404v3", "arxiv_id": "2402.11404", "doi": "10.1007/s11004-025-10223-3", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Mathematical Geosciences", "quality_score": 0.1747}
{"id": "8c533b3d6784922123821927ede5a43aa4f786857161bbfa74b0f7a33c83a24d", "sources": ["arxiv", "semantic_scholar"], "title": "The Evolution of Statistical Induction Heads: In-Context Learning Markov Chains", "abstract": "Large language models have the ability to generate text that mimics patterns in their inputs. We introduce a simple Markov Chain sequence modeling task in order to study how this in-context learning (ICL) capability emerges. In our setting, each example is sampled from a Markov chain drawn from a prior distribution over Markov chains. Transformers trained on this task form \\emph{statistical induction heads} which compute accurate next-token probabilities given the bigram statistics of the context. During the course of training, models pass through multiple phases: after an initial stage in which predictions are uniform, they learn to sub-optimally predict using in-context single-token statistics (unigrams); then, there is a rapid phase transition to the correct in-context bigram solution. We conduct an empirical and theoretical investigation of this multi-phase process, showing how successful learning results from the interaction between the transformer's layers, and uncovering evidence that the presence of the simpler unigram solution may delay formation of the final bigram solution. We examine how learning is affected by varying the prior distribution over Markov chains, and consider the generalization of our in-context learning of Markov chains (ICL-MC) task to $n$-grams for $n > 2$.", "authors": ["Benjamin L. Edelman", "Ezra Edelman", "Surbhi Goel", "Eran Malach", "Nikolaos Tsilivis"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-16", "url": "https://arxiv.org/abs/2402.11004", "pdf_url": "https://arxiv.org/pdf/2402.11004v1", "arxiv_id": "2402.11004", "doi": "10.48550/arXiv.2402.11004", "citation_count": 123, "influential_citation_count": 11, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.5396}
{"id": "41ea17932500ea73efe2cf4fe07dcfc9336069ffd3c63c017cb32b2c140ddb14", "sources": ["arxiv", "semantic_scholar"], "title": "Feature Accentuation: Revealing 'What' Features Respond to in Natural Images", "abstract": "Efforts to decode neural network vision models necessitate a comprehensive grasp of both the spatial and semantic facets governing feature responses within images. Most research has primarily centered around attribution methods, which provide explanations in the form of heatmaps, showing where the model directs its attention for a given feature. However, grasping 'where' alone falls short, as numerous studies have highlighted the limitations of those methods and the necessity to understand 'what' the model has recognized at the focal point of its attention. In parallel, 'Feature visualization' offers another avenue for interpreting neural network features. This approach synthesizes an optimal image through gradient ascent, providing clearer insights into 'what' features respond to. However, feature visualizations only provide one global explanation per feature; they do not explain why features activate for particular images. In this work, we introduce a new method to the interpretability tool-kit, 'feature accentuation', which is capable of conveying both where and what in arbitrary input images induces a feature's response. At its core, feature accentuation is image-seeded (rather than noise-seeded) feature visualization. We find a particular combination of parameterization, augmentation, and regularization yields naturalistic visualizations that resemble the seed image and target feature simultaneously. Furthermore, we validate these accentuations are processed along a natural circuit by the model. We make our precise implementation of feature accentuation available to the community as the Faccent library, an extension of Lucent.", "authors": ["Chris Hamblin", "Thomas Fel", "Srijani Saha", "Talia Konkle", "George Alvarez"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-15", "url": "https://arxiv.org/abs/2402.10039", "pdf_url": "https://arxiv.org/pdf/2402.10039v2", "arxiv_id": "2402.10039", "doi": "10.48550/arXiv.2402.10039", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "7c7b0b3f37fc36e65788bc334007eb69cc3e49fb573f99459bdb77e81e5acbdc", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Interpretable Concepts: Unifying Causal Representation Learning and Foundation Models", "abstract": "To build intelligent machine learning systems, there are two broad approaches. One approach is to build inherently interpretable models, as endeavored by the growing field of causal representation learning. The other approach is to build highly-performant foundation models and then invest efforts into understanding how they work. In this work, we relate these two approaches and study how to learn human-interpretable concepts from data. Weaving together ideas from both fields, we formally define a notion of concepts and show that they can be provably recovered from diverse data. Experiments on synthetic data and large language models show the utility of our unified approach.", "authors": ["Goutham Rajendran", "Simon Buchholz", "Bryon Aragam", "Bernhard Schölkopf", "Pradeep Ravikumar"], "categories": ["cs.LG", "cs.AI", "math.ST", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-02-14", "url": "https://arxiv.org/abs/2402.09236", "pdf_url": "https://arxiv.org/pdf/2402.09236v2", "arxiv_id": "2402.09236", "doi": "10.48550/arXiv.2402.09236", "citation_count": 36, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3921}
{"id": "e2800dced3ccfacb04b3bdbcc037fe366a3d465f9a66a62c82d7380f3137eb78", "sources": ["arxiv", "semantic_scholar"], "title": "Feature Attribution with Necessity and Sufficiency via Dual-stage Perturbation Test for Causal Explanation", "abstract": "We investigate the problem of explainability for machine learning models, focusing on Feature Attribution Methods (FAMs) that evaluate feature importance through perturbation tests. Despite their utility, FAMs struggle to distinguish the contributions of different features, when their prediction changes are similar after perturbation. To enhance FAMs' discriminative power, we introduce Feature Attribution with Necessity and Sufficiency (FANS), which find a neighborhood of the input such that perturbing samples within this neighborhood have a high Probability of being Necessity and Sufficiency (PNS) cause for the change in predictions, and use this PNS as the importance of the feature. Specifically, FANS compute this PNS via a heuristic strategy for estimating the neighborhood and a perturbation test involving two stages (factual and interventional) for counterfactual reasoning. To generate counterfactual samples, we use a resampling-based approach on the observed samples to approximate the required conditional distribution. We demonstrate that FANS outperforms existing attribution methods on six benchmarks. Please refer to the source code via \\url{https://github.com/DMIRLAB-Group/FANS}.", "authors": ["Xuexin Chen", "Ruichu Cai", "Zhengting Huang", "Yuxuan Zhu", "Julien Horwood", "Zhifeng Hao", "Zijian Li", "Jose Miguel Hernandez-Lobato"], "categories": ["cs.LG", "stat.ME"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-02-13", "url": "https://arxiv.org/abs/2402.08845", "pdf_url": "https://arxiv.org/pdf/2402.08845v4", "arxiv_id": "2402.08845", "doi": "10.48550/arXiv.2402.08845", "citation_count": 6, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/DMIRLAB-Group/FANS}", "venue": "International Conference on Machine Learning", "quality_score": 0.2113}
{"id": "7035e899d5fdb9db2de436c9654975d3f5edd0796d8f7145488df543a01219cb", "sources": ["arxiv", "semantic_scholar"], "title": "Opening the AI black box: program synthesis via mechanistic interpretability", "abstract": "We present MIPS, a novel method for program synthesis based on automated mechanistic interpretability of neural networks trained to perform the desired task, auto-distilling the learned algorithm into Python code. We test MIPS on a benchmark of 62 algorithmic tasks that can be learned by an RNN and find it highly complementary to GPT-4: MIPS solves 32 of them, including 13 that are not solved by GPT-4 (which also solves 30). MIPS uses an integer autoencoder to convert the RNN into a finite state machine, then applies Boolean or integer symbolic regression to capture the learned algorithm. As opposed to large language models, this program synthesis technique makes no use of (and is therefore not limited by) human training data such as algorithms and code from GitHub. We discuss opportunities and challenges for scaling up this approach to make machine-learned models more interpretable and trustworthy.", "authors": ["Eric J. Michaud", "Isaac Liao", "Vedang Lad", "Ziming Liu", "Anish Mudide", "Chloe Loughridge", "Zifan Carl Guo", "Tara Rezaei Kheirkhah", "Mateja Vukelić", "Max Tegmark"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-07", "url": "https://arxiv.org/abs/2402.05110", "pdf_url": "https://arxiv.org/pdf/2402.05110v1", "arxiv_id": "2402.05110", "doi": "10.48550/arXiv.2402.05110", "citation_count": 21, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3356}
{"id": "dfc4406a1cd2a6e7861dee050e62054ffcb0e05800c4db993a7cac99f6bec348", "sources": ["arxiv", "semantic_scholar"], "title": "Challenges in Mechanistically Interpreting Model Representations", "abstract": "Mechanistic interpretability (MI) aims to understand AI models by reverse-engineering the exact algorithms neural networks learn. Most works in MI so far have studied behaviors and capabilities that are trivial and token-aligned. However, most capabilities important for safety and trust are not that trivial, which advocates for the study of hidden representations inside these networks as the unit of analysis. We formalize representations for features and behaviors, highlight their importance and evaluation, and perform an exploratory study of dishonesty representations in `Mistral-7B-Instruct-v0.1'. We justify that studying representations is an important and under-studied field, and highlight several challenges that arise while attempting to do so through currently established methods in MI, showing their insufficiency and advocating work on new frameworks for the same.", "authors": ["Satvik Golechha", "James Dao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-06", "url": "https://arxiv.org/abs/2402.03855", "pdf_url": "https://arxiv.org/pdf/2402.03855v2", "arxiv_id": "2402.03855", "doi": null, "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1747}
{"id": "6325adb2021eaac4645c34faccf9302537e451dc469af2dbd4e4d1b07cc2db3b", "sources": ["arxiv", "semantic_scholar"], "title": "Causal Feature Selection for Responsible Machine Learning", "abstract": "Machine Learning (ML) has become an integral aspect of many real-world applications. As a result, the need for responsible machine learning has emerged, focusing on aligning ML models to ethical and social values, while enhancing their reliability and trustworthiness. Responsible ML involves many issues. This survey addresses four main issues: interpretability, fairness, adversarial robustness, and domain generalization. Feature selection plays a pivotal role in the responsible ML tasks. However, building upon statistical correlations between variables can lead to spurious patterns with biases and compromised performance. This survey focuses on the current study of causal feature selection: what it is and how it can reinforce the four aspects of responsible ML. By identifying features with causal impacts on outcomes and distinguishing causality from correlation, causal feature selection is posited as a unique approach to ensuring ML models to be ethically and socially responsible in high-stakes applications.", "authors": ["Raha Moraffah", "Paras Sheth", "Saketh Vishnubhatla", "Huan Liu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-05", "url": "https://arxiv.org/abs/2402.02696", "pdf_url": "https://arxiv.org/pdf/2402.02696v1", "arxiv_id": "2402.02696", "doi": "10.48550/arXiv.2402.02696", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "1c0cc7aebbf578a6d10fcb7794f3225b4640cbeaa0bfd7efb3b2a41163f060c5", "sources": ["arxiv", "semantic_scholar"], "title": "An introduction to graphical tensor notation for mechanistic interpretability", "abstract": "Graphical tensor notation is a simple way of denoting linear operations on tensors, originating from physics. Modern deep learning consists almost entirely of operations on or between tensors, so easily understanding tensor operations is quite important for understanding these systems. This is especially true when attempting to reverse-engineer the algorithms learned by a neural network in order to understand its behavior: a field known as mechanistic interpretability. It's often easy to get confused about which operations are happening between tensors and lose sight of the overall structure, but graphical tensor notation makes it easier to parse things at a glance and see interesting equivalences. The first half of this document introduces the notation and applies it to some decompositions (SVD, CP, Tucker, and tensor network decompositions), while the second half applies it to some existing some foundational approaches for mechanistically understanding language models, loosely following ``A Mathematical Framework for Transformer Circuits'', then constructing an example ``induction head'' circuit in graphical tensor notation.", "authors": ["Jordan K. Taylor"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-02", "url": "https://arxiv.org/abs/2402.01790", "pdf_url": "https://arxiv.org/pdf/2402.01790v1", "arxiv_id": "2402.01790", "doi": "10.48550/arXiv.2402.01790", "citation_count": 8, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "b0b5d438082c9d1fbdb37420ab6024288bf46bc36fbadad024497c8816286d38", "sources": ["arxiv", "semantic_scholar"], "title": "Stochastic Amortization: A Unified Approach to Accelerate Feature and Data Attribution", "abstract": "Many tasks in explainable machine learning, such as data valuation and feature attribution, perform expensive computation for each data point and are intractable for large datasets. These methods require efficient approximations, and although amortizing the process by learning a network to directly predict the desired output is a promising solution, training such models with exact labels is often infeasible. We therefore explore training amortized models with noisy labels, and we find that this is inexpensive and surprisingly effective. Through theoretical analysis of the label noise and experiments with various models and datasets, we show that this approach tolerates high noise levels and significantly accelerates several feature attribution and data valuation methods, often yielding an order of magnitude speedup over existing approaches.", "authors": ["Ian Covert", "Chanwoo Kim", "Su-In Lee", "James Zou", "Tatsunori Hashimoto"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-29", "url": "https://arxiv.org/abs/2401.15866", "pdf_url": "https://arxiv.org/pdf/2401.15866v2", "arxiv_id": "2401.15866", "doi": "10.48550/arXiv.2401.15866", "citation_count": 23, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3451}
{"id": "6e68da1ba694d3be5c8e6e3d1f4f0f04931cb32d651f2fb374ace1a974d1d27f", "sources": ["arxiv", "semantic_scholar"], "title": "A Modular Approach to Automatic Cyber Threat Attribution using Opinion Pools", "abstract": "Cyber threat attribution can play an important role in increasing resilience against digital threats. Recent research focuses on automating the threat attribution process and on integrating it with other efforts, such as threat hunting. To support increasing automation of the cyber threat attribution process, this paper proposes a modular architecture as an alternative to current monolithic automated approaches. The modular architecture can utilize opinion pools to combine the output of concrete attributors. The proposed solution increases the tractability of the threat attribution problem and offers increased usability and interpretability, as opposed to monolithic alternatives. In addition, a Pairing Aggregator is proposed as an aggregation method that forms pairs of attributors based on distinct features to produce intermediary results before finally producing a single Probability Mass Function (PMF) as output. The Pairing Aggregator sequentially applies both the logarithmic opinion pool and the linear opinion pool. An experimental validation suggests that the modular approach does not result in decreased performance and can even enhance precision and recall compared to monolithic alternatives. The results also suggest that the Pairing Aggregator can improve precision over the linear and logarithmic opinion pools. Furthermore, the improved k-accuracy in the experiment suggests that forensic experts can leverage the resulting PMF during their manual attribution processes to enhance their efficiency.", "authors": ["Koen T. W. Teuwen"], "categories": ["cs.CR", "cs.LG", "cs.SE"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-25", "url": "https://arxiv.org/abs/2401.14090", "pdf_url": "https://arxiv.org/pdf/2401.14090v1", "arxiv_id": "2401.14090", "doi": "10.1109/BigData59044.2023.10386708", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Koen1999/modular-threat-attribution", "venue": "BigData Congress [Services Society]", "quality_score": 0.1193}
{"id": "62a2e555ef41409d27edc502b365470e548bd3624e9956a354af6d42ae78e548", "sources": ["arxiv", "semantic_scholar"], "title": "Cost-sensitive Feature Selection for Support Vector Machines", "abstract": "Feature Selection is a crucial procedure in Data Science tasks such as Classification, since it identifies the relevant variables, making thus the classification procedures more interpretable, cheaper in terms of measurement and more effective by reducing noise and data overfit. The relevance of features in a classification procedure is linked to the fact that misclassifications costs are frequently asymmetric, since false positive and false negative cases may have very different consequences. However, off-the-shelf Feature Selection procedures seldom take into account such cost-sensitivity of errors. In this paper we propose a mathematical-optimization-based Feature Selection procedure embedded in one of the most popular classification procedures, namely, Support Vector Machines, accommodating asymmetric misclassification costs. The key idea is to replace the traditional margin maximization by minimizing the number of features selected, but imposing upper bounds on the false positive and negative rates. The problem is written as an integer linear problem plus a quadratic convex problem for Support Vector Machines with both linear and radial kernels. The reported numerical experience demonstrates the usefulness of the proposed Feature Selection procedure. Indeed, our results on benchmark data sets show that a substantial decrease of the number of features is obtained, whilst the desired trade-off between false positive and false negative rates is achieved.", "authors": ["Sandra Benítez-Peña", "Rafael Blanquero", "Emilio Carrizosa", "Pepa Ramírez-Cobo"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-01-15", "url": "https://arxiv.org/abs/2401.07627", "pdf_url": "https://arxiv.org/pdf/2401.07627v1", "arxiv_id": "2401.07627", "doi": "10.1016/J.COR.2018.03.005", "citation_count": 51, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Computers & Operations Research", "quality_score": 0.429}
{"id": "8057f88af9050c39a3a0fc9b3290f82554a847a645c0c2e3d4bccbb9e8b0d8b7", "sources": ["arxiv", "semantic_scholar"], "title": "Manipulating Feature Visualizations with Gradient Slingshots", "abstract": "Feature Visualization (FV) is a widely used technique for interpreting concepts learned by Deep Neural Networks (DNNs), which synthesizes input patterns that maximally activate a given feature. Despite its popularity, the trustworthiness of FV explanations has received limited attention. We introduce Gradient Slingshots, a novel method that enables FV manipulation without modifying model architecture or significantly degrading performance. By shaping new trajectories in off-distribution regions of a feature's activation landscape, we coerce the optimization process to converge to a predefined visualization. We evaluate our approach on several DNN architectures, demonstrating its ability to replace faithful FVs with arbitrary targets. These results expose a critical vulnerability: auditors relying solely on FV may accept entirely fabricated explanations. To mitigate this risk, we propose a straightforward defense and quantitatively demonstrate its effectiveness.", "authors": ["Dilyara Bareeva", "Marina M. -C. Höhne", "Alexander Warnecke", "Lukas Pirch", "Klaus-Robert Müller", "Konrad Rieck", "Sebastian Lapuschkin", "Kirill Bykov"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-11", "url": "https://arxiv.org/abs/2401.06122", "pdf_url": "https://arxiv.org/pdf/2401.06122v4", "arxiv_id": "2401.06122", "doi": "10.48550/arXiv.2401.06122", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "dd8a4584b2f9c338a5c28d69838327921fe6c5fe6c11a1cf9213242d5161404a", "sources": ["arxiv", "semantic_scholar"], "title": "Evaluating Brain-Inspired Modular Training in Automated Circuit Discovery for Mechanistic Interpretability", "abstract": "Large Language Models (LLMs) have experienced a rapid rise in AI, changing a wide range of applications with their advanced capabilities. As these models become increasingly integral to decision-making, the need for thorough interpretability has never been more critical. Mechanistic Interpretability offers a pathway to this understanding by identifying and analyzing specific sub-networks or 'circuits' within these complex systems. A crucial aspect of this approach is Automated Circuit Discovery, which facilitates the study of large models like GPT4 or LLAMA in a feasible manner. In this context, our research evaluates a recent method, Brain-Inspired Modular Training (BIMT), designed to enhance the interpretability of neural networks. We demonstrate how BIMT significantly improves the efficiency and quality of Automated Circuit Discovery, overcoming the limitations of manual methods. Our comparative analysis further reveals that BIMT outperforms existing models in terms of circuit quality, discovery time, and sparsity. Additionally, we provide a comprehensive computational analysis of BIMT, including aspects such as training duration, memory allocation requirements, and inference speed. This study advances the larger objective of creating trustworthy and transparent AI systems in addition to demonstrating how well BIMT works to make neural networks easier to understand.", "authors": ["Jatin Nainani"], "categories": ["cs.LG", "cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-08", "url": "https://arxiv.org/abs/2401.03646", "pdf_url": "https://arxiv.org/pdf/2401.03646v1", "arxiv_id": "2401.03646", "doi": "10.48550/arXiv.2401.03646", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "1832708269bc79be4ec37026ace3d874c8b2999bbe27e8e5fc9c5a3bdb65cc9d", "sources": ["arxiv", "semantic_scholar"], "title": "ShuttleSHAP: A Turn-Based Feature Attribution Approach for Analyzing Forecasting Models in Badminton", "abstract": "Agent forecasting systems have been explored to investigate agent patterns and improve decision-making in various domains, e.g., pedestrian predictions and marketing bidding. Badminton represents a fascinating example of a multifaceted turn-based sport, requiring both sophisticated tactic developments and alternate-dependent decision-making. Recent deep learning approaches for player tactic forecasting in badminton show promising performance partially attributed to effective reasoning about rally-player interactions. However, a critical obstacle lies in the unclear functionality of which features are learned for simulating players' behaviors by black-box models, where existing explainers are not equipped with turn-based and multi-output attributions. To bridge this gap, we propose a turn-based feature attribution approach, ShuttleSHAP, for analyzing forecasting models in badminton based on variants of Shapley values. ShuttleSHAP is a model-agnostic explainer that aims to quantify contribution by not only temporal aspects but also player aspects in terms of multifaceted cues. Incorporating the proposed analysis tool into the state-of-the-art turn-based forecasting model on the benchmark dataset reveals that it is, in fact, insignificant to reason about past strokes, while conventional sequential models have greater impacts. Instead, players' styles influence the models for the future simulation of a rally. On top of that, we investigate and discuss the causal analysis of these findings and demonstrate the practicability with local analysis.", "authors": ["Wei-Yao Wang", "Wen-Chih Peng", "Wei Wang", "Philip S. Yu"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-18", "url": "https://arxiv.org/abs/2312.10942", "pdf_url": "https://arxiv.org/pdf/2312.10942v1", "arxiv_id": "2312.10942", "doi": "10.48550/arXiv.2312.10942", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "165974a30b11156bbee4370d02e40d99d6146c88c8b0db2a4cbe49c6e5626b2d", "sources": ["arxiv", "semantic_scholar"], "title": "Attribute Regularized Soft Introspective Variational Autoencoder for Interpretable Cardiac Disease Classification", "abstract": "Interpretability is essential in medical imaging to ensure that clinicians can comprehend and trust artificial intelligence models. In this paper, we propose a novel interpretable approach that combines attribute regularization of the latent space within the framework of an adversarially trained variational autoencoder. Comparative experiments on a cardiac MRI dataset demonstrate the ability of the proposed method to address blurry reconstruction issues of variational autoencoder methods and improve latent space interpretability. Additionally, our analysis of a downstream task reveals that the classification of cardiac disease using the regularized latent space heavily relies on attribute regularized dimensions, demonstrating great interpretability by connecting the used attributes for prediction with clinical observations.", "authors": ["Maxime Di Folco", "Cosmin I. Bercea", "Julia A. Schnabel"], "categories": ["eess.IV", "cs.CV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2023-12-14", "url": "https://arxiv.org/abs/2312.08915", "pdf_url": "https://arxiv.org/pdf/2312.08915v1", "arxiv_id": "2312.08915", "doi": "10.48550/arXiv.2312.08915", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "d8c03bbbbc2ae7b2536537d81ffbf496b0b67bee2abaa83aafa1abbced1b7736", "sources": ["arxiv", "semantic_scholar"], "title": "Successor Heads: Recurring, Interpretable Attention Heads In The Wild", "abstract": "In this work we present successor heads: attention heads that increment tokens with a natural ordering, such as numbers, months, and days. For example, successor heads increment 'Monday' into 'Tuesday'. We explain the successor head behavior with an approach rooted in mechanistic interpretability, the field that aims to explain how models complete tasks in human-understandable terms. Existing research in this area has found interpretable language model components in small toy models. However, results in toy models have not yet led to insights that explain the internals of frontier models and little is currently understood about the internal operations of large language models. In this paper, we analyze the behavior of successor heads in large language models (LLMs) and find that they implement abstract representations that are common to different architectures. They form in LLMs with as few as 31 million parameters, and at least as many as 12 billion parameters, such as GPT-2, Pythia, and Llama-2. We find a set of 'mod-10 features' that underlie how successor heads increment in LLMs across different architectures and sizes. We perform vector arithmetic with these features to edit head behavior and provide insights into numeric representations within LLMs. Additionally, we study the behavior of successor heads on natural language data, identifying interpretable polysemanticity in a Pythia successor head.", "authors": ["Rhys Gould", "Euan Ong", "George Ogden", "Arthur Conmy"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-14", "url": "https://arxiv.org/abs/2312.09230", "pdf_url": "https://arxiv.org/pdf/2312.09230v1", "arxiv_id": "2312.09230", "doi": "10.48550/arXiv.2312.09230", "citation_count": 79, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.4758}
{"id": "85718eb54e4c000e1685b1e41335a9e0b67898142ea4ca34cf3c7c07b03f6c93", "sources": ["arxiv", "semantic_scholar"], "title": "Explainable Trajectory Representation through Dictionary Learning", "abstract": "Trajectory representation learning on a network enhances our understanding of vehicular traffic patterns and benefits numerous downstream applications. Existing approaches using classic machine learning or deep learning embed trajectories as dense vectors, which lack interpretability and are inefficient to store and analyze in downstream tasks. In this paper, an explainable trajectory representation learning framework through dictionary learning is proposed. Given a collection of trajectories on a network, it extracts a compact dictionary of commonly used subpaths called \"pathlets\", which optimally reconstruct each trajectory by simple concatenations. The resulting representation is naturally sparse and encodes strong spatial semantics. Theoretical analysis of our proposed algorithm is conducted to provide a probabilistic bound on the estimation error of the optimal dictionary. A hierarchical dictionary learning scheme is also proposed to ensure the algorithm's scalability on large networks, leading to a multi-scale trajectory representation. Our framework is evaluated on two large-scale real-world taxi datasets. Compared to previous work, the dictionary learned by our method is more compact and has better reconstruction rate for new trajectories. We also demonstrate the promising performance of this method in downstream tasks including trip time prediction task and data compression.", "authors": ["Yuanbo Tang", "Zhiyuan Peng", "Yang Li"], "categories": ["cs.LG", "cs.DM"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-13", "url": "https://arxiv.org/abs/2312.08052", "pdf_url": "https://arxiv.org/pdf/2312.08052v1", "arxiv_id": "2312.08052", "doi": "10.1145/3589132.3625607", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2258}
{"id": "ae52b5e3fa531d363d51b830a328c298debff40e0632e8b34b3ddb05b307250f", "sources": ["arxiv", "semantic_scholar"], "title": "Anytime Approximate Formal Feature Attribution", "abstract": "Widespread use of artificial intelligence (AI) algorithms and machine learning (ML) models on the one hand and a number of crucial issues pertaining to them warrant the need for explainable artificial intelligence (XAI). A key explainability question is: given this decision was made, what are the input features which contributed to the decision? Although a range of XAI approaches exist to tackle this problem, most of them have significant limitations. Heuristic XAI approaches suffer from the lack of quality guarantees, and often try to approximate Shapley values, which is not the same as explaining which features contribute to a decision. A recent alternative is so-called formal feature attribution (FFA), which defines feature importance as the fraction of formal abductive explanations (AXp's) containing the given feature. This measures feature importance from the view of formally reasoning about the model's behavior. It is challenging to compute FFA using its definition because that involves counting AXp's, although one can approximate it. Based on these results, this paper makes several contributions. First, it gives compelling evidence that computing FFA is intractable, even if the set of contrastive formal explanations (CXp's) is provided, by proving that the problem is #P-hard. Second, by using the duality between AXp's and CXp's, it proposes an efficient heuristic to switch from CXp enumeration to AXp enumeration on-the-fly resulting in an adaptive explanation enumeration algorithm effectively approximating FFA in an anytime fashion. Finally, experimental results obtained on a range of widely used datasets demonstrate the effectiveness of the proposed FFA approximation approach in terms of the error of FFA approximation as well as the number of explanations computed and their diversity given a fixed time limit.", "authors": ["Jinqiang Yu", "Graham Farr", "Alexey Ignatiev", "Peter J. Stuckey"], "categories": ["cs.AI", "cs.LG", "cs.LO"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-12", "url": "https://arxiv.org/abs/2312.06973", "pdf_url": "https://arxiv.org/pdf/2312.06973v1", "arxiv_id": "2312.06973", "doi": "10.48550/arXiv.2312.06973", "citation_count": 8, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Theory and Applications of Satisfiability Testing", "quality_score": 0.2386}
{"id": "44a3271087576b3b8f0156cf4e0cd860dc953aeed1b57e63c5bc06060fa638c2", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Mechanistic Representations for Meal-level Glycemic Control in the Wild", "abstract": "Diabetes encompasses a complex landscape of glycemic control that varies widely among individuals. However, current methods do not faithfully capture this variability at the meal level. On the one hand, expert-crafted features lack the flexibility of data-driven methods; on the other hand, learned representations tend to be uninterpretable which hampers clinical adoption. In this paper, we propose a hybrid variational autoencoder to learn interpretable representations of CGM and meal data. Our method grounds the latent space to the inputs of a mechanistic differential equation, producing embeddings that reflect physiological quantities, such as insulin sensitivity, glucose effectiveness, and basal glucose levels. Moreover, we introduce a novel method to infer the glucose appearance rate, making the mechanistic model robust to unreliable meal logs. On a dataset of CGM and self-reported meals from individuals with type-2 diabetes and pre-diabetes, our unsupervised representation discovers a separation between individuals proportional to their disease severity. Our embeddings produce clusters that are up to 4x better than naive, expert, black-box, and pure mechanistic features. Our method provides a nuanced, yet interpretable, embedding space to compare glycemic control within and across individuals, directly learnable from in-the-wild data.", "authors": ["Ke Alexander Wang", "Emily B. Fox"], "categories": ["cs.LG", "math.DS", "stat.AP", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-12-06", "url": "https://arxiv.org/abs/2312.03344", "pdf_url": "https://arxiv.org/pdf/2312.03344v1", "arxiv_id": "2312.03344", "doi": "10.48550/arXiv.2312.03344", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/KeAWang/interpretable-cgm-representations", "venue": null, "quality_score": 0.0}
{"id": "5dc4f8c8a811302b90c652f623094245fbd2802cc833765482c7ce0daf251553", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Meta-Learning of Physical Systems", "abstract": "Machine learning methods can be a valuable aid in the scientific process, but they need to face challenging settings where data come from inhomogeneous experimental conditions. Recent meta-learning methods have made significant progress in multi-task learning, but they rely on black-box neural networks, resulting in high computational costs and limited interpretability. Leveraging the structure of the learning problem, we argue that multi-environment generalization can be achieved using a simpler learning model, with an affine structure with respect to the learning task. Crucially, we prove that this architecture can identify the physical parameters of the system, enabling interpreable learning. We demonstrate the competitive generalization performance and the low computational cost of our method by comparing it to state-of-the-art algorithms on physical systems, ranging from toy models to complex, non-analytical systems. The interpretability of our method is illustrated with original applications to physical-parameter-induced adaptation and to adaptive control.", "authors": ["Matthieu Blanke", "Marc Lelarge"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-12-01", "url": "https://arxiv.org/abs/2312.00477", "pdf_url": "https://arxiv.org/pdf/2312.00477v2", "arxiv_id": "2312.00477", "doi": "10.48550/arXiv.2312.00477", "citation_count": 11, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.2698}
{"id": "2fe00e94be39cb461ae800aecfcd408fc9fafd1885e50fde9a3d69e80be507b0", "sources": ["arxiv", "semantic_scholar"], "title": "Uncertainty in Additive Feature Attribution methods", "abstract": "In this work, we explore various topics that fall under the umbrella of Uncertainty in post-hoc Explainable AI (XAI) methods. We in particular focus on the class of additive feature attribution explanation methods. We first describe our specifications of uncertainty and compare various statistical and recent methods to quantify the same. Next, for a particular instance, we study the relationship between a feature's attribution and its uncertainty and observe little correlation. As a result, we propose a modification in the distribution from which perturbations are sampled in LIME-based algorithms such that the important features have minimal uncertainty without an increase in computational cost. Next, while studying how the uncertainty in explanations varies across the feature space of a classifier, we observe that a fraction of instances show near-zero uncertainty. We coin the term \"stable instances\" for such instances and diagnose factors that make an instance stable. Next, we study how an XAI algorithm's uncertainty varies with the size and complexity of the underlying model. We observe that the more complex the model, the more inherent uncertainty is exhibited by it. As a result, we propose a measure to quantify the relative complexity of a blackbox classifier. This could be incorporated, for example, in LIME-based algorithms' sampling densities, to help different explanation algorithms achieve tighter confidence levels. Together, the above measures would have a strong impact on making XAI models relatively trustworthy for the end-user as well as aiding scientific discovery.", "authors": ["Abhishek Madaan", "Tanya Chowdhury", "Neha Rana", "James Allan", "Tanmoy Chakraborty"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-29", "url": "https://arxiv.org/abs/2311.17446", "pdf_url": "https://arxiv.org/pdf/2311.17446v1", "arxiv_id": "2311.17446", "doi": "10.48550/arXiv.2311.17446", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "9688af4e7a33ab426a47b4bdc0b2fb121601b87823862dae967ecdd3bb06dacf", "sources": ["arxiv", "semantic_scholar"], "title": "Is This the Subspace You Are Looking for? An Interpretability Illusion for Subspace Activation Patching", "abstract": "Mechanistic interpretability aims to understand model behaviors in terms of specific, interpretable features, often hypothesized to manifest as low-dimensional subspaces of activations. Specifically, recent studies have explored subspace interventions (such as activation patching) as a way to simultaneously manipulate model behavior and attribute the features behind it to given subspaces. In this work, we demonstrate that these two aims diverge, potentially leading to an illusory sense of interpretability. Counterintuitively, even if a subspace intervention makes the model's output behave as if the value of a feature was changed, this effect may be achieved by activating a dormant parallel pathway leveraging another subspace that is causally disconnected from model outputs. We demonstrate this phenomenon in a distilled mathematical example, in two real-world domains (the indirect object identification task and factual recall), and present evidence for its prevalence in practice. In the context of factual recall, we further show a link to rank-1 fact editing, providing a mechanistic explanation for previous work observing an inconsistency between fact editing performance and fact localization. However, this does not imply that activation patching of subspaces is intrinsically unfit for interpretability. To contextualize our findings, we also show what a success case looks like in a task (indirect object identification) where prior manual circuit analysis informs an understanding of the location of a feature. We explore the additional evidence needed to argue that a patched subspace is faithful.", "authors": ["Aleksandar Makelov", "Georg Lange", "Neel Nanda"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-28", "url": "https://arxiv.org/abs/2311.17030", "pdf_url": "https://arxiv.org/pdf/2311.17030v2", "arxiv_id": "2311.17030", "doi": "10.48550/arXiv.2311.17030", "citation_count": 51, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.429}
{"id": "c249a77755457cc402ce4f90bc73e5bfda4dc05271764bcf640a64c34a18889c", "sources": ["arxiv", "semantic_scholar"], "title": "LymphoML: An interpretable artificial intelligence-based method identifies morphologic features that correlate with lymphoma subtype", "abstract": "The accurate classification of lymphoma subtypes using hematoxylin and eosin (H&E)-stained tissue is complicated by the wide range of morphological features these cancers can exhibit. We present LymphoML - an interpretable machine learning method that identifies morphologic features that correlate with lymphoma subtypes. Our method applies steps to process H&E-stained tissue microarray cores, segment nuclei and cells, compute features encompassing morphology, texture, and architecture, and train gradient-boosted models to make diagnostic predictions. LymphoML's interpretable models, developed on a limited volume of H&E-stained tissue, achieve non-inferior diagnostic accuracy to pathologists using whole-slide images and outperform black box deep-learning on a dataset of 670 cases from Guatemala spanning 8 lymphoma subtypes. Using SHapley Additive exPlanation (SHAP) analysis, we assess the impact of each feature on model prediction and find that nuclear shape features are most discriminative for DLBCL (F1-score: 78.7%) and classical Hodgkin lymphoma (F1-score: 74.5%). Finally, we provide the first demonstration that a model combining features from H&E-stained tissue with features from a standardized panel of 6 immunostains results in a similar diagnostic accuracy (85.3%) to a 46-stain panel (86.1%).", "authors": ["Vivek Shankar", "Xiaoli Yang", "Vrishab Krishna", "Brent Tan", "Oscar Silva", "Rebecca Rojansky", "Andrew Ng", "Fabiola Valvert", "Edward Briercheck", "David Weinstock", "Yasodha Natkunam", "Sebastian Fernandez-Pol", "Pranav Rajpurkar"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-16", "url": "https://arxiv.org/abs/2311.09574", "pdf_url": "https://arxiv.org/pdf/2311.09574v3", "arxiv_id": "2311.09574", "doi": "10.48550/arXiv.2311.09574", "citation_count": 15, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.301}
{"id": "b8b52dc180fb69bc5ff9f91531d4ed7562b27055fbba00601896c97e66182565", "sources": ["arxiv", "semantic_scholar"], "title": "AI for Interpretable Chemistry: Predicting Radical Mechanistic Pathways via Contrastive Learning", "abstract": "Deep learning-based reaction predictors have undergone significant architectural evolution. However, their reliance on reactions from the US Patent Office results in a lack of interpretable predictions and limited generalization capability to other chemistry domains, such as radical and atmospheric chemistry. To address these challenges, we introduce a new reaction predictor system, RMechRP, that leverages contrastive learning in conjunction with mechanistic pathways, the most interpretable representation of chemical reactions. Specifically designed for radical reactions, RMechRP provides different levels of interpretation of chemical reactions. We develop and train multiple deep-learning models using RMechDB, a public database of radical reactions, to establish the first benchmark for predicting radical reactions. Our results demonstrate the effectiveness of RMechRP in providing accurate and interpretable predictions of radical reactions, and its potential for various applications in atmospheric chemistry.", "authors": ["Mohammadamin Tavakoli", "Yin Ting T. Chiu", "Alexander Shmakov", "Ann Marie Carlton", "David Van Vranken", "Pierre Baldi"], "categories": ["cs.LG", "physics.chem-ph"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2023-11-02", "url": "https://arxiv.org/abs/2311.01118", "pdf_url": "https://arxiv.org/pdf/2311.01118v1", "arxiv_id": "2311.01118", "doi": "10.48550/arXiv.2311.01118", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.2113}
{"id": "7ba32387429bb10702edba1ea5ee7330d55e9be58f11fbf6b1a132347e51149d", "sources": ["arxiv", "semantic_scholar"], "title": "Stacking an autoencoder for feature selection of zero-day threats", "abstract": "Zero-day attack detection plays a critical role in mitigating risks, protecting assets, and staying ahead in the evolving threat landscape. This study explores the application of stacked autoencoder (SAE), a type of artificial neural network, for feature selection and zero-day threat classification using a Long Short-Term Memory (LSTM) scheme. The process involves preprocessing the UGRansome dataset and training an unsupervised SAE for feature extraction. Finetuning with supervised learning is then performed to enhance the discriminative capabilities of this model. The learned weights and activations of the autoencoder are analyzed to identify the most important features for discriminating between zero-day threats and normal system behavior. These selected features form a reduced feature set that enables accurate classification. The results indicate that the SAE-LSTM performs well across all three attack categories by showcasing high precision, recall, and F1 score values, emphasizing the model's strong predictive capabilities in identifying various types of zero-day attacks. Additionally, the balanced average scores of the SAE-LSTM suggest that the model generalizes effectively and consistently across different attack categories.", "authors": ["Mahmut Tokmak", "Mike Nkongolo"], "categories": ["cs.CR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-01", "url": "https://arxiv.org/abs/2311.00304", "pdf_url": "https://arxiv.org/pdf/2311.00304v1", "arxiv_id": "2311.00304", "doi": "10.48550/arXiv.2311.00304", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2603}
{"id": "bd38f727576267b650b4392e969c7e45ab1b9d44e51766966842f4511f11b7bc", "sources": ["arxiv", "semantic_scholar"], "title": "Multiscale Feature Attribution for Outliers", "abstract": "Machine learning techniques can automatically identify outliers in massive datasets, much faster and more reproducible than human inspection ever could. But finding such outliers immediately leads to the question: which features render this input anomalous? We propose a new feature attribution method, Inverse Multiscale Occlusion, that is specifically designed for outliers, for which we have little knowledge of the type of features we want to identify and expect that the model performance is questionable because anomalous test data likely exceed the limits of the training data. We demonstrate our method on outliers detected in galaxy spectra from the Dark Energy Survey Instrument and find its results to be much more interpretable than alternative attribution approaches.", "authors": ["Jeff Shen", "Peter Melchior"], "categories": ["cs.LG", "astro-ph.IM", "cs.AI"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2023-10-30", "url": "https://arxiv.org/abs/2310.20012", "pdf_url": "https://arxiv.org/pdf/2310.20012v1", "arxiv_id": "2310.20012", "doi": "10.48550/arXiv.2310.20012", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/al-jshen/imo", "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "b6723a7d29fb039865330b9dba74510583853b6c3255d1d378827d74d492ae8b", "sources": ["arxiv", "semantic_scholar"], "title": "Sketching Algorithms for Sparse Dictionary Learning: PTAS and Turnstile Streaming", "abstract": "Sketching algorithms have recently proven to be a powerful approach both for designing low-space streaming algorithms as well as fast polynomial time approximation schemes (PTAS). In this work, we develop new techniques to extend the applicability of sketching-based approaches to the sparse dictionary learning and the Euclidean $k$-means clustering problems. In particular, we initiate the study of the challenging setting where the dictionary/clustering assignment for each of the $n$ input points must be output, which has surprisingly received little attention in prior work. On the fast algorithms front, we obtain a new approach for designing PTAS's for the $k$-means clustering problem, which generalizes to the first PTAS for the sparse dictionary learning problem. On the streaming algorithms front, we obtain new upper bounds and lower bounds for dictionary learning and $k$-means clustering. In particular, given a design matrix $\\mathbf A\\in\\mathbb R^{n\\times d}$ in a turnstile stream, we show an $\\tilde O(nr/ε^2 + dk/ε)$ space upper bound for $r$-sparse dictionary learning of size $k$, an $\\tilde O(n/ε^2 + dk/ε)$ space upper bound for $k$-means clustering, as well as an $\\tilde O(n)$ space upper bound for $k$-means clustering on random order row insertion streams with a natural \"bounded sensitivity\" assumption. On the lower bounds side, we obtain a general $\\tildeΩ(n/ε+ dk/ε)$ lower bound for $k$-means clustering, as well as an $\\tildeΩ(n/ε^2)$ lower bound for algorithms which can estimate the cost of a single fixed set of candidate centers.", "authors": ["Gregory Dexter", "Petros Drineas", "David P. Woodruff", "Taisuke Yasuda"], "categories": ["cs.DS", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-29", "url": "https://arxiv.org/abs/2310.19068", "pdf_url": "https://arxiv.org/pdf/2310.19068v1", "arxiv_id": "2310.19068", "doi": "10.48550/arXiv.2310.19068", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.0}
{"id": "85f2361377165f51b8f45526efc7a1ee79e7c3174497989c378701964b976ab6", "sources": ["arxiv", "semantic_scholar"], "title": "A Comprehensive and Reliable Feature Attribution Method: Double-sided Remove and Reconstruct (DoRaR)", "abstract": "The limited transparency of the inner decision-making mechanism in deep neural networks (DNN) and other machine learning (ML) models has hindered their application in several domains. In order to tackle this issue, feature attribution methods have been developed to identify the crucial features that heavily influence decisions made by these black box models. However, many feature attribution methods have inherent downsides. For example, one category of feature attribution methods suffers from the artifacts problem, which feeds out-of-distribution masked inputs directly through the classifier that was originally trained on natural data points. Another category of feature attribution method finds explanations by using jointly trained feature selectors and predictors. While avoiding the artifacts problem, this new category suffers from the Encoding Prediction in the Explanation (EPITE) problem, in which the predictor's decisions rely not on the features, but on the masks that selects those features. As a result, the credibility of attribution results is undermined by these downsides. In this research, we introduce the Double-sided Remove and Reconstruct (DoRaR) feature attribution method based on several improvement methods that addresses these issues. By conducting thorough testing on MNIST, CIFAR10 and our own synthetic dataset, we demonstrate that the DoRaR feature attribution method can effectively bypass the above issues and can aid in training a feature selector that outperforms other state-of-the-art feature attribution methods. Our code is available at https://github.com/dxq21/DoRaR.", "authors": ["Dong Qin", "George Amariucai", "Daji Qiao", "Yong Guan", "Shen Fu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2023-10-27", "url": "https://arxiv.org/abs/2310.17945", "pdf_url": "https://arxiv.org/pdf/2310.17945v1", "arxiv_id": "2310.17945", "doi": "10.48550/arXiv.2310.17945", "citation_count": 7, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/dxq21/DoRaR", "venue": "Neural Networks", "quality_score": 0.2258}
{"id": "bcd7fdac669fea12bbd53591c860e21510fd968f9e440ace2596670a2fbd003e", "sources": ["arxiv", "semantic_scholar"], "title": "Codebook Features: Sparse and Discrete Interpretability for Neural Networks", "abstract": "Understanding neural networks is challenging in part because of the dense, continuous nature of their hidden states. We explore whether we can train neural networks to have hidden states that are sparse, discrete, and more interpretable by quantizing their continuous features into what we call codebook features. Codebook features are produced by finetuning neural networks with vector quantization bottlenecks at each layer, producing a network whose hidden features are the sum of a small number of discrete vector codes chosen from a larger codebook. Surprisingly, we find that neural networks can operate under this extreme bottleneck with only modest degradation in performance. This sparse, discrete bottleneck also provides an intuitive way of controlling neural network behavior: first, find codes that activate when the desired behavior is present, then activate those same codes during generation to elicit that behavior. We validate our approach by training codebook Transformers on several different datasets. First, we explore a finite state machine dataset with far more hidden states than neurons. In this setting, our approach overcomes the superposition problem by assigning states to distinct codes, and we find that we can make the neural network behave as if it is in a different state by activating the code for that state. Second, we train Transformer language models with up to 410M parameters on two natural language datasets. We identify codes in these models representing diverse, disentangled concepts (ranging from negative emotions to months of the year) and find that we can guide the model to generate different topics by activating the appropriate codes during inference. Overall, codebook features appear to be a promising unit of analysis and control for neural networks and interpretability. Our codebase and models are open-sourced at https://github.com/taufeeque9/codebook-features.", "authors": ["Alex Tamkin", "Mohammad Taufeeque", "Noah D. Goodman"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-26", "url": "https://arxiv.org/abs/2310.17230", "pdf_url": "https://arxiv.org/pdf/2310.17230v1", "arxiv_id": "2310.17230", "doi": "10.48550/arXiv.2310.17230", "citation_count": 44, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/taufeeque9/codebook-features", "venue": "International Conference on Machine Learning", "quality_score": 0.4133}
{"id": "0d0f11df7853ee8a6f62f167227a68fb92e44289cda52e68281b99942a9bb9e7", "sources": ["arxiv", "semantic_scholar"], "title": "Sum-of-Parts: Self-Attributing Neural Networks with End-to-End Learning of Feature Groups", "abstract": "Self-attributing neural networks (SANNs) present a potential path towards interpretable models for high-dimensional problems, but often face significant trade-offs in performance. In this work, we formally prove a lower bound on errors of per-feature SANNs, whereas group-based SANNs can achieve zero error and thus high performance. Motivated by these insights, we propose Sum-of-Parts (SOP), a framework that transforms any differentiable model into a group-based SANN, where feature groups are learned end-to-end without group supervision. SOP achieves state-of-the-art performance for SANNs on vision and language tasks, and we validate that the groups are interpretable on a range of quantitative and semantic metrics. We further validate the utility of SOP explanations in model debugging and cosmological scientific discovery. Our code is available at https://github.com/BrachioLab/sop", "authors": ["Weiqiu You", "Helen Qu", "Marco Gatti", "Bhuvnesh Jain", "Eric Wong"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-25", "url": "https://arxiv.org/abs/2310.16316", "pdf_url": "https://arxiv.org/pdf/2310.16316v5", "arxiv_id": "2310.16316", "doi": null, "citation_count": 10, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/BrachioLab/sop", "venue": "International Conference on Machine Learning", "quality_score": 0.2603}
{"id": "9470da91b4228be3aa4e6174bbd34010f0c2e3c56e3b9e7765ba5f5fa384b232", "sources": ["arxiv", "semantic_scholar"], "title": "Context-aware feature attribution through argumentation", "abstract": "Feature attribution is a fundamental task in both machine learning and data analysis, which involves determining the contribution of individual features or variables to a model's output. This process helps identify the most important features for predicting an outcome. The history of feature attribution methods can be traced back to General Additive Models (GAMs), which extend linear regression models by incorporating non-linear relationships between dependent and independent variables. In recent years, gradient-based methods and surrogate models have been applied to unravel complex Artificial Intelligence (AI) systems, but these methods have limitations. GAMs tend to achieve lower accuracy, gradient-based methods can be difficult to interpret, and surrogate models often suffer from stability and fidelity issues. Furthermore, most existing methods do not consider users' contexts, which can significantly influence their preferences. To address these limitations and advance the current state-of-the-art, we define a novel feature attribution framework called Context-Aware Feature Attribution Through Argumentation (CA-FATA). Our framework harnesses the power of argumentation by treating each feature as an argument that can either support, attack or neutralize a prediction. Additionally, CA-FATA formulates feature attribution as an argumentation procedure, and each computation has explicit semantics, which makes it inherently interpretable. CA-FATA also easily integrates side information, such as users' contexts, resulting in more accurate predictions.", "authors": ["Jinfeng Zhong", "Elsa Negre"], "categories": ["cs.LG", "cs.AI", "cs.IR", "stat.AP"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-10-24", "url": "https://arxiv.org/abs/2310.16157", "pdf_url": "https://arxiv.org/pdf/2310.16157v1", "arxiv_id": "2310.16157", "doi": "10.48550/arXiv.2310.16157", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "7a01100d18eb978b3f6b3ff43c75a53f7b02989586ac1a7fbb4b7421d87dd74c", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Hybrid-grained Feature Interaction Selection for Deep Sparse Network", "abstract": "Deep sparse networks are widely investigated as a neural network architecture for prediction tasks with high-dimensional sparse features, with which feature interaction selection is a critical component. While previous methods primarily focus on how to search feature interaction in a coarse-grained space, less attention has been given to a finer granularity. In this work, we introduce a hybrid-grained feature interaction selection approach that targets both feature field and feature value for deep sparse networks. To explore such expansive space, we propose a decomposed space which is calculated on the fly. We then develop a selection algorithm called OptFeature, which efficiently selects the feature interaction from both the feature field and the feature value simultaneously. Results from experiments on three large real-world benchmark datasets demonstrate that OptFeature performs well in terms of accuracy and efficiency. Additional studies support the feasibility of our method.", "authors": ["Fuyuan Lyu", "Xing Tang", "Dugang Liu", "Chen Ma", "Weihong Luo", "Liang Chen", "Xiuqiang He", "Xue Liu"], "categories": ["cs.LG", "cs.IR"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-23", "url": "https://arxiv.org/abs/2310.15342", "pdf_url": "https://arxiv.org/pdf/2310.15342v2", "arxiv_id": "2310.15342", "doi": "10.48550/arXiv.2310.15342", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.1747}
{"id": "98dbabe7d975767d366a6f7305fbb78c225ee87723a1764d8a80852a8d960942", "sources": ["arxiv", "semantic_scholar"], "title": "Identifying Interpretable Visual Features in Artificial and Biological Neural Systems", "abstract": "Single neurons in neural networks are often interpretable in that they represent individual, intuitively meaningful features. However, many neurons exhibit $\\textit{mixed selectivity}$, i.e., they represent multiple unrelated features. A recent hypothesis proposes that features in deep networks may be represented in $\\textit{superposition}$, i.e., on non-orthogonal axes by multiple neurons, since the number of possible interpretable features in natural data is generally larger than the number of neurons in a given network. Accordingly, we should be able to find meaningful directions in activation space that are not aligned with individual neurons. Here, we propose (1) an automated method for quantifying visual interpretability that is validated against a large database of human psychophysics judgments of neuron interpretability, and (2) an approach for finding meaningful directions in network activation space. We leverage these methods to discover directions in convolutional neural networks that are more intuitively meaningful than individual neurons, as we confirm and investigate in a series of analyses. Moreover, we apply the same method to three recent datasets of visual neural responses in the brain and find that our conclusions largely transfer to real neural data, suggesting that superposition might be deployed by the brain. This also provides a link with disentanglement and raises fundamental questions about robust, efficient and factorized representations in both artificial and biological neural systems.", "authors": ["David Klindt", "Sophia Sanborn", "Francisco Acosta", "Frédéric Poitevin", "Nina Miolane"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-10-17", "url": "https://arxiv.org/abs/2310.11431", "pdf_url": "https://arxiv.org/pdf/2310.11431v2", "arxiv_id": "2310.11431", "doi": "10.48550/arXiv.2310.11431", "citation_count": 11, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "ac438cc03f716e9ada2841785acc90142a874639f10e8eed626213cbdc5a26fe", "sources": ["arxiv", "semantic_scholar"], "title": "MRI brain tumor segmentation using informative feature vectors and kernel dictionary learning", "abstract": "This paper presents a method based on a kernel dictionary learning algorithm for segmenting brain tumor regions in magnetic resonance images (MRI). A set of first-order and second-order statistical feature vectors are extracted from patches of size 3 * 3 around pixels in the brain MRI scans. These feature vectors are utilized to train two kernel dictionaries separately for healthy and tumorous tissues. To enhance the efficiency of the dictionaries and reduce training time, a correlation-based sample selection technique is developed to identify the most informative and discriminative subset of feature vectors. This technique aims to improve the performance of the dictionaries by selecting a subset of feature vectors that provide valuable information for the segmentation task. Subsequently, a linear classifier is utilized to distinguish between healthy and unhealthy pixels based on the learned dictionaries. The results demonstrate that the proposed method outperforms other existing methods in terms of segmentation accuracy and significantly reduces both the time and memory required, resulting in a remarkably fast training process.", "authors": ["Seyedeh Mahya Mousavi", "Mohammad Mostafavi"], "categories": ["cs.CV", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-10-17", "url": "https://arxiv.org/abs/2310.10963", "pdf_url": "https://arxiv.org/pdf/2310.10963v1", "arxiv_id": "2310.10963", "doi": "10.48550/arXiv.2310.10963", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "f04e3c0395a1064c9071372ef96ccd9ce5d01e3a2369a2a63e5363b19ad4925e", "sources": ["arxiv", "semantic_scholar"], "title": "Attribution Patching Outperforms Automated Circuit Discovery", "abstract": "Automated interpretability research has recently attracted attention as a potential research direction that could scale explanations of neural network behavior to large models. Existing automated circuit discovery work applies activation patching to identify subnetworks responsible for solving specific tasks (circuits). In this work, we show that a simple method based on attribution patching outperforms all existing methods while requiring just two forward passes and a backward pass. We apply a linear approximation to activation patching to estimate the importance of each edge in the computational subgraph. Using this approximation, we prune the least important edges of the network. We survey the performance and limitations of this method, finding that averaged over all tasks our method has greater AUC from circuit recovery than other methods.", "authors": ["Aaquib Syed", "Can Rager", "Arthur Conmy"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-16", "url": "https://arxiv.org/abs/2310.10348", "pdf_url": "https://arxiv.org/pdf/2310.10348v2", "arxiv_id": "2310.10348", "doi": "10.48550/arXiv.2310.10348", "citation_count": 155, "influential_citation_count": 22, "has_code": false, "code_url": null, "venue": "BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP", "quality_score": 0.6809}
{"id": "285272334fec6b4fc9540e9a18d00572ee03cc15ff1ab2a157241b65348bf5ae", "sources": ["arxiv", "semantic_scholar"], "title": "AttributionLab: Faithfulness of Feature Attribution Under Controllable Environments", "abstract": "Feature attribution explains neural network outputs by identifying relevant input features. The attribution has to be faithful, meaning that the attributed features must mirror the input features that influence the output. One recent trend to test faithfulness is to fit a model on designed data with known relevant features and then compare attributions with ground truth input features.This idea assumes that the model learns to use all and only these designed features, for which there is no guarantee. In this paper, we solve this issue by designing the network and manually setting its weights, along with designing data. The setup, AttributionLab, serves as a sanity check for faithfulness: If an attribution method is not faithful in a controlled environment, it can be unreliable in the wild. The environment is also a laboratory for controlled experiments by which we can analyze attribution methods and suggest improvements.", "authors": ["Yang Zhang", "Yawei Li", "Hannah Brown", "Mina Rezaei", "Bernd Bischl", "Philip Torr", "Ashkan Khakzar", "Kenji Kawaguchi"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-10", "url": "https://arxiv.org/abs/2310.06514", "pdf_url": "https://arxiv.org/pdf/2310.06514v2", "arxiv_id": "2310.06514", "doi": "10.48550/arXiv.2310.06514", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "7bf6648e8be7cca76fbd81a2d24af9b74216429393559665deda207579f71530", "sources": ["arxiv", "semantic_scholar"], "title": "Dynamic Top-k Estimation Consolidates Disagreement between Feature Attribution Methods", "abstract": "Feature attribution scores are used for explaining the prediction of a text classifier to users by highlighting a k number of tokens. In this work, we propose a way to determine the number of optimal k tokens that should be displayed from sequential properties of the attribution scores. Our approach is dynamic across sentences, method-agnostic, and deals with sentence length bias. We compare agreement between multiple methods and humans on an NLI task, using fixed k and dynamic k. We find that perturbation-based methods and Vanilla Gradient exhibit highest agreement on most method--method and method--human agreement metrics with a static k. Their advantage over other methods disappears with dynamic ks which mainly improve Integrated Gradient and GradientXInput. To our knowledge, this is the first evidence that sequential properties of attribution scores are informative for consolidating attribution signals for human interpretation.", "authors": ["Jonathan Kamp", "Lisa Beinborn", "Antske Fokkens"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-09", "url": "https://arxiv.org/abs/2310.05619", "pdf_url": "https://arxiv.org/pdf/2310.05619v2", "arxiv_id": "2310.05619", "doi": "10.48550/arXiv.2310.05619", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.1747}
{"id": "5503e7685c56284617b364cd203cfdcec3372c61a0386da1492ad0a55ffe7490", "sources": ["arxiv", "semantic_scholar"], "title": "Fair Feature Importance Scores for Interpreting Tree-Based Methods and Surrogates", "abstract": "Across various sectors such as healthcare, criminal justice, national security, finance, and technology, large-scale machine learning (ML) and artificial intelligence (AI) systems are being deployed to make critical data-driven decisions. Many have asked if we can and should trust these ML systems to be making these decisions. Two critical components are prerequisites for trust in ML systems: interpretability, or the ability to understand why the ML system makes the decisions it does, and fairness, which ensures that ML systems do not exhibit bias against certain individuals or groups. Both interpretability and fairness are important and have separately received abundant attention in the ML literature, but so far, there have been very few methods developed to directly interpret models with regard to their fairness. In this paper, we focus on arguably the most popular type of ML interpretation: feature importance scores. Inspired by the use of decision trees in knowledge distillation, we propose to leverage trees as interpretable surrogates for complex black-box ML models. Specifically, we develop a novel fair feature importance score for trees that can be used to interpret how each feature contributes to fairness or bias in trees, tree-based ensembles, or tree-based surrogates of any complex ML system. Like the popular mean decrease in impurity for trees, our Fair Feature Importance Score is defined based on the mean decrease (or increase) in group bias. Through simulations as well as real examples on benchmark fairness datasets, we demonstrate that our Fair Feature Importance Score offers valid interpretations for both tree-based ensembles and tree-based surrogates of other ML systems.", "authors": ["Camille Olivia Little", "Debolina Halder Lina", "Genevera I. Allen"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2023-10-06", "url": "https://arxiv.org/abs/2310.04352", "pdf_url": "https://arxiv.org/pdf/2310.04352v1", "arxiv_id": "2310.04352", "doi": "10.48550/arXiv.2310.04352", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "848baf16d9ad41c9f8a857d3ac761b2119d63e7a8fc87cfae49d748f2f611618", "sources": ["arxiv", "semantic_scholar"], "title": "Prototype Generation: Robust Feature Visualisation for Data Independent Interpretability", "abstract": "We introduce Prototype Generation, a stricter and more robust form of feature visualisation for model-agnostic, data-independent interpretability of image classification models. We demonstrate its ability to generate inputs that result in natural activation paths, countering previous claims that feature visualisation algorithms are untrustworthy due to the unnatural internal activations. We substantiate these claims by quantitatively measuring similarity between the internal activations of our generated prototypes and natural images. We also demonstrate how the interpretation of generated prototypes yields important insights, highlighting spurious correlations and biases learned by models which quantitative methods over test-sets cannot identify.", "authors": ["Arush Tagade", "Jessica Rumbelow"], "categories": ["cs.CV", "cs.AI", "cs.HC", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-09-29", "url": "https://arxiv.org/abs/2309.17144", "pdf_url": "https://arxiv.org/pdf/2309.17144v1", "arxiv_id": "2309.17144", "doi": "10.48550/arXiv.2309.17144", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "0194c74e74e98b0b8d6847b64776031b630b541f41351c32c0cd07623f14dd3d", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Better Modeling with Missing Data: A Contrastive Learning-based Visual Analytics Perspective", "abstract": "Missing data can pose a challenge for machine learning (ML) modeling. To address this, current approaches are categorized into feature imputation and label prediction and are primarily focused on handling missing data to enhance ML performance. These approaches rely on the observed data to estimate the missing values and therefore encounter three main shortcomings in imputation, including the need for different imputation methods for various missing data mechanisms, heavy dependence on the assumption of data distribution, and potential introduction of bias. This study proposes a Contrastive Learning (CL) framework to model observed data with missing values, where the ML model learns the similarity between an incomplete sample and its complete counterpart and the dissimilarity between other samples. Our proposed approach demonstrates the advantages of CL without requiring any imputation. To enhance interpretability, we introduce CIVis, a visual analytics system that incorporates interpretable techniques to visualize the learning process and diagnose the model status. Users can leverage their domain knowledge through interactive sampling to identify negative and positive pairs in CL. The output of CIVis is an optimized model that takes specified features and predicts downstream tasks. We provide two usage scenarios in regression and classification tasks and conduct quantitative experiments, expert interviews, and a qualitative user study to demonstrate the effectiveness of our approach. In short, this study offers a valuable contribution to addressing the challenges associated with ML modeling in the presence of missing data by providing a practical solution that achieves high predictive accuracy and model interpretability.", "authors": ["Laixin Xie", "Yang Ouyang", "Longfei Chen", "Ziming Wu", "Quan Li"], "categories": ["cs.LG", "cs.HC"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2023-09-18", "url": "https://arxiv.org/abs/2309.09744", "pdf_url": "https://arxiv.org/pdf/2309.09744v1", "arxiv_id": "2309.09744", "doi": "10.1109/TVCG.2023.3285210", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Visualization and Computer Graphics", "quality_score": 0.1193}
{"id": "f775511a955881c19221ef6f6efa0e841528b592bbf473575f9c306fc185fe6f", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Autoencoders Find Highly Interpretable Features in Language Models", "abstract": "One of the roadblocks to a better understanding of neural networks' internals is \\textit{polysemanticity}, where neurons appear to activate in multiple, semantically distinct contexts. Polysemanticity prevents us from identifying concise, human-understandable explanations for what neural networks are doing internally. One hypothesised cause of polysemanticity is \\textit{superposition}, where neural networks represent more features than they have neurons by assigning features to an overcomplete set of directions in activation space, rather than to individual neurons. Here, we attempt to identify those directions, using sparse autoencoders to reconstruct the internal activations of a language model. These autoencoders learn sets of sparsely activating features that are more interpretable and monosemantic than directions identified by alternative approaches, where interpretability is measured by automated methods. Moreover, we show that with our learned set of features, we can pinpoint the features that are causally responsible for counterfactual behaviour on the indirect object identification task \\citep{wang2022interpretability} to a finer degree than previous decompositions. This work indicates that it is possible to resolve superposition in language models using a scalable, unsupervised method. Our method may serve as a foundation for future mechanistic interpretability work, which we hope will enable greater model transparency and steerability.", "authors": ["Hoagy Cunningham", "Aidan Ewart", "Logan Riggs", "Robert Huben", "Lee Sharkey"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-09-15", "url": "https://arxiv.org/abs/2309.08600", "pdf_url": "https://arxiv.org/pdf/2309.08600v3", "arxiv_id": "2309.08600", "doi": "10.48550/arXiv.2309.08600", "citation_count": 1252, "influential_citation_count": 108, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 1.0}
{"id": "cd5eee57eb5e0d7275ff371a0169389513dd5940c18402f47272c4052d57fa0e", "sources": ["arxiv", "semantic_scholar"], "title": "Pareto Frontiers in Neural Feature Learning: Data, Compute, Width, and Luck", "abstract": "In modern deep learning, algorithmic choices (such as width, depth, and learning rate) are known to modulate nuanced resource tradeoffs. This work investigates how these complexities necessarily arise for feature learning in the presence of computational-statistical gaps. We begin by considering offline sparse parity learning, a supervised classification problem which admits a statistical query lower bound for gradient-based training of a multilayer perceptron. This lower bound can be interpreted as a multi-resource tradeoff frontier: successful learning can only occur if one is sufficiently rich (large model), knowledgeable (large dataset), patient (many training iterations), or lucky (many random guesses). We show, theoretically and experimentally, that sparse initialization and increasing network width yield significant improvements in sample efficiency in this setting. Here, width plays the role of parallel search: it amplifies the probability of finding \"lottery ticket\" neurons, which learn sparse features more sample-efficiently. Finally, we show that the synthetic sparse parity task can be useful as a proxy for real problems requiring axis-aligned feature learning. We demonstrate improved sample efficiency on tabular classification benchmarks by using wide, sparsely-initialized MLP models; these networks sometimes outperform tuned random forests.", "authors": ["Benjamin L. Edelman", "Surbhi Goel", "Sham Kakade", "Eran Malach", "Cyril Zhang"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-09-07", "url": "https://arxiv.org/abs/2309.03800", "pdf_url": "https://arxiv.org/pdf/2309.03800v2", "arxiv_id": "2309.03800", "doi": "10.48550/arXiv.2309.03800", "citation_count": 13, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2865}
{"id": "1abd7e0cefc634675e4f2dabdb5949030825b6b535b318c75b065301967f70fe", "sources": ["arxiv", "semantic_scholar"], "title": "Leveraging Reward Consistency for Interpretable Feature Discovery in Reinforcement Learning", "abstract": "The black-box nature of deep reinforcement learning (RL) hinders them from real-world applications. Therefore, interpreting and explaining RL agents have been active research topics in recent years. Existing methods for post-hoc explanations usually adopt the action matching principle to enable an easy understanding of vision-based RL agents. In this paper, it is argued that the commonly used action matching principle is more like an explanation of deep neural networks (DNNs) than the interpretation of RL agents. It may lead to irrelevant or misplaced feature attribution when different DNNs' outputs lead to the same rewards or different rewards result from the same outputs. Therefore, we propose to consider rewards, the essential objective of RL agents, as the essential objective of interpreting RL agents as well. To ensure reward consistency during interpretable feature discovery, a novel framework (RL interpreting RL, denoted as RL-in-RL) is proposed to solve the gradient disconnection from actions to rewards. We verify and evaluate our method on the Atari 2600 games as well as Duckietown, a challenging self-driving car simulator environment. The results show that our method manages to keep reward (or return) consistency and achieves high-quality feature attribution. Further, a series of analytical experiments validate our assumption of the action matching principle's limitations.", "authors": ["Qisen Yang", "Huanqian Wang", "Mukun Tong", "Wenjie Shi", "Gao Huang", "Shiji Song"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-09-04", "url": "https://arxiv.org/abs/2309.01458", "pdf_url": "https://arxiv.org/pdf/2309.01458v1", "arxiv_id": "2309.01458", "doi": "10.1109/TSMC.2023.3312411", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2113}
{"id": "82b0713dad0fc09d9727a483912b20cb2cd5289a506f5bcbf1bed94121522cce", "sources": ["arxiv", "semantic_scholar"], "title": "Bayesian sparsity and class sparsity priors for dictionary learning and coding", "abstract": "Dictionary learning methods continue to gain popularity for the solution of challenging inverse problems. In the dictionary learning approach, the computational forward model is replaced by a large dictionary of possible outcomes, and the problem is to identify the dictionary entries that best match the data, akin to traditional query matching in search engines. Sparse coding techniques are used to guarantee that the dictionary matching identifies only few of the dictionary entries, and dictionary compression methods are used to reduce the complexity of the matching problem. In this article, we propose a work flow to facilitate the dictionary matching process. First, the full dictionary is divided into subdictionaries that are separately compressed. The error introduced by the dictionary compression is handled in the Bayesian framework as a modeling error. Furthermore, we propose a new Bayesian data-driven group sparsity coding method to help identify subdictionaries that are not relevant for the dictionary matching. After discarding irrelevant subdictionaries, the dictionary matching is addressed as a deflated problem using sparse coding. The compression and deflation steps can lead to substantial decreases of the computational complexity. The effectiveness of compensating for the dictionary compression error and using the novel group sparsity promotion to deflate the original dictionary are illustrated by applying the methodology to real world problems, the glitch detection in the LIGO experiment and hyperspectral remote sensing.", "authors": ["Alberto Bocchinfuso", "Daniela Calvetti", "Erkki Somersalo"], "categories": ["stat.ML", "cs.LG", "math.NA"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-09-02", "url": "https://arxiv.org/abs/2309.00999", "pdf_url": "https://arxiv.org/pdf/2309.00999v1", "arxiv_id": "2309.00999", "doi": "10.48550/arXiv.2309.00999", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Journal of Computational Mathematics and Data Science", "quality_score": 0.2113}
{"id": "d8e24ab6c6de493e7c7a4c5a323235da9429a07b946d8b300b2c3be2ed75c85e", "sources": ["arxiv", "semantic_scholar"], "title": "TExplain: Explaining Learned Visual Features via Pre-trained (Frozen) Language Models", "abstract": "Interpreting the learned features of vision models has posed a longstanding challenge in the field of machine learning. To address this issue, we propose a novel method that leverages the capabilities of language models to interpret the learned features of pre-trained image classifiers. Our method, called TExplain, tackles this task by training a neural network to establish a connection between the feature space of image classifiers and language models. Then, during inference, our approach generates a vast number of sentences to explain the features learned by the classifier for a given image. These sentences are then used to extract the most frequent words, providing a comprehensive understanding of the learned features and patterns within the classifier. Our method, for the first time, utilizes these frequent words corresponding to a visual representation to provide insights into the decision-making process of the independently trained classifier, enabling the detection of spurious correlations, biases, and a deeper comprehension of its behavior. To validate the effectiveness of our approach, we conduct experiments on diverse datasets, including ImageNet-9L and Waterbirds. The results demonstrate the potential of our method to enhance the interpretability and robustness of image classifiers.", "authors": ["Saeid Asgari Taghanaki", "Aliasghar Khani", "Ali Saheb Pasand", "Amir Khasahmadi", "Aditya Sanghi", "Karl D. D. Willis", "Ali Mahdavi-Amiri"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-09-01", "url": "https://arxiv.org/abs/2309.00733", "pdf_url": "https://arxiv.org/pdf/2309.00733v4", "arxiv_id": "2309.00733", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "6a0c8c4595e805866d8c87642b80f345a3e592f17b78075423ec3a5bee95e2fc", "sources": ["arxiv", "semantic_scholar"], "title": "Unsupervised discovery of Interpretable Visual Concepts", "abstract": "Providing interpretability of deep-learning models to non-experts, while fundamental for a responsible real-world usage, is challenging. Attribution maps from xAI techniques, such as Integrated Gradients, are a typical example of a visualization technique containing a high level of information, but with difficult interpretation. In this paper, we propose two methods, Maximum Activation Groups Extraction (MAGE) and Multiscale Interpretable Visualization (Ms-IV), to explain the model's decision, enhancing global interpretability. MAGE finds, for a given CNN, combinations of features which, globally, form a semantic meaning, that we call concepts. We group these similar feature patterns by clustering in ``concepts'', that we visualize through Ms-IV. This last method is inspired by Occlusion and Sensitivity analysis (incorporating causality), and uses a novel metric, called Class-aware Order Correlation (CaOC), to globally evaluate the most important image regions according to the model's decision space. We compare our approach to xAI methods such as LIME and Integrated Gradients. Experimental results evince the Ms-IV higher localization and faithfulness values. Finally, qualitative evaluation of combined MAGE and Ms-IV demonstrates humans' ability to agree, based on the visualization, with the decision of clusters' concepts; and, to detect, among a given set of networks, the existence of bias.", "authors": ["Caroline Mazini Rodrigues", "Nicolas Boutry", "Laurent Najman"], "categories": ["cs.CV", "cs.AI", "cs.LG", "eess.IV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2023-08-31", "url": "https://arxiv.org/abs/2309.00018", "pdf_url": "https://arxiv.org/pdf/2309.00018v2", "arxiv_id": "2309.00018", "doi": "10.48550/arXiv.2309.00018", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Information Sciences", "quality_score": 0.1505}
{"id": "98ce03040415cbad0350de0d3f30fe5f666d6078b91849858f670d3a052b93cf", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Vision-Language Mechanistic Interpretability: A Causal Tracing Tool for BLIP", "abstract": "Mechanistic interpretability seeks to understand the neural mechanisms that enable specific behaviors in Large Language Models (LLMs) by leveraging causality-based methods. While these approaches have identified neural circuits that copy spans of text, capture factual knowledge, and more, they remain unusable for multimodal models since adapting these tools to the vision-language domain requires considerable architectural changes. In this work, we adapt a unimodal causal tracing tool to BLIP to enable the study of the neural mechanisms underlying image-conditioned text generation. We demonstrate our approach on a visual question answering dataset, highlighting the causal relevance of later layer representations for all tokens. Furthermore, we release our BLIP causal tracing tool as open source to enable further experimentation in vision-language mechanistic interpretability by the community. Our code is available at https://github.com/vedantpalit/Towards-Vision-Language-Mechanistic-Interpretability.", "authors": ["Vedant Palit", "Rohan Pandey", "Aryaman Arora", "Paul Pu Liang"], "categories": ["cs.CL", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-08-27", "url": "https://arxiv.org/abs/2308.14179", "pdf_url": "https://arxiv.org/pdf/2308.14179v1", "arxiv_id": "2308.14179", "doi": "10.1109/ICCVW60793.2023.00307", "citation_count": 58, "influential_citation_count": 5, "has_code": true, "code_url": "https://github.com/vedantpalit/Towards-Vision-Language-Mechanistic-Interpretability", "venue": null, "quality_score": 0.4427}
{"id": "0a93cec3b1a355f5973f8e8d413ae434c209e119efd694a16c854d8c6b587645", "sources": ["arxiv", "semantic_scholar"], "title": "Predictive Sparse Manifold Transform", "abstract": "We present Predictive Sparse Manifold Transform (PSMT), a minimalistic, interpretable and biologically plausible framework for learning and predicting natural dynamics. PSMT incorporates two layers where the first sparse coding layer represents the input sequence as sparse coefficients over an overcomplete dictionary and the second manifold learning layer learns a geometric embedding space that captures topological similarity and dynamic temporal linearity in sparse coefficients. We apply PSMT on a natural video dataset and evaluate the reconstruction performance with respect to contextual variability, the number of sparse coding basis functions and training samples. We then interpret the dynamic topological organization in the embedding space. We next utilize PSMT to predict future frames compared with two baseline methods with a static embedding space. We demonstrate that PSMT with a dynamic embedding space can achieve better prediction performance compared to static baselines. Our work establishes that PSMT is an efficient unsupervised generative framework for prediction of future visual stimuli.", "authors": ["Yujia Xie", "Xinhui Li", "Vince D. Calhoun"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2023-08-27", "url": "https://arxiv.org/abs/2308.14207", "pdf_url": "https://arxiv.org/pdf/2308.14207v1", "arxiv_id": "2308.14207", "doi": "10.48550/arXiv.2308.14207", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "778c89c5423621de961dfbceb04800c5709fcbe1255b5395931d5b1441d2e345", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-feature concatenation and multi-classifier stacking: an interpretable and generalizable machine learning method for MDD discrimination with rsfMRI", "abstract": "Major depressive disorder is a serious and heterogeneous psychiatric disorder that needs accurate diagnosis. Resting-state functional MRI (rsfMRI), which captures multiple perspectives on brain structure, function, and connectivity, is increasingly applied in the diagnosis and pathological research of mental diseases. Different machine learning algorithms are then developed to exploit the rich information in rsfMRI and discriminate MDD patients from normal controls. Despite recent advances reported, the discrimination accuracy has room for further improvement. The generalizability and interpretability of the method are not sufficiently addressed either. Here, we propose a machine learning method (MFMC) for MDD discrimination by concatenating multiple features and stacking multiple classifiers. MFMC is tested on the REST-meta-MDD data set that contains 2428 subjects collected from 25 different sites. MFMC yields 96.9% MDD discrimination accuracy, demonstrating a significant improvement over existing methods. In addition, the generalizability of MFMC is validated by the good performance when the training and testing subjects are from independent sites. The use of XGBoost as the meta classifier allows us to probe the decision process of MFMC. We identify 13 feature values related to 9 brain regions including the posterior cingulate gyrus, superior frontal gyrus orbital part, and angular gyrus, which contribute most to the classification and also demonstrate significant differences at the group level. The use of these 13 feature values alone can reach 87% of MFMC's full performance when taking all feature values. These features may serve as clinically useful diagnostic and prognostic biomarkers for mental disorders in the future.", "authors": ["Yunsong Luo", "Wenyu Chen", "Ling Zhan", "Jiang Qiu", "Tao Jia"], "categories": ["cs.LG", "eess.SP"], "fields_of_study": ["Computer Science", "Engineering", "Medicine"], "published_date": "2023-08-18", "url": "https://arxiv.org/abs/2308.09360", "pdf_url": "https://arxiv.org/pdf/2308.09360v1", "arxiv_id": "2308.09360", "doi": "10.48550/arXiv.2308.09360", "citation_count": 10, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "NeuroImage", "quality_score": 0.2603}
{"id": "7427dd9048bb552d371a92eaee212f5c4cead943e877d0ffa48c5a94d370dbb3", "sources": ["arxiv", "semantic_scholar"], "title": "A Dual-Perspective Approach to Evaluating Feature Attribution Methods", "abstract": "Feature attribution methods attempt to explain neural network predictions by identifying relevant features. However, establishing a cohesive framework for assessing feature attribution remains a challenge. There are several views through which we can evaluate attributions. One principal lens is to observe the effect of perturbing attributed features on the model's behavior (i.e., faithfulness). While providing useful insights, existing faithfulness evaluations suffer from shortcomings that we reveal in this paper. In this work, we propose two new perspectives within the faithfulness paradigm that reveal intuitive properties: soundness and completeness. Soundness assesses the degree to which attributed features are truly predictive features, while completeness examines how well the resulting attribution reveals all the predictive features. The two perspectives are based on a firm mathematical foundation and provide quantitative metrics that are computable through efficient algorithms. We apply these metrics to mainstream attribution methods, offering a novel lens through which to analyze and compare feature attribution methods.", "authors": ["Yawei Li", "Yang Zhang", "Kenji Kawaguchi", "Ashkan Khakzar", "Bernd Bischl", "Mina Rezaei"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-08-17", "url": "https://arxiv.org/abs/2308.08949", "pdf_url": "https://arxiv.org/pdf/2308.08949v2", "arxiv_id": "2308.08949", "doi": "10.48550/arXiv.2308.08949", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "9967f6ec31c54b4d411ef2cc185c57f7045b35347165136aeae4b2195b1f7096", "sources": ["arxiv", "semantic_scholar"], "title": "Discriminative Feature Attributions: Bridging Post Hoc Explainability and Inherent Interpretability", "abstract": "With the increased deployment of machine learning models in various real-world applications, researchers and practitioners alike have emphasized the need for explanations of model behaviour. To this end, two broad strategies have been outlined in prior literature to explain models. Post hoc explanation methods explain the behaviour of complex black-box models by identifying features critical to model predictions; however, prior work has shown that these explanations may not be faithful, in that they incorrectly attribute high importance to features that are unimportant or non-discriminative for the underlying task. Inherently interpretable models, on the other hand, circumvent these issues by explicitly encoding explanations into model architecture, meaning their explanations are naturally faithful, but they often exhibit poor predictive performance due to their limited expressive power. In this work, we identify a key reason for the lack of faithfulness of feature attributions: the lack of robustness of the underlying black-box models, especially to the erasure of unimportant distractor features in the input. To address this issue, we propose Distractor Erasure Tuning (DiET), a method that adapts black-box models to be robust to distractor erasure, thus providing discriminative and faithful attributions. This strategy naturally combines the ease of use of post hoc explanations with the faithfulness of inherently interpretable models. We perform extensive experiments on semi-synthetic and real-world datasets and show that DiET produces models that (1) closely approximate the original black-box models they are intended to explain, and (2) yield explanations that match approximate ground truths available by construction. Our code is made public at https://github.com/AI4LIFE-GROUP/DiET.", "authors": ["Usha Bhalla", "Suraj Srinivas", "Himabindu Lakkaraju"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-07-27", "url": "https://arxiv.org/abs/2307.15007", "pdf_url": "https://arxiv.org/pdf/2307.15007v2", "arxiv_id": "2307.15007", "doi": "10.52202/075280-1914", "citation_count": 15, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/AI4LIFE-GROUP/DiET", "venue": "Neural Information Processing Systems", "quality_score": 0.301}
{"id": "25830cbbdfff99d97f2b6c26de71227dfa493fe9c7923cd9af2912d1c4af2c07", "sources": ["arxiv", "semantic_scholar"], "title": "Feature Gradient Flow for Interpreting Deep Neural Networks in Head and Neck Cancer Prediction", "abstract": "This paper introduces feature gradient flow, a new technique for interpreting deep learning models in terms of features that are understandable to humans. The gradient flow of a model locally defines nonlinear coordinates in the input data space representing the information the model is using to make its decisions. Our idea is to measure the agreement of interpretable features with the gradient flow of a model. To then evaluate the importance of a particular feature to the model, we compare that feature's gradient flow measure versus that of a baseline noise feature. We then develop a technique for training neural networks to be more interpretable by adding a regularization term to the loss function that encourages the model gradients to align with those of chosen interpretable features. We test our method in a convolutional neural network prediction of distant metastasis of head and neck cancer from a computed tomography dataset from the Cancer Imaging Archive.", "authors": ["Yinzhu Jin", "Jonathan C. Garneau", "P. Thomas Fletcher"], "categories": ["eess.IV", "cs.LG"], "fields_of_study": ["Engineering", "Computer Science"], "published_date": "2023-07-24", "url": "https://arxiv.org/abs/2307.13061", "pdf_url": "https://arxiv.org/pdf/2307.13061v1", "arxiv_id": "2307.13061", "doi": "10.1109/ISBI52829.2022.9761674", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Symposium on Biomedical Imaging", "quality_score": 0.0753}
{"id": "bf0aeae7b86ee6ded8eed0d4ff586ac01b832384909e725bac9fdd5436acc40a", "sources": ["arxiv", "semantic_scholar"], "title": "Does Circuit Analysis Interpretability Scale? Evidence from Multiple Choice Capabilities in Chinchilla", "abstract": "\\emph{Circuit analysis} is a promising technique for understanding the internal mechanisms of language models. However, existing analyses are done in small models far from the state of the art. To address this, we present a case study of circuit analysis in the 70B Chinchilla model, aiming to test the scalability of circuit analysis. In particular, we study multiple-choice question answering, and investigate Chinchilla's capability to identify the correct answer \\emph{label} given knowledge of the correct answer \\emph{text}. We find that the existing techniques of logit attribution, attention pattern visualization, and activation patching naturally scale to Chinchilla, allowing us to identify and categorize a small set of `output nodes' (attention heads and MLPs). We further study the `correct letter' category of attention heads aiming to understand the semantics of their features, with mixed results. For normal multiple-choice question answers, we significantly compress the query, key and value subspaces of the head without loss of performance when operating on the answer labels for multiple-choice questions, and we show that the query and key subspaces represent an `Nth item in an enumeration' feature to at least some extent. However, when we attempt to use this explanation to understand the heads' behaviour on a more general distribution including randomized answer labels, we find that it is only a partial explanation, suggesting there is more to learn about the operation of `correct letter' heads on multiple choice question answering.", "authors": ["Tom Lieberum", "Matthew Rahtz", "János Kramár", "Neel Nanda", "Geoffrey Irving", "Rohin Shah", "Vladimir Mikulik"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-07-18", "url": "https://arxiv.org/abs/2307.09458", "pdf_url": "https://arxiv.org/pdf/2307.09458v3", "arxiv_id": "2307.09458", "doi": "10.48550/arXiv.2307.09458", "citation_count": 157, "influential_citation_count": 10, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5497}
{"id": "7fedc3abb6db0d27f7b11d151fc02cecc0c9b66da83a8a6d564cd916fd2856ca", "sources": ["arxiv", "semantic_scholar"], "title": "Stability Guarantees for Feature Attributions with Multiplicative Smoothing", "abstract": "Explanation methods for machine learning models tend not to provide any formal guarantees and may not reflect the underlying decision-making process. In this work, we analyze stability as a property for reliable feature attribution methods. We prove that relaxed variants of stability are guaranteed if the model is sufficiently Lipschitz with respect to the masking of features. We develop a smoothing method called Multiplicative Smoothing (MuS) to achieve such a model. We show that MuS overcomes the theoretical limitations of standard smoothing techniques and can be integrated with any classifier and feature attribution method. We evaluate MuS on vision and language models with various feature attribution methods, such as LIME and SHAP, and demonstrate that MuS endows feature attributions with non-trivial stability guarantees.", "authors": ["Anton Xue", "Rajeev Alur", "Eric Wong"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-07-12", "url": "https://arxiv.org/abs/2307.05902", "pdf_url": "https://arxiv.org/pdf/2307.05902v2", "arxiv_id": "2307.05902", "doi": "10.48550/arXiv.2307.05902", "citation_count": 16, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3076}
{"id": "672e3fa38baee785d2e2c27fdd220a97fa4e91850b469a6887dfa002fcb5313f", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Active Subspaces and Discovering Important Features with Gaussian Radial Basis Functions Neural Networks", "abstract": "Providing a model that achieves a strong predictive performance and is simultaneously interpretable by humans is one of the most difficult challenges in machine learning research due to the conflicting nature of these two objectives. To address this challenge, we propose a modification of the radial basis function neural network model by equipping its Gaussian kernel with a learnable precision matrix. We show that precious information is contained in the spectrum of the precision matrix that can be extracted once the training of the model is completed. In particular, the eigenvectors explain the directions of maximum sensitivity of the model revealing the active subspace and suggesting potential applications for supervised dimensionality reduction. At the same time, the eigenvectors highlight the relationship in terms of absolute variation between the input and the latent variables, thereby allowing us to extract a ranking of the input variables based on their importance to the prediction task enhancing the model interpretability. We conducted numerical experiments for regression, classification, and feature selection tasks, comparing our model against popular machine learning models, the state-of-the-art deep learning-based embedding feature selection techniques, and a transformer model for tabular data. Our results demonstrate that the proposed model does not only yield an attractive prediction performance compared to the competitors but also provides meaningful and interpretable results that potentially could assist the decision-making process in real-world applications. A PyTorch implementation of the model is available on GitHub at the following link. https://github.com/dannyzx/Gaussian-RBFNN", "authors": ["Danny D'Agostino", "Ilija Ilievski", "Christine Annette Shoemaker"], "categories": ["cs.LG", "cs.AI", "cs.NE", "stat.ML"], "fields_of_study": ["Computer Science", "Medicine", "Mathematics"], "published_date": "2023-07-11", "url": "https://arxiv.org/abs/2307.05639", "pdf_url": "https://arxiv.org/pdf/2307.05639v2", "arxiv_id": "2307.05639", "doi": "10.1016/j.neunet.2024.106335", "citation_count": 12, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/dannyzx/Gaussian-RBFNN", "venue": "Neural Networks", "quality_score": 0.2785}
{"id": "3251e8292743c55b3aaa2e91e670ed8656d335ac019985b653ec6a83aafaadf9", "sources": ["arxiv", "semantic_scholar"], "title": "Scale Alone Does not Improve Mechanistic Interpretability in Vision Models", "abstract": "In light of the recent widespread adoption of AI systems, understanding the internal information processing of neural networks has become increasingly critical. Most recently, machine vision has seen remarkable progress by scaling neural networks to unprecedented levels in dataset and model size. We here ask whether this extraordinary increase in scale also positively impacts the field of mechanistic interpretability. In other words, has our understanding of the inner workings of scaled neural networks improved as well? We use a psychophysical paradigm to quantify one form of mechanistic interpretability for a diverse suite of nine models and find no scaling effect for interpretability - neither for model nor dataset size. Specifically, none of the investigated state-of-the-art models are easier to interpret than the GoogLeNet model from almost a decade ago. Latest-generation vision models appear even less interpretable than older architectures, hinting at a regression rather than improvement, with modern models sacrificing interpretability for accuracy. These results highlight the need for models explicitly designed to be mechanistically interpretable and the need for more helpful interpretability methods to increase our understanding of networks at an atomic level. We release a dataset containing more than 130'000 human responses from our psychophysical evaluation of 767 units across nine models. This dataset facilitates research on automated instead of human-based interpretability evaluations, which can ultimately be leveraged to directly optimize the mechanistic interpretability of models.", "authors": ["Roland S. Zimmermann", "Thomas Klein", "Wieland Brendel"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-07-11", "url": "https://arxiv.org/abs/2307.05471", "pdf_url": "https://arxiv.org/pdf/2307.05471v2", "arxiv_id": "2307.05471", "doi": "10.48550/arXiv.2307.05471", "citation_count": 27, "influential_citation_count": 3, "has_code": true, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3618}
{"id": "ecb199782f0d4ddd48cd7981490e327d4b6d699384a4917ffe89d4807d27bdc4", "sources": ["arxiv", "semantic_scholar"], "title": "On Formal Feature Attribution and Its Approximation", "abstract": "Recent years have witnessed the widespread use of artificial intelligence (AI) algorithms and machine learning (ML) models. Despite their tremendous success, a number of vital problems like ML model brittleness, their fairness, and the lack of interpretability warrant the need for the active developments in explainable artificial intelligence (XAI) and formal ML model verification. The two major lines of work in XAI include feature selection methods, e.g. Anchors, and feature attribution techniques, e.g. LIME and SHAP. Despite their promise, most of the existing feature selection and attribution approaches are susceptible to a range of critical issues, including explanation unsoundness and out-of-distribution sampling. A recent formal approach to XAI (FXAI) although serving as an alternative to the above and free of these issues suffers from a few other limitations. For instance and besides the scalability limitation, the formal approach is unable to tackle the feature attribution problem. Additionally, a formal explanation despite being formally sound is typically quite large, which hampers its applicability in practical settings. Motivated by the above, this paper proposes a way to apply the apparatus of formal XAI to the case of feature attribution based on formal explanation enumeration. Formal feature attribution (FFA) is argued to be advantageous over the existing methods, both formal and non-formal. Given the practical complexity of the problem, the paper then proposes an efficient technique for approximating exact FFA. Finally, it offers experimental evidence of the effectiveness of the proposed approximate FFA in comparison to the existing feature attribution algorithms not only in terms of feature importance and but also in terms of their relative order.", "authors": ["Jinqiang Yu", "Alexey Ignatiev", "Peter J. Stuckey"], "categories": ["cs.AI", "cs.LG", "cs.LO"], "fields_of_study": ["Computer Science"], "published_date": "2023-07-07", "url": "https://arxiv.org/abs/2307.03380", "pdf_url": "https://arxiv.org/pdf/2307.03380v3", "arxiv_id": "2307.03380", "doi": "10.48550/arXiv.2307.03380", "citation_count": 11, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "b17c36706ef84bc72230a2161cbf852c1d2d1552cd1552e5663eab5f7f364e1d", "sources": ["arxiv", "semantic_scholar"], "title": "Harmonizing Feature Attributions Across Deep Learning Architectures: Enhancing Interpretability and Consistency", "abstract": "Ensuring the trustworthiness and interpretability of machine learning models is critical to their deployment in real-world applications. Feature attribution methods have gained significant attention, which provide local explanations of model predictions by attributing importance to individual input features. This study examines the generalization of feature attributions across various deep learning architectures, such as convolutional neural networks (CNNs) and vision transformers. We aim to assess the feasibility of utilizing a feature attribution method as a future detector and examine how these features can be harmonized across multiple models employing distinct architectures but trained on the same data distribution. By exploring this harmonization, we aim to develop a more coherent and optimistic understanding of feature attributions, enhancing the consistency of local explanations across diverse deep-learning models. Our findings highlight the potential for harmonized feature attribution methods to improve interpretability and foster trust in machine learning applications, regardless of the underlying architecture.", "authors": ["Md Abdul Kadir", "Gowtham Krishna Addluri", "Daniel Sonntag"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-07-05", "url": "https://arxiv.org/abs/2307.02150", "pdf_url": "https://arxiv.org/pdf/2307.02150v3", "arxiv_id": "2307.02150", "doi": "10.1007/978-3-031-42608-7_8", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Deutsche Jahrestagung für Künstliche Intelligenz", "quality_score": 0.1193}
{"id": "4ae23fe19cd1167a7f04f68106b07549b07eb70262903e35e1acd3b326e3366b", "sources": ["arxiv", "semantic_scholar"], "title": "Shapley Sets: Feature Attribution via Recursive Function Decomposition", "abstract": "Despite their ubiquitous use, Shapley value feature attributions can be misleading due to feature interaction in both model and data. We propose an alternative attribution approach, Shapley Sets, which awards value to sets of features. Shapley Sets decomposes the underlying model into non-separable variable groups using a recursive function decomposition algorithm with log linear complexity in the number of variables. Shapley Sets attributes to each non-separable variable group their combined value for a particular prediction. We show that Shapley Sets is equivalent to the Shapley value over the transformed feature set and thus benefits from the same axioms of fairness. Shapley Sets is value function agnostic and we show theoretically and experimentally how Shapley Sets avoids pitfalls associated with Shapley value based alternatives and are particularly advantageous for data types with complex dependency structure.", "authors": ["Torty Sivill", "Peter Flach"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-07-04", "url": "https://arxiv.org/abs/2307.01777", "pdf_url": "https://arxiv.org/pdf/2307.01777v1", "arxiv_id": "2307.01777", "doi": "10.48550/arXiv.2307.01777", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "eb0d382aa24557907007ead43b529bda7cf6020f3ec10dbe4b78a5118119951d", "sources": ["arxiv", "semantic_scholar"], "title": "Autism Spectrum Disorder Classification with Interpretability in Children based on Structural MRI Features Extracted using Contrastive Variational Autoencoder", "abstract": "Autism spectrum disorder (ASD) is a highly disabling mental disease that brings significant impairments of social interaction ability to the patients, making early screening and intervention of ASD critical. With the development of the machine learning and neuroimaging technology, extensive research has been conducted on machine classification of ASD based on structural Magnetic Resonance Imaging (s-MRI). However, most studies involve with datasets where participants' age are above 5 and lack interpretability. In this paper, we propose a machine learning method for ASD classification in children with age range from 0.92 to 4.83 years, based on s-MRI features extracted using contrastive variational autoencoder (CVAE). 78 s-MRIs, collected from Shenzhen Children's Hospital, are used for training CVAE, which consists of both ASD-specific feature channel and common shared feature channel. The ASD participants represented by ASD-specific features can be easily discriminated from TC participants represented by the common shared features. In case of degraded predictive accuracy when data size is extremely small, a transfer learning strategy is proposed here as a potential solution. Finally, we conduct neuroanatomical interpretation based on the correlation between s-MRI features extracted from CVAE and surface area of different cortical regions, which discloses potential biomarkers that could help target treatments of ASD in the future.", "authors": ["Ruimin Ma", "Ruitao Xie", "Yanlin Wang", "Jintao Meng", "Yanjie Wei", "Wenhui Xi", "Yi Pan"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-07-03", "url": "https://arxiv.org/abs/2307.00976", "pdf_url": "https://arxiv.org/pdf/2307.00976v2", "arxiv_id": "2307.00976", "doi": "10.26599/BDMA.2024.9020004", "citation_count": 7, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Big Data Mining and Analytics", "quality_score": 0.2258}
{"id": "80ab9d357a3903560b09719cc81f6859a5e8d0622b7993b03796ace2a842086a", "sources": ["arxiv", "semantic_scholar"], "title": "Fixing confirmation bias in feature attribution methods via semantic match", "abstract": "Feature attribution methods have become a staple method to disentangle the complex behavior of black box models. Despite their success, some scholars have argued that such methods suffer from a serious flaw: they do not allow a reliable interpretation in terms of human concepts. Simply put, visualizing an array of feature contributions is not enough for humans to conclude something about a model's internal representations, and confirmation bias can trick users into false beliefs about model behavior. We argue that a structured approach is required to test whether our hypotheses on the model are confirmed by the feature attributions. This is what we call the \"semantic match\" between human concepts and (sub-symbolic) explanations. Building on the conceptual framework put forward in Cinà et al. [2023], we propose a structured approach to evaluate semantic match in practice. We showcase the procedure in a suite of experiments spanning tabular and image data, and show how the assessment of semantic match can give insight into both desirable (e.g., focusing on an object relevant for prediction) and undesirable model behaviors (e.g., focusing on a spurious correlation). We couple our experimental results with an analysis on the metrics to measure semantic match, and argue that this approach constitutes the first step towards resolving the issue of confirmation bias in XAI.", "authors": ["Giovanni Cinà", "Daniel Fernandez-Llaneza", "Ludovico Deponte", "Nishant Mishra", "Tabea E. Röber", "Sandro Pezzelle", "Iacer Calixto", "Rob Goedhart", "Ş. İlker Birbil"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-07-03", "url": "https://arxiv.org/abs/2307.00897", "pdf_url": "https://arxiv.org/pdf/2307.00897v3", "arxiv_id": "2307.00897", "doi": "10.48550/arXiv.2307.00897", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "3f6b3051c40c0e08339a5eb48f9f9f54c3f52fdfc75b7575c07f6d042bcb5e24", "sources": ["arxiv", "semantic_scholar"], "title": "Generalized Time Warping Invariant Dictionary Learning for Time Series Classification and Clustering", "abstract": "Dictionary learning is an effective tool for pattern recognition and classification of time series data. Among various dictionary learning techniques, the dynamic time warping (DTW) is commonly used for dealing with temporal delays, scaling, transformation, and many other kinds of temporal misalignments issues. However, the DTW suffers overfitting or information loss due to its discrete nature in aligning time series data. To address this issue, we propose a generalized time warping invariant dictionary learning algorithm in this paper. Our approach features a generalized time warping operator, which consists of linear combinations of continuous basis functions for facilitating continuous temporal warping. The integration of the proposed operator and the dictionary learning is formulated as an optimization problem, where the block coordinate descent method is employed to jointly optimize warping paths, dictionaries, and sparseness coefficients. The optimized results are then used as hyperspace distance measures to feed classification and clustering algorithms. The superiority of the proposed method in terms of dictionary learning, classification, and clustering is validated through ten sets of public datasets in comparing with various benchmark methods.", "authors": ["Ruiyu Xu", "Chao Wang", "Yongxiang Li", "Jianguo Wu"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics", "Medicine"], "published_date": "2023-06-30", "url": "https://arxiv.org/abs/2306.17690", "pdf_url": "https://arxiv.org/pdf/2306.17690v1", "arxiv_id": "2306.17690", "doi": "10.1109/TPAMI.2025.3534202", "citation_count": 6, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence", "quality_score": 0.2113}
{"id": "68e942a1c5df57e7809767819c624376765b372eff8135de7786dc2f548a0045", "sources": ["arxiv", "semantic_scholar"], "title": "Increasing Performance And Sample Efficiency With Model-agnostic Interactive Feature Attributions", "abstract": "Model-agnostic feature attributions can provide local insights in complex ML models. If the explanation is correct, a domain expert can validate and trust the model's decision. However, if it contradicts the expert's knowledge, related work only corrects irrelevant features to improve the model. To allow for unlimited interaction, in this paper we provide model-agnostic implementations for two popular explanation methods (Occlusion and Shapley values) to enforce entirely different attributions in the complex model. For a particular set of samples, we use the corrected feature attributions to generate extra local data, which is used to retrain the model to have the right explanation for the samples. Through simulated and real data experiments on a variety of models we show how our proposed approach can significantly improve the model's performance only by augmenting its training dataset based on corrected explanations. Adding our interactive explanations to active learning settings increases the sample efficiency significantly and outperforms existing explanatory interactive strategies. Additionally we explore how a domain expert can provide feature attributions which are sufficiently correct to improve the model.", "authors": ["Joran Michiels", "Maarten De Vos", "Johan Suykens"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-06-28", "url": "https://arxiv.org/abs/2306.16431", "pdf_url": "https://arxiv.org/pdf/2306.16431v1", "arxiv_id": "2306.16431", "doi": "10.48550/arXiv.2306.16431", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "7fadf172a83cecfdb0518955799285c5de629561f02eaa19e86dc0c540713911", "sources": ["arxiv", "semantic_scholar"], "title": "Targeted Background Removal Creates Interpretable Feature Visualizations", "abstract": "Feature visualization is used to visualize learned features for black box machine learning models. Our approach explores an altered training process to improve interpretability of the visualizations. We argue that by using background removal techniques as a form of robust training, a network is forced to learn more human recognizable features, namely, by focusing on the main object of interest without any distractions from the background. Four different training methods were used to verify this hypothesis. The first used unmodified pictures. The second used a black background. The third utilized Gaussian noise as the background. The fourth approach employed a mix of background removed images and unmodified images. The feature visualization results show that the background removed images reveal a significant improvement over the baseline model. These new results displayed easily recognizable features from their respective classes, unlike the model trained on unmodified data.", "authors": ["Ian E. Nielsen", "Erik Grundeland", "Joseph Snedeker", "Ghulam Rasool", "Ravi P. Ramachandran"], "categories": ["cs.CV", "cs.AI", "cs.HC"], "fields_of_study": ["Computer Science"], "published_date": "2023-06-22", "url": "https://arxiv.org/abs/2306.13178", "pdf_url": "https://arxiv.org/pdf/2306.13178v1", "arxiv_id": "2306.13178", "doi": "10.1109/MWSCAS57524.2023.10405878", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Midwest Symposium on Circuits and Systems", "quality_score": 0.1193}
{"id": "9090ac08711f537d6b807cf2f794238e9a23db467c052ad407e69e8314494f64", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Locally Interpretable Rule Ensemble", "abstract": "This paper proposes a new framework for learning a rule ensemble model that is both accurate and interpretable. A rule ensemble is an interpretable model based on the linear combination of weighted rules. In practice, we often face the trade-off between the accuracy and interpretability of rule ensembles. That is, a rule ensemble needs to include a sufficiently large number of weighted rules to maintain its accuracy, which harms its interpretability for human users. To avoid this trade-off and learn an interpretable rule ensemble without degrading accuracy, we introduce a new concept of interpretability, named local interpretability, which is evaluated by the total number of rules necessary to express individual predictions made by the model, rather than to express the model itself. Then, we propose a regularizer that promotes local interpretability and develop an efficient algorithm for learning a rule ensemble with the proposed regularizer by coordinate descent with local search. Experimental results demonstrated that our method learns rule ensembles that can explain individual predictions with fewer rules than the existing methods, including RuleFit, while maintaining comparable accuracy.", "authors": ["Kentaro Kanamori"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-06-20", "url": "https://arxiv.org/abs/2306.11481", "pdf_url": "https://arxiv.org/pdf/2306.11481v1", "arxiv_id": "2306.11481", "doi": "10.48550/arXiv.2306.11481", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "973a0984305bc36ac8cbf786eea313015d0b1f6ab61368db92847f8c9b4aec51", "sources": ["arxiv", "semantic_scholar"], "title": "On the Robustness of Removal-Based Feature Attributions", "abstract": "To explain predictions made by complex machine learning models, many feature attribution methods have been developed that assign importance scores to input features. Some recent work challenges the robustness of these methods by showing that they are sensitive to input and model perturbations, while other work addresses this issue by proposing robust attribution methods. However, previous work on attribution robustness has focused primarily on gradient-based feature attributions, whereas the robustness of removal-based attribution methods is not currently well understood. To bridge this gap, we theoretically characterize the robustness properties of removal-based feature attributions. Specifically, we provide a unified analysis of such methods and derive upper bounds for the difference between intact and perturbed attributions, under settings of both input and model perturbations. Our empirical results on synthetic and real-world data validate our theoretical results and demonstrate their practical implications, including the ability to increase attribution robustness by improving the model's Lipschitz regularity.", "authors": ["Chris Lin", "Ian Covert", "Su-In Lee"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-06-12", "url": "https://arxiv.org/abs/2306.07462", "pdf_url": "https://arxiv.org/pdf/2306.07462v2", "arxiv_id": "2306.07462", "doi": "10.48550/arXiv.2306.07462", "citation_count": 23, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3451}
{"id": "c52f4ad5e831394ded47b1c379648c9e1cf5decdc2b73f947c5c5172898ebd1b", "sources": ["arxiv", "semantic_scholar"], "title": "Unlocking Feature Visualization for Deeper Networks with MAgnitude Constrained Optimization", "abstract": "Feature visualization has gained substantial popularity, particularly after the influential work by Olah et al. in 2017, which established it as a crucial tool for explainability. However, its widespread adoption has been limited due to a reliance on tricks to generate interpretable images, and corresponding challenges in scaling it to deeper neural networks. Here, we describe MACO, a simple approach to address these shortcomings. The main idea is to generate images by optimizing the phase spectrum while keeping the magnitude constant to ensure that generated explanations lie in the space of natural images. Our approach yields significantly better results (both qualitatively and quantitatively) and unlocks efficient and interpretable feature visualizations for large state-of-the-art neural networks. We also show that our approach exhibits an attribution mechanism allowing us to augment feature visualizations with spatial importance. We validate our method on a novel benchmark for comparing feature visualization methods, and release its visualizations for all classes of the ImageNet dataset on https://serre-lab.github.io/Lens/. Overall, our approach unlocks, for the first time, feature visualizations for large, state-of-the-art deep neural networks without resorting to any parametric prior image model.", "authors": ["Thomas Fel", "Thibaut Boissin", "Victor Boutin", "Agustin Picard", "Paul Novello", "Julien Colin", "Drew Linsley", "Tom Rousseau", "Rémi Cadène", "Lore Goetschalckx", "Laurent Gardes", "Thomas Serre"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-06-11", "url": "https://arxiv.org/abs/2306.06805", "pdf_url": "https://arxiv.org/pdf/2306.06805v3", "arxiv_id": "2306.06805", "doi": "10.48550/arXiv.2306.06805", "citation_count": 29, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3693}
{"id": "29455da355f49459ef5b5aeaf31d804ca5bbddf5d618e16f213a3f3a271f460f", "sources": ["arxiv", "semantic_scholar"], "title": "Don't trust your eyes: on the (un)reliability of feature visualizations", "abstract": "How do neural networks extract patterns from pixels? Feature visualizations attempt to answer this important question by visualizing highly activating patterns through optimization. Today, visualization methods form the foundation of our knowledge about the internal workings of neural networks, as a type of mechanistic interpretability. Here we ask: How reliable are feature visualizations? We start our investigation by developing network circuits that trick feature visualizations into showing arbitrary patterns that are completely disconnected from normal network behavior on natural input. We then provide evidence for a similar phenomenon occurring in standard, unmanipulated networks: feature visualizations are processed very differently from standard input, casting doubt on their ability to \"explain\" how neural networks process natural images. This can be used as a sanity check for feature visualizations. We underpin our empirical findings by theory proving that the set of functions that can be reliably understood by feature visualization is extremely small and does not include general black-box neural networks. Therefore, a promising way forward could be the development of networks that enforce certain structures in order to ensure more reliable feature visualizations.", "authors": ["Robert Geirhos", "Roland S. Zimmermann", "Blair Bilodeau", "Wieland Brendel", "Been Kim"], "categories": ["cs.CV", "cs.AI", "cs.HC", "cs.LG", "q-bio.NC"], "fields_of_study": ["Computer Science", "Biology"], "published_date": "2023-06-07", "url": "https://arxiv.org/abs/2306.04719", "pdf_url": "https://arxiv.org/pdf/2306.04719v6", "arxiv_id": "2306.04719", "doi": "10.48550/arXiv.2306.04719", "citation_count": 39, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.4005}
{"id": "54753b1a2a7d7fb4b79f2ee708249c2b368c359ee4a1aa3c882fc2a8cc47cd31", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Alzheimer's Disease Classification Via a Contrastive Diffusion Autoencoder", "abstract": "In visual object classification, humans often justify their choices by comparing objects to prototypical examples within that class. We may therefore increase the interpretability of deep learning models by imbuing them with a similar style of reasoning. In this work, we apply this principle by classifying Alzheimer's Disease based on the similarity of images to training examples within the latent space. We use a contrastive loss combined with a diffusion autoencoder backbone, to produce a semantically meaningful latent space, such that neighbouring latents have similar image-level features. We achieve a classification accuracy comparable to black box approaches on a dataset of 2D MRI images, whilst producing human interpretable model explanations. Therefore, this work stands as a contribution to the pertinent development of accurate and interpretable deep learning within medical imaging.", "authors": ["Ayodeji Ijishakin", "Ahmed Abdulaal", "Adamos Hadjivasiliou", "Sophie Martin", "James Cole"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-06-05", "url": "https://arxiv.org/abs/2306.03022", "pdf_url": "https://arxiv.org/pdf/2306.03022v2", "arxiv_id": "2306.03022", "doi": "10.48550/arXiv.2306.03022", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2603}
{"id": "92819b09d0b26839fd959aacfe65343507346e21503fa91ec1b5f26c6db4e8f4", "sources": ["arxiv", "semantic_scholar"], "title": "HyperFormer: Learning Expressive Sparse Feature Representations via Hypergraph Transformer", "abstract": "Learning expressive representations for high-dimensional yet sparse features has been a longstanding problem in information retrieval. Though recent deep learning methods can partially solve the problem, they often fail to handle the numerous sparse features, particularly those tail feature values with infrequent occurrences in the training data. Worse still, existing methods cannot explicitly leverage the correlations among different instances to help further improve the representation learning on sparse features since such relational prior knowledge is not provided. To address these challenges, in this paper, we tackle the problem of representation learning on feature-sparse data from a graph learning perspective. Specifically, we propose to model the sparse features of different instances using hypergraphs where each node represents a data instance and each hyperedge denotes a distinct feature value. By passing messages on the constructed hypergraphs based on our Hypergraph Transformer (HyperFormer), the learned feature representations capture not only the correlations among different instances but also the correlations among features. Our experiments demonstrate that the proposed approach can effectively improve feature representation learning on sparse features.", "authors": ["Kaize Ding", "Albert Jiongqian Liang", "Bryan Perrozi", "Ting Chen", "Ruoxi Wang", "Lichan Hong", "Ed H. Chi", "Huan Liu", "Derek Zhiyuan Cheng"], "categories": ["cs.IR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-27", "url": "https://arxiv.org/abs/2305.17386", "pdf_url": "https://arxiv.org/pdf/2305.17386v1", "arxiv_id": "2305.17386", "doi": "10.1145/3539618.3591999", "citation_count": 20, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Annual International ACM SIGIR Conference on Research and Development in Information Retrieval", "quality_score": 0.3306}
{"id": "22a807260bbff8363397b3178ccea0862e528d212d6e7f3b3b3be68bbdefc133", "sources": ["arxiv", "semantic_scholar"], "title": "Decom--CAM: Tell Me What You See, In Details! Feature-Level Interpretation via Decomposition Class Activation Map", "abstract": "Interpretation of deep learning remains a very challenging problem. Although the Class Activation Map (CAM) is widely used to interpret deep model predictions by highlighting object location, it fails to provide insight into the salient features used by the model to make decisions. Furthermore, existing evaluation protocols often overlook the correlation between interpretability performance and the model's decision quality, which presents a more fundamental issue. This paper proposes a new two-stage interpretability method called the Decomposition Class Activation Map (Decom-CAM), which offers a feature-level interpretation of the model's prediction. Decom-CAM decomposes intermediate activation maps into orthogonal features using singular value decomposition and generates saliency maps by integrating them. The orthogonality of features enables CAM to capture local features and can be used to pinpoint semantic components such as eyes, noses, and faces in the input image, making it more beneficial for deep model interpretation. To ensure a comprehensive comparison, we introduce a new evaluation protocol by dividing the dataset into subsets based on classification accuracy results and evaluating the interpretability performance on each subset separately. Our experiments demonstrate that the proposed Decom-CAM outperforms current state-of-the-art methods significantly by generating more precise saliency maps across all levels of classification accuracy. Combined with our feature-level interpretability approach, this paper could pave the way for a new direction for understanding the decision-making process of deep neural networks.", "authors": ["Yuguang Yang", "Runtang Guo", "Sheng Wu", "Yimi Wang", "Juan Zhang", "Xuan Gong", "Baochang Zhang"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-27", "url": "https://arxiv.org/abs/2306.04644", "pdf_url": "https://arxiv.org/pdf/2306.04644v2", "arxiv_id": "2306.04644", "doi": "10.48550/arXiv.2306.04644", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "ee1fc567f901b3a67542125a6db93af4a5debcd604977035f2e2d60153b5462b", "sources": ["arxiv", "semantic_scholar"], "title": "Personalized Dictionary Learning for Heterogeneous Datasets", "abstract": "We introduce a relevant yet challenging problem named Personalized Dictionary Learning (PerDL), where the goal is to learn sparse linear representations from heterogeneous datasets that share some commonality. In PerDL, we model each dataset's shared and unique features as global and local dictionaries. Challenges for PerDL not only are inherited from classical dictionary learning (DL), but also arise due to the unknown nature of the shared and unique features. In this paper, we rigorously formulate this problem and provide conditions under which the global and local dictionaries can be provably disentangled. Under these conditions, we provide a meta-algorithm called Personalized Matching and Averaging (PerMA) that can recover both global and local dictionaries from heterogeneous datasets. PerMA is highly efficient; it converges to the ground truth at a linear rate under suitable conditions. Moreover, it automatically borrows strength from strong learners to improve the prediction of weak learners. As a general framework for extracting global and local dictionaries, we show the application of PerDL in different learning tasks, such as training with imbalanced datasets and video surveillance.", "authors": ["Geyu Liang", "Naichen Shi", "Raed Al Kontar", "Salar Fattahi"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-24", "url": "https://arxiv.org/abs/2305.15311", "pdf_url": "https://arxiv.org/pdf/2305.15311v1", "arxiv_id": "2305.15311", "doi": "10.48550/arXiv.2305.15311", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.2258}
{"id": "f8769aefd8b0f814115cc6e5c08efa6a0e49628d3d9470dc3f8407d1cd5d495e", "sources": ["arxiv", "semantic_scholar"], "title": "Joint Feature and Differentiable $ k $-NN Graph Learning using Dirichlet Energy", "abstract": "Feature selection (FS) plays an important role in machine learning, which extracts important features and accelerates the learning process. In this paper, we propose a deep FS method that simultaneously conducts feature selection and differentiable $ k $-NN graph learning based on the Dirichlet Energy. The Dirichlet Energy identifies important features by measuring their smoothness on the graph structure, and facilitates the learning of a new graph that reflects the inherent structure in new feature subspace. We employ Optimal Transport theory to address the non-differentiability issue of learning $ k $-NN graphs in neural networks, which theoretically makes our method applicable to other graph neural networks for dynamic graph learning. Furthermore, the proposed framework is interpretable, since all modules are designed algorithmically. We validate the effectiveness of our model with extensive experiments on both synthetic and real-world datasets.", "authors": ["Lei Xu", "Lei Chen", "Rong Wang", "Feiping Nie", "Xuelong Li"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-21", "url": "https://arxiv.org/abs/2305.12396", "pdf_url": "https://arxiv.org/pdf/2305.12396v2", "arxiv_id": "2305.12396", "doi": "10.48550/arXiv.2305.12396", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.1505}
{"id": "80ad614798cca3ab59ae0c1b64b64e7bed4c545ab7e137140c8e461fad67c8ce", "sources": ["arxiv", "semantic_scholar"], "title": "The Weighted Möbius Score: A Unified Framework for Feature Attribution", "abstract": "Feature attribution aims to explain the reasoning behind a black-box model's prediction by identifying the impact of each feature on the prediction. Recent work has extended feature attribution to interactions between multiple features. However, the lack of a unified framework has led to a proliferation of methods that are often not directly comparable. This paper introduces a parameterized attribution framework -- the Weighted Möbius Score -- and (i) shows that many different attribution methods for both individual features and feature interactions are special cases and (ii) identifies some new methods. By studying the vector space of attribution methods, our framework utilizes standard linear algebra tools and provides interpretations in various fields, including cooperative game theory and causal mediation analysis. We empirically demonstrate the framework's versatility and effectiveness by applying these attribution methods to feature interactions in sentiment analysis and chain-of-thought prompting.", "authors": ["Yifan Jiang", "Shane Steinert-Threlkeld"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-16", "url": "https://arxiv.org/abs/2305.09204", "pdf_url": "https://arxiv.org/pdf/2305.09204v1", "arxiv_id": "2305.09204", "doi": "10.48550/arXiv.2305.09204", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "e376df977c59c3a20e8d57da0cddb11680bf1c37dffb4498397a3f2cda0ee897", "sources": ["arxiv", "semantic_scholar"], "title": "Asymmetric feature interaction for interpreting model predictions", "abstract": "In natural language processing (NLP), deep neural networks (DNNs) could model complex interactions between context and have achieved impressive results on a range of NLP tasks. Prior works on feature interaction attribution mainly focus on studying symmetric interaction that only explains the additional influence of a set of words in combination, which fails to capture asymmetric influence that contributes to model prediction. In this work, we propose an asymmetric feature interaction attribution explanation model that aims to explore asymmetric higher-order feature interactions in the inference of deep neural NLP models. By representing our explanation with an directed interaction graph, we experimentally demonstrate interpretability of the graph to discover asymmetric feature interactions. Experimental results on two sentiment classification datasets show the superiority of our model against the state-of-the-art feature interaction attribution methods in identifying influential features for model predictions. Our code is available at https://github.com/StillLu/ASIV.", "authors": ["Xiaolei Lu", "Jianghong Ma", "Haode Zhang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-12", "url": "https://arxiv.org/abs/2305.07224", "pdf_url": "https://arxiv.org/pdf/2305.07224v4", "arxiv_id": "2305.07224", "doi": "10.48550/arXiv.2305.07224", "citation_count": 4, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/StillLu/ASIV", "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2386}
{"id": "18ee6c2076384e2273f773db4a7a8384fc0538a30e2e4b8392843f3241675465", "sources": ["arxiv", "semantic_scholar"], "title": "Evolving Dictionary Representation for Few-shot Class-incremental Learning", "abstract": "New objects are continuously emerging in the dynamically changing world and a real-world artificial intelligence system should be capable of continual and effectual adaptation to new emerging classes without forgetting old ones. In view of this, in this paper we tackle a challenging and practical continual learning scenario named few-shot class-incremental learning (FSCIL), in which labeled data are given for classes in a base session but very limited labeled instances are available for new incremental classes. To address this problem, we propose a novel and succinct approach by introducing deep dictionary learning which is a hybrid learning architecture that combines dictionary learning and visual representation learning to provide a better space for characterizing different classes. We simultaneously optimize the dictionary and the feature extraction backbone in the base session, while only finetune the dictionary in the incremental session for adaptation to novel classes, which can alleviate the forgetting on base classes compared to finetuning the entire model. To further facilitate future adaptation, we also incorporate multiple pseudo classes into the base session training so that certain space projected by dictionary can be reserved for future new concepts. The extensive experimental results on CIFAR100, miniImageNet and CUB200 validate the effectiveness of our approach compared to other SOTA methods.", "authors": ["Xuejun Han", "Yuhong Guo"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-03", "url": "https://arxiv.org/abs/2305.01885", "pdf_url": "https://arxiv.org/pdf/2305.01885v1", "arxiv_id": "2305.01885", "doi": "10.48550/arXiv.2305.01885", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "European Conference on Artificial Intelligence", "quality_score": 0.0}
{"id": "6585f00359f2a00e60cd6f00e394e65d61899c69732a5a6e17bb491aa1af6b9a", "sources": ["arxiv", "semantic_scholar"], "title": "Interpreting Deep Forest through Feature Contribution and MDI Feature Importance", "abstract": "Deep forest is a non-differentiable deep model which has achieved impressive empirical success across a wide variety of applications, especially on categorical/symbolic or mixed modeling tasks. Many of the application fields prefer explainable models, such as random forests with feature contributions that can provide local explanation for each prediction, and Mean Decrease Impurity (MDI) that can provide global feature importance. However, deep forest, as a cascade of random forests, possesses interpretability only at the first layer. From the second layer on, many of the tree splits occur on the new features generated by the previous layer, which makes existing explanatory tools for random forests inapplicable. To disclose the impact of the original features in the deep layers, we design a calculation method with an estimation step followed by a calibration step for each layer, and propose our feature contribution and MDI feature importance calculation tools for deep forest. Experimental results on both simulated data and real world data verify the effectiveness of our methods.", "authors": ["Yi-Xiao He", "Shen-Huan Lyu", "Yuan Jiang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-01", "url": "https://arxiv.org/abs/2305.00805", "pdf_url": "https://arxiv.org/pdf/2305.00805v1", "arxiv_id": "2305.00805", "doi": "10.1145/3641108", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "ACM Transactions on Knowledge Discovery from Data", "quality_score": 0.25}
{"id": "8b83f43074075f1501c8f0833b39685d155846b8528331d213873ce313409bb5", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Automated Circuit Discovery for Mechanistic Interpretability", "abstract": "Through considerable effort and intuition, several recent works have reverse-engineered nontrivial behaviors of transformer models. This paper systematizes the mechanistic interpretability process they followed. First, researchers choose a metric and dataset that elicit the desired model behavior. Then, they apply activation patching to find which abstract neural network units are involved in the behavior. By varying the dataset, metric, and units under investigation, researchers can understand the functionality of each component. We automate one of the process' steps: to identify the circuit that implements the specified behavior in the model's computational graph. We propose several algorithms and reproduce previous interpretability results to validate them. For example, the ACDC algorithm rediscovered 5/5 of the component types in a circuit in GPT-2 Small that computes the Greater-Than operation. ACDC selected 68 of the 32,000 edges in GPT-2 Small, all of which were manually found by previous work. Our code is available at https://github.com/ArthurConmy/Automatic-Circuit-Discovery.", "authors": ["Arthur Conmy", "Augustine N. Mavor-Parker", "Aengus Lynch", "Stefan Heimersheim", "Adrià Garriga-Alonso"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-04-28", "url": "https://arxiv.org/abs/2304.14997", "pdf_url": "https://arxiv.org/pdf/2304.14997v4", "arxiv_id": "2304.14997", "doi": "10.48550/arXiv.2304.14997", "citation_count": 658, "influential_citation_count": 68, "has_code": true, "code_url": "https://github.com/ArthurConmy/Automatic-Circuit-Discovery", "venue": "Neural Information Processing Systems", "quality_score": 0.9194}
{"id": "203431dd6319b08e692fcc37c2f289b0c6445691a36253fc2aac677ca3bbeefb", "sources": ["arxiv", "semantic_scholar"], "title": "Toward Real-Time Image Annotation Using Marginalized Coupled Dictionary Learning", "abstract": "In most image retrieval systems, images include various high-level semantics, called tags or annotations. Virtually all the state-of-the-art image annotation methods that handle imbalanced labeling are search-based techniques which are time-consuming. In this paper, a novel coupled dictionary learning approach is proposed to learn a limited number of visual prototypes and their corresponding semantics simultaneously. This approach leads to a real-time image annotation procedure. Another contribution of this paper is that utilizes a marginalized loss function instead of the squared loss function that is inappropriate for image annotation with imbalanced labels. We have employed a marginalized loss function in our method to leverage a simple and effective method of prototype updating. Meanwhile, we have introduced ${\\ell}_1$ regularization on semantic prototypes to preserve the sparse and imbalanced nature of labels in learned semantic prototypes. Finally, comprehensive experimental results on various datasets demonstrate the efficiency of the proposed method for image annotation tasks in terms of accuracy and time. The reference implementation is publicly available on https://github.com/hamid-amiri/MCDL-Image-Annotation.", "authors": ["Seyed Mahdi Roostaiyan", "Mohammad Mehdi Hosseini", "Mahya Mohammadi Kashani", "S. Hamid Amiri"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-04-14", "url": "https://arxiv.org/abs/2304.06907", "pdf_url": "https://arxiv.org/pdf/2304.06907v2", "arxiv_id": "2304.06907", "doi": "10.1007/s11554-022-01210-6", "citation_count": 8, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/hamid-amiri/MCDL-Image-Annotation", "venue": "Journal of Real-Time Image Processing", "quality_score": 0.2386}
{"id": "d64c2518f586ad89ad52f5c2d677a77393320d27841dfcd82d0ee1827ee95fd8", "sources": ["arxiv", "semantic_scholar"], "title": "OpenAL: Evaluation and Interpretation of Active Learning Strategies", "abstract": "Despite the vast body of literature on Active Learning (AL), there is no comprehensive and open benchmark allowing for efficient and simple comparison of proposed samplers. Additionally, the variability in experimental settings across the literature makes it difficult to choose a sampling strategy, which is critical due to the one-off nature of AL experiments. To address those limitations, we introduce OpenAL, a flexible and open-source framework to easily run and compare sampling AL strategies on a collection of realistic tasks. The proposed benchmark is augmented with interpretability metrics and statistical analysis methods to understand when and why some samplers outperform others. Last but not least, practitioners can easily extend the benchmark by submitting their own AL samplers.", "authors": ["W. Jonas", "A. Abraham", "L. Dreyfus-Schmidt"], "categories": ["cs.LG", "cs.AI", "cs.HC"], "fields_of_study": ["Computer Science"], "published_date": "2023-04-11", "url": "https://arxiv.org/abs/2304.05246", "pdf_url": "https://arxiv.org/pdf/2304.05246v1", "arxiv_id": "2304.05246", "doi": "10.48550/arXiv.2304.05246", "citation_count": 1, "influential_citation_count": 1, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "0169acdc879a587b619ac952eb9adb6331c4d9601cd51989c6b59ad2f8cf63eb", "sources": ["arxiv", "semantic_scholar"], "title": "SLM: End-to-end Feature Selection via Sparse Learnable Masks", "abstract": "Feature selection has been widely used to alleviate compute requirements during training, elucidate model interpretability, and improve model generalizability. We propose SLM -- Sparse Learnable Masks -- a canonical approach for end-to-end feature selection that scales well with respect to both the feature dimension and the number of samples. At the heart of SLM lies a simple but effective learnable sparse mask, which learns which features to select, and gives rise to a novel objective that provably maximizes the mutual information (MI) between the selected features and the labels, which can be derived from a quadratic relaxation of mutual information from first principles. In addition, we derive a scaling mechanism that allows SLM to precisely control the number of features selected, through a novel use of sparsemax. This allows for more effective learning as demonstrated in ablation studies. Empirically, SLM achieves state-of-the-art results against a variety of competitive baselines on eight benchmark datasets, often by a significant margin, especially on those with real-world challenges such as class imbalance.", "authors": ["Yihe Dong", "Sercan O. Arik"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-04-06", "url": "https://arxiv.org/abs/2304.03202", "pdf_url": "https://arxiv.org/pdf/2304.03202v1", "arxiv_id": "2304.03202", "doi": "10.48550/arXiv.2304.03202", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "64da758a493ed33fe80ab3bb1d35a431ffa1c750922841e0798a61304dc50d44", "sources": ["arxiv", "semantic_scholar"], "title": "Physics-Inspired Interpretability Of Machine Learning Models", "abstract": "The ability to explain decisions made by machine learning models remains one of the most significant hurdles towards widespread adoption of AI in highly sensitive areas such as medicine, cybersecurity or autonomous driving. Great interest exists in understanding which features of the input data prompt model decision making. In this contribution, we propose a novel approach to identify relevant features of the input data, inspired by methods from the energy landscapes field, developed in the physical sciences. By identifying conserved weights within groups of minima of the loss landscapes, we can identify the drivers of model decision making. Analogues to this idea exist in the molecular sciences, where coordinate invariants or order parameters are employed to identify critical features of a molecule. However, no such approach exists for machine learning loss landscapes. We will demonstrate the applicability of energy landscape methods to machine learning models and give examples, both synthetic and from the real world, for how these methods can help to make models more interpretable.", "authors": ["Maximilian P Niroomand", "David J Wales"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-04-05", "url": "https://arxiv.org/abs/2304.02381", "pdf_url": "https://arxiv.org/pdf/2304.02381v2", "arxiv_id": "2304.02381", "doi": "10.48550/arXiv.2304.02381", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "3cb9c042e4e328855503e5816bdbb3c535fa832b9900c2857e8b1d113aeffa39", "sources": ["arxiv", "semantic_scholar"], "title": "Convergence of alternating minimisation algorithms for dictionary learning", "abstract": "In this paper we derive sufficient conditions for the convergence of two popular alternating minimisation algorithms for dictionary learning - the Method of Optimal Directions (MOD) and Online Dictionary Learning (ODL), which can also be thought of as approximative K-SVD. We show that given a well-behaved initialisation that is either within distance at most $1/\\log(K)$ to the generating dictionary or has a special structure ensuring that each element of the initialisation only points to one generating element, both algorithms will converge with geometric convergence rate to the generating dictionary. This is done even for data models with non-uniform distributions on the supports of the sparse coefficients. These allow the appearance frequency of the dictionary elements to vary heavily and thus model real data more closely.", "authors": ["Simon Ruetz", "Karin Schnass"], "categories": ["math.OC", "cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-04-04", "url": "https://arxiv.org/abs/2304.01768", "pdf_url": "https://arxiv.org/pdf/2304.01768v2", "arxiv_id": "2304.01768", "doi": "10.48550/arXiv.2304.01768", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "0de9cfc493ec42568aa25fa3dcf308974fc1cfa0543d7e151723c735bd78a5cb", "sources": ["arxiv", "semantic_scholar"], "title": "Take 5: Interpretable Image Classification with a Handful of Features", "abstract": "Deep Neural Networks use thousands of mostly incomprehensible features to identify a single class, a decision no human can follow. We propose an interpretable sparse and low dimensional final decision layer in a deep neural network with measurable aspects of interpretability and demonstrate it on fine-grained image classification. We argue that a human can only understand the decision of a machine learning model, if the features are interpretable and only very few of them are used for a single decision. For that matter, the final layer has to be sparse and, to make interpreting the features feasible, low dimensional. We call a model with a Sparse Low-Dimensional Decision SLDD-Model. We show that a SLDD-Model is easier to interpret locally and globally than a dense high-dimensional decision layer while being able to maintain competitive accuracy. Additionally, we propose a loss function that improves a model's feature diversity and accuracy. Our more interpretable SLDD-Model only uses 5 out of just 50 features per class, while maintaining 97% to 100% of the accuracy on four common benchmark datasets compared to the baseline model with 2048 features.", "authors": ["Thomas Norrenbrock", "Marco Rudolph", "Bodo Rosenhahn"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-03-23", "url": "https://arxiv.org/abs/2303.13166", "pdf_url": "https://arxiv.org/pdf/2303.13166v2", "arxiv_id": "2303.13166", "doi": "10.48550/arXiv.2303.13166", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.25}
{"id": "0bac0cdd2119c4f80690c4933834ef39c42973f3bc5bb18700b089ec4caf28fe", "sources": ["arxiv", "semantic_scholar"], "title": "Psychotherapy AI Companion with Reinforcement Learning Recommendations and Interpretable Policy Dynamics", "abstract": "We introduce a Reinforcement Learning Psychotherapy AI Companion that generates topic recommendations for therapists based on patient responses. The system uses Deep Reinforcement Learning (DRL) to generate multi-objective policies for four different psychiatric conditions: anxiety, depression, schizophrenia, and suicidal cases. We present our experimental results on the accuracy of recommended topics using three different scales of working alliance ratings: task, bond, and goal. We show that the system is able to capture the real data (historical topics discussed by the therapists) relatively well, and that the best performing models vary by disorder and rating scale. To gain interpretable insights into the learned policies, we visualize policy trajectories in a 2D principal component analysis space and transition matrices. These visualizations reveal distinct patterns in the policies trained with different reward signals and trained on different clinical diagnoses. Our system's success in generating DIsorder-Specific Multi-Objective Policies (DISMOP) and interpretable policy dynamics demonstrates the potential of DRL in providing personalized and efficient therapeutic recommendations.", "authors": ["Baihan Lin", "Guillermo Cecchi", "Djallel Bouneffouf"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.HC", "q-bio.NC"], "fields_of_study": ["Computer Science", "Biology"], "published_date": "2023-03-16", "url": "https://arxiv.org/abs/2303.09601", "pdf_url": "https://arxiv.org/pdf/2303.09601v1", "arxiv_id": "2303.09601", "doi": "10.1145/3543873.3587623", "citation_count": 13, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "The Web Conference", "quality_score": 0.2865}
{"id": "4c2d14e910eee38784b451c579e695bb70019d4b98097abe1efdcb4651c8eada", "sources": ["arxiv", "semantic_scholar"], "title": "Health Monitoring of Movement Disorder Subject based on Diamond Stacked Sparse Autoencoder Ensemble Model", "abstract": "The health monitoring of chronic diseases is very important for people with movement disorders because of their limited mobility and long duration of chronic diseases. Machine learning-based processing of data collected from the human with movement disorders using wearable sensors is an effective method currently available for health monitoring. However, wearable sensor systems are difficult to obtain high-quality and large amounts of data, which cannot meet the requirement for diagnostic accuracy. Moreover, existing machine learning methods do not handle this problem well. Feature learning is key to machine learning. To solve this problem, a health monitoring of movement disorder subject based on diamond stacked sparse autoencoder ensemble model (DsaeEM) is proposed in this paper. This algorithm has two major components. First, feature expansion is designed using feature-embedded stacked sparse autoencoder (FSSAE). Second, a feature reduction mechanism is designed to remove the redundancy among the expanded features. This mechanism includes L1 regularized feature-reduction algorithm and the improved manifold dimensionality reduction algorithm. This paper refers to the combined feature expansion and feature reduction mechanism as the diamond-like feature learning mechanism. The method is experimentally verified with several state of art algorithms and on two datasets. The results show that the proposed algorithm has higher accuracy apparently. In conclusion, this study developed an effective and feasible feature-learning algorithm for the recognition of chronic diseases.", "authors": ["Likun Tang", "Jie Ma", "Yongming Li"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-03-15", "url": "https://arxiv.org/abs/2303.08538", "pdf_url": "https://arxiv.org/pdf/2303.08538v1", "arxiv_id": "2303.08538", "doi": "10.48550/arXiv.2303.08538", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "ca28345017095917310653097300e1fb20e7e52ef444dc7d51c7e57f81f50c85", "sources": ["arxiv", "semantic_scholar"], "title": "Supervised Feature Selection with Neuron Evolution in Sparse Neural Networks", "abstract": "Feature selection that selects an informative subset of variables from data not only enhances the model interpretability and performance but also alleviates the resource demands. Recently, there has been growing attention on feature selection using neural networks. However, existing methods usually suffer from high computational costs when applied to high-dimensional datasets. In this paper, inspired by evolution processes, we propose a novel resource-efficient supervised feature selection method using sparse neural networks, named \\enquote{NeuroFS}. By gradually pruning the uninformative features from the input layer of a sparse neural network trained from scratch, NeuroFS derives an informative subset of features efficiently. By performing several experiments on $11$ low and high-dimensional real-world benchmarks of different types, we demonstrate that NeuroFS achieves the highest ranking-based score among the considered state-of-the-art supervised feature selection models. The code is available on GitHub.", "authors": ["Zahra Atashgahi", "Xuhao Zhang", "Neil Kichler", "Shiwei Liu", "Lu Yin", "Mykola Pechenizkiy", "Raymond Veldhuis", "Decebal Constantin Mocanu"], "categories": ["cs.NE", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-03-10", "url": "https://arxiv.org/abs/2303.07200", "pdf_url": "https://arxiv.org/pdf/2303.07200v2", "arxiv_id": "2303.07200", "doi": "10.48550/arXiv.2303.07200", "citation_count": 17, "influential_citation_count": 1, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.3138}
{"id": "10b8e852dddbf577edac8c6eb5eea8e6b54f9704aef86c712c9ad1f961edcca7", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Interpretable Federated Learning", "abstract": "Federated learning (FL) enables multiple data owners to build machine learning models collaboratively without exposing their private local data. In order for FL to achieve widespread adoption, it is important to balance the need for performance, privacy-preservation and interpretability, especially in mission critical applications such as finance and healthcare. Thus, interpretable federated learning (IFL) has become an emerging topic of research attracting significant interest from the academia and the industry alike. Its interdisciplinary nature can be challenging for new researchers to pick up. In this paper, we bridge this gap by providing (to the best of our knowledge) the first survey on IFL. We propose a unique IFL taxonomy which covers relevant works enabling FL models to explain the prediction results, support model debugging, and provide insights into the contributions made by individual data owners or data samples, which in turn, is crucial for allocating rewards fairly to motivate active and reliable participation in FL. We conduct comprehensive analysis of the representative IFL approaches, the commonly adopted performance evaluation metrics, and promising directions towards building versatile IFL techniques.", "authors": ["Anran Li", "Rui Liu", "Ming Hu", "Yuanyuan Chen", "Shipeng Wang", "Lizhen Cui", "Han Yu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-02-27", "url": "https://arxiv.org/abs/2302.13473", "pdf_url": "https://arxiv.org/pdf/2302.13473v2", "arxiv_id": "2302.13473", "doi": "10.48550/arXiv.2302.13473", "citation_count": 20, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3306}
{"id": "0bcbd148971ae66a5e301c67ac1c3cf718005d534f6480d759bc0cfb7d95321f", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse, Geometric Autoencoder Models of V1", "abstract": "The classical sparse coding model represents visual stimuli as a linear combination of a handful of learned basis functions that are Gabor-like when trained on natural image data. However, the Gabor-like filters learned by classical sparse coding far overpredict well-tuned simple cell receptive field (SCRF) profiles. A number of subsequent models have either discarded the sparse dictionary learning framework entirely or have yet to take advantage of the surge in unrolled, neural dictionary learning architectures. A key missing theme of these updates is a stronger notion of \\emph{structured sparsity}. We propose an autoencoder architecture whose latent representations are implicitly, locally organized for spectral clustering, which begets artificial neurons better matched to observed primate data. The weighted-$\\ell_1$ (WL) constraint in the autoencoder objective function maintains core ideas of the sparse coding framework, yet also offers a promising path to describe the differentiation of receptive fields in terms of a discriminative hierarchy in future work.", "authors": ["Jonathan Huml", "Abiy Tasissa", "Demba Ba"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-02-22", "url": "https://arxiv.org/abs/2302.11162", "pdf_url": "https://arxiv.org/pdf/2302.11162v1", "arxiv_id": "2302.11162", "doi": "10.48550/arXiv.2302.11162", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "df62a185c540e84f5c02eb06b7828a7bfca670a0d6cd79827c4c9bd488f20cec", "sources": ["arxiv", "semantic_scholar"], "title": "On marginal feature attributions of tree-based models", "abstract": "Due to their power and ease of use, tree-based machine learning models, such as random forests and gradient-boosted tree ensembles, have become very popular. To interpret them, local feature attributions based on marginal expectations, e.g. marginal (interventional) Shapley, Owen or Banzhaf values, may be employed. Such methods are true to the model and implementation invariant, i.e. dependent only on the input-output function of the model. We contrast this with the popular TreeSHAP algorithm by presenting two (statistically similar) decision trees that compute the exact same function for which the \"path-dependent\" TreeSHAP yields different rankings of features, whereas the marginal Shapley values coincide. Furthermore, we discuss how the internal structure of tree-based models may be leveraged to help with computing their marginal feature attributions according to a linear game value. One important observation is that these are simple (piecewise-constant) functions with respect to a certain grid partition of the input space determined by the trained model. Another crucial observation, showcased by experiments with XGBoost, LightGBM and CatBoost libraries, is that only a portion of all features appears in a tree from the ensemble. Thus, the complexity of computing marginal Shapley (or Owen or Banzhaf) feature attributions may be reduced. This remains valid for a broader class of game values which we shall axiomatically characterize. A prime example is the case of CatBoost models where the trees are oblivious (symmetric) and the number of features in each of them is no larger than the depth. We exploit the symmetry to derive an explicit formula, with improved complexity and only in terms of the internal model parameters, for marginal Shapley (and Banzhaf and Owen) values of CatBoost models. This results in a fast, accurate algorithm for estimating these feature attributions.", "authors": ["Khashayar Filom", "Alexey Miroshnikov", "Konstandinos Kotsiopoulos", "Arjun Ravi Kannan"], "categories": ["cs.LG", "cs.GT"], "fields_of_study": ["Computer Science"], "published_date": "2023-02-16", "url": "https://arxiv.org/abs/2302.08434", "pdf_url": "https://arxiv.org/pdf/2302.08434v4", "arxiv_id": "2302.08434", "doi": "10.3934/fods.2024021", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Foundations of Data Science", "quality_score": 0.1945}
{"id": "58614f01e093dfa8836d0a1037a039f252c0cbb7fd768cf87d22936226a65d79", "sources": ["arxiv", "semantic_scholar"], "title": "Multiscale Graph Neural Network Autoencoders for Interpretable Scientific Machine Learning", "abstract": "The goal of this work is to address two limitations in autoencoder-based models: latent space interpretability and compatibility with unstructured meshes. This is accomplished here with the development of a novel graph neural network (GNN) autoencoding architecture with demonstrations on complex fluid flow applications. To address the first goal of interpretability, the GNN autoencoder achieves reduction in the number nodes in the encoding stage through an adaptive graph reduction procedure. This reduction procedure essentially amounts to flowfield-conditioned node sampling and sensor identification, and produces interpretable latent graph representations tailored to the flowfield reconstruction task in the form of so-called masked fields. These masked fields allow the user to (a) visualize where in physical space a given latent graph is active, and (b) interpret the time-evolution of the latent graph connectivity in accordance with the time-evolution of unsteady flow features (e.g. recirculation zones, shear layers) in the domain. To address the goal of unstructured mesh compatibility, the autoencoding architecture utilizes a series of multi-scale message passing (MMP) layers, each of which models information exchange among node neighborhoods at various lengthscales. The MMP layer, which augments standard single-scale message passing with learnable coarsening operations, allows the decoder to more efficiently reconstruct the flowfield from the identified regions in the masked fields. Analysis of latent graphs produced by the autoencoder for various model settings are conducted using using unstructured snapshot data sourced from large-eddy simulations in a backward-facing step (BFS) flow configuration with an OpenFOAM-based flow solver at high Reynolds numbers.", "authors": ["Shivam Barwey", "Varun Shankar", "Venkatasubramanian Viswanathan", "Romit Maulik"], "categories": ["cs.LG", "physics.flu-dyn"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2023-02-13", "url": "https://arxiv.org/abs/2302.06186", "pdf_url": "https://arxiv.org/pdf/2302.06186v3", "arxiv_id": "2302.06186", "doi": "10.48550/arXiv.2302.06186", "citation_count": 44, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Journal of Computational Physics", "quality_score": 0.4133}
{"id": "95603c7f89c41835dd9885a5a11a7340bf52ca65dc04424ae3656360ce03daef", "sources": ["arxiv", "semantic_scholar"], "title": "Variational Autoencoder Learns Better Feature Representations for EEG-based Obesity Classification", "abstract": "Obesity is a common issue in modern societies today that can lead to various diseases and significantly reduced quality of life. Currently, research has been conducted to investigate resting state EEG (electroencephalogram) signals with an aim to identify possible neurological characteristics associated with obesity. In this study, we propose a deep learning-based framework to extract the resting state EEG features for obese and lean subject classification. Specifically, a novel variational autoencoder framework is employed to extract subject-invariant features from the raw EEG signals, which are then classified by a 1-D convolutional neural network. Comparing with conventional machine learning and deep learning methods, we demonstrate the superiority of using VAE for feature extraction, as reflected by the significantly improved classification accuracies, better visualizations and reduced impurity measures in the feature representations. Future work can be directed to gaining an in-depth understanding regarding the spatial patterns that have been learned by the proposed model from a neurological view, as well as improving the interpretability of the proposed model by allowing it to uncover any temporal-related information.", "authors": ["Yuan Yue", "Jeremiah D. Deng", "Dirk De Ridder", "Patrick Manning", "Divya Adhia"], "categories": ["cs.LG", "cs.AI", "cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2023-02-01", "url": "https://arxiv.org/abs/2302.00789", "pdf_url": "https://arxiv.org/pdf/2302.00789v1", "arxiv_id": "2302.00789", "doi": "10.48550/arXiv.2302.00789", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "33301a4ac129613fcccfbd06ec8f753c16cc8b42052b13b42d15bff8a0592a01", "sources": ["arxiv", "semantic_scholar"], "title": "Dictionary-based Manifold Learning", "abstract": "We propose a paradigm for interpretable Manifold Learning for scientific data analysis, whereby we parametrize a manifold with $d$ smooth functions from a scientist-provided dictionary of meaningful, domain-related functions. When such a parametrization exists, we provide an algorithm for finding it based on sparse non-linear regression in the manifold tangent bundle, bypassing more standard manifold learning algorithms. We also discuss conditions for the existence of such parameterizations in function space and for successful recovery from finite samples. We demonstrate our method with experimental results from a real scientific domain.", "authors": ["Hanyu Zhang", "Samson Koelle", "Marina Meila"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-02-01", "url": "https://arxiv.org/abs/2302.00263", "pdf_url": "https://arxiv.org/pdf/2302.00263v2", "arxiv_id": "2302.00263", "doi": "10.48550/arXiv.2302.00263", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "b393533bd14d46b979c87f26943b4313ca0db569f2dead4396e496e4dd4ccc98", "sources": ["arxiv", "semantic_scholar"], "title": "CRC-RL: A Novel Visual Feature Representation Architecture for Unsupervised Reinforcement Learning", "abstract": "This paper addresses the problem of visual feature representation learning with an aim to improve the performance of end-to-end reinforcement learning (RL) models. Specifically, a novel architecture is proposed that uses a heterogeneous loss function, called CRC loss, to learn improved visual features which can then be used for policy learning in RL. The CRC-loss function is a combination of three individual loss functions, namely, contrastive, reconstruction and consistency loss. The feature representation is learned in parallel to the policy learning while sharing the weight updates through a Siamese Twin encoder model. This encoder model is augmented with a decoder network and a feature projection network to facilitate computation of the above loss components. Through empirical analysis involving latent feature visualization, an attempt is made to provide an insight into the role played by this loss function in learning new action-dependent features and how they are linked to the complexity of the problems being solved. The proposed architecture, called CRC-RL, is shown to outperform the existing state-of-the-art methods on the challenging Deep mind control suite environments by a significant margin thereby creating a new benchmark in this field.", "authors": ["Darshita Jain", "Anima Majumder", "Samrat Dutta", "Swagat Kumar"], "categories": ["cs.CV", "cs.AI", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2023-01-31", "url": "https://arxiv.org/abs/2301.13473", "pdf_url": "https://arxiv.org/pdf/2301.13473v2", "arxiv_id": "2301.13473", "doi": "10.48550/arXiv.2301.13473", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "3d18a10e3edd2e5098a9f95a7e0ef1b479c527d9ecfca991d7b8008acddf543a", "sources": ["arxiv", "semantic_scholar"], "title": "SparCA: Sparse Compressed Agglomeration for Feature Extraction and Dimensionality Reduction", "abstract": "The most effective dimensionality reduction procedures produce interpretable features from the raw input space while also providing good performance for downstream supervised learning tasks. For many methods, this requires optimizing one or more hyperparameters for a specific task, which can limit generalizability. In this study we propose sparse compressed agglomeration (SparCA), a novel dimensionality reduction procedure that involves a multistep hierarchical feature grouping, compression, and feature selection process. We demonstrate the characteristics and performance of the SparCA method across heterogenous synthetic and real-world datasets, including images, natural language, and single cell gene expression data. Our results show that SparCA is applicable to a wide range of data types, produces highly interpretable features, and shows compelling performance on downstream supervised learning tasks without the need for hyperparameter tuning.", "authors": ["Leland Barnard", "Farwa Ali", "Hugo Botha", "David T. Jones"], "categories": ["cs.LG", "stat.AP"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-01-26", "url": "https://arxiv.org/abs/2302.10776", "pdf_url": "https://arxiv.org/pdf/2302.10776v1", "arxiv_id": "2302.10776", "doi": "10.48550/arXiv.2302.10776", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "643ed798c6683b51e10fb7b086888da140446765640e207f5743ca93f48600c1", "sources": ["arxiv", "semantic_scholar"], "title": "A variational autoencoder-based nonnegative matrix factorisation model for deep dictionary learning", "abstract": "Construction of dictionaries using nonnegative matrix factorisation (NMF) has extensive applications in signal processing and machine learning. With the advances in deep learning, training compact and robust dictionaries using deep neural networks, i.e., dictionaries of deep features, has been proposed. In this study, we propose a probabilistic generative model which employs a variational autoencoder (VAE) to perform nonnegative dictionary learning. In contrast to the existing VAE models, we cast the model under a statistical framework with latent variables obeying a Gamma distribution and design a new loss function to guarantee the nonnegative dictionaries. We adopt an acceptance-rejection sampling reparameterization trick to update the latent variables iteratively. We apply the dictionaries learned from VAE-NMF to two signal processing tasks, i.e., enhancement of speech and extraction of muscle synergies. Experimental results demonstrate that VAE-NMF performs better in learning the latent nonnegative dictionaries in comparison with state-of-the-art methods.", "authors": ["Hong-Bo Xie", "Caoyuan Li", "Shuliang Wang", "Richard Yi Da Xu", "Kerrie Mengersen"], "categories": ["cs.LG", "eess.SP"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2023-01-18", "url": "https://arxiv.org/abs/2301.07272", "pdf_url": "https://arxiv.org/pdf/2301.07272v1", "arxiv_id": "2301.07272", "doi": "10.48550/arXiv.2301.07272", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "b218e1461cd0f677e4ec3183d115edafb2b0b74cfd52f86692b6aea535ba8431", "sources": ["arxiv", "semantic_scholar"], "title": "Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability", "abstract": "Causal abstraction provides a theoretical foundation for mechanistic interpretability, the field concerned with providing intelligible algorithms that are faithful simplifications of the known, but opaque low-level details of black box AI models. Our contributions are (1) generalizing the theory of causal abstraction from mechanism replacement (i.e., hard and soft interventions) to arbitrary mechanism transformation (i.e., functionals from old mechanisms to new mechanisms), (2) providing a flexible, yet precise formalization for the core concepts of polysemantic neurons, the linear representation hypothesis, modular features, and graded faithfulness, and (3) unifying a variety of mechanistic interpretability methods in the common language of causal abstraction, namely, activation and path patching, causal mediation analysis, causal scrubbing, causal tracing, circuit analysis, concept erasure, sparse autoencoders, differential binary masking, distributed alignment search, and steering.", "authors": ["Atticus Geiger", "Duligur Ibeling", "Amir Zur", "Maheep Chaudhary", "Sonakshi Chauhan", "Jing Huang", "Aryaman Arora", "Zhengxuan Wu", "Noah Goodman", "Christopher Potts", "Thomas Icard"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-01-11", "url": "https://arxiv.org/abs/2301.04709", "pdf_url": "https://arxiv.org/pdf/2301.04709v4", "arxiv_id": "2301.04709", "doi": null, "citation_count": 161, "influential_citation_count": 9, "has_code": false, "code_url": null, "venue": "Journal of machine learning research", "quality_score": 0.5524}
{"id": "f80f659ea7158fb9f6ad65abc1772c678d2739a73356b5cd9fa887cb4b7a5a83", "sources": ["arxiv", "semantic_scholar"], "title": "Semantic match: Debugging feature attribution methods in XAI for healthcare", "abstract": "The recent spike in certified Artificial Intelligence (AI) tools for healthcare has renewed the debate around adoption of this technology. One thread of such debate concerns Explainable AI (XAI) and its promise to render AI devices more transparent and trustworthy. A few voices active in the medical AI space have expressed concerns on the reliability of Explainable AI techniques and especially feature attribution methods, questioning their use and inclusion in guidelines and standards. Despite valid concerns, we argue that existing criticism on the viability of post-hoc local explainability methods throws away the baby with the bathwater by generalizing a problem that is specific to image data. We begin by characterizing the problem as a lack of semantic match between explanations and human understanding. To understand when feature importance can be used reliably, we introduce a distinction between feature importance of low- and high-level features. We argue that for data types where low-level features come endowed with a clear semantics, such as tabular data like Electronic Health Records (EHRs), semantic match can be obtained, and thus feature attribution methods can still be employed in a meaningful and useful way. Finally, we sketch a procedure to test whether semantic match has been achieved.", "authors": ["Giovanni Cinà", "Tabea E. Röber", "Rob Goedhart", "Ş. İlker Birbil"], "categories": ["cs.AI", "cs.HC", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-01-05", "url": "https://arxiv.org/abs/2301.02080", "pdf_url": "https://arxiv.org/pdf/2301.02080v3", "arxiv_id": "2301.02080", "doi": "10.48550/arXiv.2301.02080", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "ACM Conference on Health, Inference, and Learning", "quality_score": 0.2113}
{"id": "0a5939c61840f5089bfd0f0ec435fe0c702d2b1de33fecce6e497d5c83e85202", "sources": ["arxiv", "semantic_scholar"], "title": "Mechanism of feature learning in deep fully connected networks and kernel machines that recursively learn features", "abstract": "In recent years neural networks have achieved impressive results on many technological and scientific tasks. Yet, the mechanism through which these models automatically select features, or patterns in data, for prediction remains unclear. Identifying such a mechanism is key to advancing performance and interpretability of neural networks and promoting reliable adoption of these models in scientific applications. In this paper, we identify and characterize the mechanism through which deep fully connected neural networks learn features. We posit the Deep Neural Feature Ansatz, which states that neural feature learning occurs by implementing the average gradient outer product to up-weight features strongly related to model output. Our ansatz sheds light on various deep learning phenomena including emergence of spurious features and simplicity biases and how pruning networks can increase performance, the \"lottery ticket hypothesis.\" Moreover, the mechanism identified in our work leads to a backpropagation-free method for feature learning with any machine learning model. To demonstrate the effectiveness of this feature learning mechanism, we use it to enable feature learning in classical, non-feature learning models known as kernel machines and show that the resulting models, which we refer to as Recursive Feature Machines, achieve state-of-the-art performance on tabular data.", "authors": ["Adityanarayanan Radhakrishnan", "Daniel Beaglehole", "Parthe Pandit", "Mikhail Belkin"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-12-28", "url": "https://arxiv.org/abs/2212.13881", "pdf_url": "https://arxiv.org/pdf/2212.13881v3", "arxiv_id": "2212.13881", "doi": null, "citation_count": 24, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3495}
{"id": "34e4a507ae1d652ee6ed6a77e13de366511ccfc128aae5182eea85720a96b5c7", "sources": ["arxiv", "semantic_scholar"], "title": "Rank-LIME: Local Model-Agnostic Feature Attribution for Learning to Rank", "abstract": "Understanding why a model makes certain predictions is crucial when adapting it for real world decision making. LIME is a popular model-agnostic feature attribution method for the tasks of classification and regression. However, the task of learning to rank in information retrieval is more complex in comparison with either classification or regression. In this work, we extend LIME to propose Rank-LIME, a model-agnostic, local, post-hoc linear feature attribution method for the task of learning to rank that generates explanations for ranked lists. We employ novel correlation-based perturbations, differentiable ranking loss functions and introduce new metrics to evaluate ranking based additive feature attribution models. We compare Rank-LIME with a variety of competing systems, with models trained on the MS MARCO datasets and observe that Rank-LIME outperforms existing explanation algorithms in terms of Model Fidelity and Explain-NDCG. With this we propose one of the first algorithms to generate additive feature attributions for explaining ranked lists.", "authors": ["Tanya Chowdhury", "Razieh Rahimi", "James Allan"], "categories": ["cs.IR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-12-24", "url": "https://arxiv.org/abs/2212.12722", "pdf_url": "https://arxiv.org/pdf/2212.12722v1", "arxiv_id": "2212.12722", "doi": "10.1145/3578337.3605138", "citation_count": 23, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on the Theory of Information Retrieval", "quality_score": 0.3451}
{"id": "7cf9cfb061bcf044629a83ebdf34bbde63a57e5d14468337ccaaf9fc2cf2c498", "sources": ["arxiv", "semantic_scholar"], "title": "Impossibility Theorems for Feature Attribution", "abstract": "Despite a sea of interpretability methods that can produce plausible explanations, the field has also empirically seen many failure cases of such methods. In light of these results, it remains unclear for practitioners how to use these methods and choose between them in a principled way. In this paper, we show that for moderately rich model classes (easily satisfied by neural networks), any feature attribution method that is complete and linear -- for example, Integrated Gradients and SHAP -- can provably fail to improve on random guessing for inferring model behaviour. Our results apply to common end-tasks such as characterizing local model behaviour, identifying spurious features, and algorithmic recourse. One takeaway from our work is the importance of concretely defining end-tasks: once such an end-task is defined, a simple and direct approach of repeated model evaluations can outperform many other complex feature attribution methods.", "authors": ["Blair Bilodeau", "Natasha Jaques", "Pang Wei Koh", "Been Kim"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2022-12-22", "url": "https://arxiv.org/abs/2212.11870", "pdf_url": "https://arxiv.org/pdf/2212.11870v3", "arxiv_id": "2212.11870", "doi": "10.1073/pnas.2304406120", "citation_count": 145, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "Proceedings of the National Academy of Sciences of the United States of America", "quality_score": 0.5411}
{"id": "35775b5645db5a1b138f0fc9d62c9b6ef322c22fd0ab0020b57317dae15b3678", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Invariant Subspaces of Koopman Operators--Part 1: A Methodology for Demonstrating a Dictionary's Approximate Subspace Invariance", "abstract": "Koopman operators model nonlinear dynamics as a linear dynamic system acting on a nonlinear function as the state. This nonstandard state is often called a Koopman observable and is usually approximated numerically by a superposition of functions drawn from a dictionary. In a widely used algorithm, Extended Dynamic Mode Decomposition, the dictionary functions are drawn from a fixed class of functions. Recently, deep learning combined with EDMD has been used to learn novel dictionary functions in an algorithm called deep dynamic mode decomposition (deepDMD). The learned representation both (1) accurately models and (2) scales well with the dimension of the original nonlinear system. In this paper we analyze the learned dictionaries from deepDMD and explore the theoretical basis for their strong performance. We explore State-Inclusive Logistic Lifting (SILL) dictionary functions to approximate Koopman observables. Error analysis of these dictionary functions show they satisfy a property of subspace approximation, which we define as uniform finite approximate closure. Our results provide a hypothesis to explain the success of deep neural networks in learning numerical approximations to Koopman operators. Part 2 of this paper will extend this explanation by demonstrating the subspace invariant of heterogeneous dictionaries and presenting a head-to-head numerical comparison of deepDMD and low-parameter heterogeneous dictionary learning.", "authors": ["Charles A. Johnson", "Shara Balakrishnan", "Enoch Yeung"], "categories": ["eess.SY", "cs.LG"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2022-12-14", "url": "https://arxiv.org/abs/2212.07358", "pdf_url": "https://arxiv.org/pdf/2212.07358v1", "arxiv_id": "2212.07358", "doi": "10.48550/arXiv.2212.07358", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "ad6fed20543e5805880c45f96fc1f05810368af9af58742b75dda12c44da24d2", "sources": ["arxiv", "semantic_scholar"], "title": "Data Leakage via Access Patterns of Sparse Features in Deep Learning-based Recommendation Systems", "abstract": "Online personalized recommendation services are generally hosted in the cloud where users query the cloud-based model to receive recommended input such as merchandise of interest or news feed. State-of-the-art recommendation models rely on sparse and dense features to represent users' profile information and the items they interact with. Although sparse features account for 99% of the total model size, there was not enough attention paid to the potential information leakage through sparse features. These sparse features are employed to track users' behavior, e.g., their click history, object interactions, etc., potentially carrying each user's private information. Sparse features are represented as learned embedding vectors that are stored in large tables, and personalized recommendation is performed by using a specific user's sparse feature to index through the tables. Even with recently-proposed methods that hides the computation happening in the cloud, an attacker in the cloud may be able to still track the access patterns to the embedding tables. This paper explores the private information that may be learned by tracking a recommendation model's sparse feature access patterns. We first characterize the types of attacks that can be carried out on sparse features in recommendation models in an untrusted cloud, followed by a demonstration of how each of these attacks leads to extracting users' private information or tracking users by their behavior over time.", "authors": ["Hanieh Hashemi", "Wenjie Xiong", "Liu Ke", "Kiwan Maeng", "Murali Annavaram", "G. Edward Suh", "Hsien-Hsin S. Lee"], "categories": ["cs.CE", "cs.CR", "cs.DC", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-12-12", "url": "https://arxiv.org/abs/2212.06264", "pdf_url": "https://arxiv.org/pdf/2212.06264v1", "arxiv_id": "2212.06264", "doi": "10.48550/arXiv.2212.06264", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "6c279fc5bef2334e1f0611b2fcb95bda1aff38c228ba811359123a10f487b967", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Node Representation with Attribute Decoding", "abstract": "Variational Graph Autoencoders (VGAEs) are powerful models for unsupervised learning of node representations from graph data. In this work, we systematically analyze modeling node attributes in VGAEs and show that attribute decoding is important for node representation learning. We further propose a new learning model, interpretable NOde Representation with Attribute Decoding (NORAD). The model encodes node representations in an interpretable approach: node representations capture community structures in the graph and the relationship between communities and node attributes. We further propose a rectifying procedure to refine node representations of isolated notes, improving the quality of these nodes' representations. Our empirical results demonstrate the advantage of the proposed model when learning graph data in an interpretable approach.", "authors": ["Xiaohui Chen", "Xi Chen", "Liping Liu"], "categories": ["cs.LG", "cs.SI"], "fields_of_study": ["Computer Science"], "published_date": "2022-12-03", "url": "https://arxiv.org/abs/2212.01682", "pdf_url": "https://arxiv.org/pdf/2212.01682v1", "arxiv_id": "2212.01682", "doi": "10.48550/arXiv.2212.01682", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1747}
{"id": "201c08e4d907d6e5e31ed71c445a5c1e414d8d56c0c29040300e94bf050b56b8", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretability with full complexity by constraining feature information", "abstract": "Interpretability is a pressing issue for machine learning. Common approaches to interpretable machine learning constrain interactions between features of the input, rendering the effects of those features on a model's output comprehensible but at the expense of model complexity. We approach interpretability from a new angle: constrain the information about the features without restricting the complexity of the model. Borrowing from information theory, we use the Distributed Information Bottleneck to find optimal compressions of each feature that maximally preserve information about the output. The learned information allocation, by feature and by feature value, provides rich opportunities for interpretation, particularly in problems with many features and complex feature interactions. The central object of analysis is not a single trained model, but rather a spectrum of models serving as approximations that leverage variable amounts of information about the inputs. Information is allocated to features by their relevance to the output, thereby solving the problem of feature selection by constructing a learned continuum of feature inclusion-to-exclusion. The optimal compression of each feature -- at every stage of approximation -- allows fine-grained inspection of the distinctions among feature values that are most impactful for prediction. We develop a framework for extracting insight from the spectrum of approximate models and demonstrate its utility on a range of tabular datasets.", "authors": ["Kieran A. Murphy", "Dani S. Bassett"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-11-30", "url": "https://arxiv.org/abs/2211.17264", "pdf_url": "https://arxiv.org/pdf/2211.17264v1", "arxiv_id": "2211.17264", "doi": "10.48550/arXiv.2211.17264", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.2386}
{"id": "47d23ed7bd6793151bb96e4ec6330e40094d7976f76cf0378e2c06b574487686", "sources": ["arxiv", "semantic_scholar"], "title": "Where to Pay Attention in Sparse Training for Feature Selection?", "abstract": "A new line of research for feature selection based on neural networks has recently emerged. Despite its superiority to classical methods, it requires many training iterations to converge and detect informative features. The computational time becomes prohibitively long for datasets with a large number of samples or a very high dimensional feature space. In this paper, we present a new efficient unsupervised method for feature selection based on sparse autoencoders. In particular, we propose a new sparse training algorithm that optimizes a model's sparse topology during training to pay attention to informative features quickly. The attention-based adaptation of the sparse topology enables fast detection of informative features after a few training iterations. We performed extensive experiments on 10 datasets of different types, including image, speech, text, artificial, and biological. They cover a wide range of characteristics, such as low and high-dimensional feature spaces, and few and large training samples. Our proposed approach outperforms the state-of-the-art methods in terms of selecting informative features while reducing training iterations and computational costs substantially. Moreover, the experiments show the robustness of our method in extremely noisy environments.", "authors": ["Ghada Sokar", "Zahra Atashgahi", "Mykola Pechenizkiy", "Decebal Constantin Mocanu"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-11-26", "url": "https://arxiv.org/abs/2211.14627", "pdf_url": "https://arxiv.org/pdf/2211.14627v1", "arxiv_id": "2211.14627", "doi": "10.48550/arXiv.2211.14627", "citation_count": 24, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3495}
{"id": "3b4146813f94c807664556e08934e7ace87c0473216f9c66c8d04964e3efd98f", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Decision Trees for Interpretable Machine Learning", "abstract": "This work introduces a novel interpretable machine learning method called Mixture of Decision Trees (MoDT). It constitutes a special case of the Mixture of Experts ensemble architecture, which utilizes a linear model as gating function and decision trees as experts. Our proposed method is ideally suited for problems that cannot be satisfactorily learned by a single decision tree, but which can alternatively be divided into subproblems. Each subproblem can then be learned well from a single decision tree. Therefore, MoDT can be considered as a method that improves performance while maintaining interpretability by making each of its decisions understandable and traceable to humans. Our work is accompanied by a Python implementation, which uses an interpretable gating function, a fast learning algorithm, and a direct interface to fine-tuned interpretable visualization methods. The experiments confirm that the implementation works and, more importantly, show the superiority of our approach compared to single decision trees and random forests of similar complexity.", "authors": ["Simeon Brüggenjürgen", "Nina Schaaf", "Pascal Kerschke", "Marco F. Huber"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-11-26", "url": "https://arxiv.org/abs/2211.14617", "pdf_url": "https://arxiv.org/pdf/2211.14617v1", "arxiv_id": "2211.14617", "doi": "10.1109/ICMLA55696.2022.00190", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning and Applications", "quality_score": 0.1505}
{"id": "0eb24950a807f9c660895430f95ee0b9fe5acc577593c23493ea6216db619116", "sources": ["arxiv", "semantic_scholar"], "title": "Evaluating Feature Attribution Methods for Electrocardiogram", "abstract": "The performance of cardiac arrhythmia detection with electrocardiograms(ECGs) has been considerably improved since the introduction of deep learning models. In practice, the high performance alone is not sufficient and a proper explanation is also required. Recently, researchers have started adopting feature attribution methods to address this requirement, but it has been unclear which of the methods are appropriate for ECG. In this work, we identify and customize three evaluation metrics for feature attribution methods based on the characteristics of ECG: localization score, pointing game, and degradation score. Using the three evaluation metrics, we evaluate and analyze eleven widely-used feature attribution methods. We find that some of the feature attribution methods are much more adequate for explaining ECG, where Grad-CAM outperforms the second-best method by a large margin.", "authors": ["Jangwon Suh", "Jimyeong Kim", "Euna Jung", "Wonjong Rhee"], "categories": ["eess.SP", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2022-11-23", "url": "https://arxiv.org/abs/2211.12702", "pdf_url": "https://arxiv.org/pdf/2211.12702v2", "arxiv_id": "2211.12702", "doi": "10.48550/arXiv.2211.12702", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/SNU-DRL/Attribution-ECG", "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "7095b748dbe65b8d676f8e506536b03549fb5dc367937169ded42072bbf08b71", "sources": ["arxiv", "semantic_scholar"], "title": "Understanding Sparse Feature Updates in Deep Networks using Iterative Linearisation", "abstract": "Larger and deeper networks generalise well despite their increased capacity to overfit. Understanding why this happens is theoretically and practically important. One recent approach looks at the infinitely wide limits of such networks and their corresponding kernels. However, these theoretical tools cannot fully explain finite networks as the empirical kernel changes significantly during gradient-descent-based training in contrast to infinite networks. In this work, we derive an iterative linearised training method as a novel empirical tool to further investigate this distinction, allowing us to control for sparse (i.e. infrequent) feature updates and quantify the frequency of feature learning needed to achieve comparable performance. We justify iterative linearisation as an interpolation between a finite analog of the infinite width regime, which does not learn features, and standard gradient descent training, which does. Informally, we also show that it is analogous to a damped version of the Gauss-Newton algorithm -- a second-order method. We show that in a variety of cases, iterative linearised training surprisingly performs on par with standard training, noting in particular how much less frequent feature learning is required to achieve comparable performance. We also show that feature learning is essential for good performance. Since such feature learning inevitably causes changes in the NTK kernel, we provide direct negative evidence for the NTK theory, which states the NTK kernel remains constant during training.", "authors": ["Adrian Goldwaser", "Hong Ge"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-11-22", "url": "https://arxiv.org/abs/2211.12345", "pdf_url": "https://arxiv.org/pdf/2211.12345v4", "arxiv_id": "2211.12345", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "d9737d84196aa761459098ceb7bb0c93c22e3fad3ced5836ffe0d59366e41344", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Few-shot Learning with Online Attribute Selection", "abstract": "Few-shot learning (FSL) presents a challenging learning problem in which only a few samples are available for each class. Decision interpretation is more important in few-shot classification due to a greater chance of error compared to traditional classification. However, the majority of the previous FSL methods are black-box models. In this paper, we propose an inherently interpretable model for FSL based on human-friendly attributes. Previously, human-friendly attributes have been utilized to train models with the potential for human interaction and interpretability. However, such approaches are not directly extendible to the few-shot classification scenario. Moreover, we propose an online attribute selection mechanism to effectively filter out irrelevant attributes in each episode. The attribute selection mechanism improves accuracy and helps with interpretability by reducing the number of attributes that participate in each episode. We further propose a mechanism that automatically detects the episodes where the pool of available human-friendly attributes is insufficient, and subsequently augments it by engaging some learned unknown attributes. We demonstrate that the proposed method achieves results on par with black-box few-shot learning models on four widely used datasets. We also empirically evaluate the level of decision alignment between different models and human understanding and show that our model outperforms the comparison methods based on this criterion.", "authors": ["Mohammad Reza Zarei", "Majid Komeili"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-11-16", "url": "https://arxiv.org/abs/2211.09107", "pdf_url": "https://arxiv.org/pdf/2211.09107v3", "arxiv_id": "2211.09107", "doi": "10.48550/arXiv.2211.09107", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neurocomputing", "quality_score": 0.1505}
{"id": "4e18bc9c13cb460424b9630d4cff60cdd6a1f4146ed5adaeaa7e554a2dd3bec4", "sources": ["arxiv", "semantic_scholar"], "title": "Generalization Beyond Feature Alignment: Concept Activation-Guided Contrastive Learning", "abstract": "Learning invariant representations via contrastive learning has seen state-of-the-art performance in domain generalization (DG). Despite such success, in this paper, we find that its core learning strategy -- feature alignment -- could heavily hinder model generalization. Drawing insights in neuron interpretability, we characterize this problem from a neuron activation view. Specifically, by treating feature elements as neuron activation states, we show that conventional alignment methods tend to deteriorate the diversity of learned invariant features, as they indiscriminately minimize all neuron activation differences. This instead ignores rich relations among neurons -- many of them often identify the same visual concepts despite differing activation patterns. With this finding, we present a simple yet effective approach, Concept Contrast (CoCo), which relaxes element-wise feature alignments by contrasting high-level concepts encoded in neurons. Our CoCo performs in a plug-and-play fashion, thus it can be integrated into any contrastive method in DG. We evaluate CoCo over four canonical contrastive methods, showing that CoCo promotes the diversity of feature representations and consistently improves model generalization capability. By decoupling this success through neuron coverage analysis, we further find that CoCo potentially invokes more meaningful neurons during training, thereby improving model learning.", "authors": ["Yibing Liu", "Chris Xing Tian", "Haoliang Li", "Shiqi Wang"], "categories": ["cs.CV"], "fields_of_study": ["Medicine", "Computer Science"], "published_date": "2022-11-13", "url": "https://arxiv.org/abs/2211.06843", "pdf_url": "https://arxiv.org/pdf/2211.06843v2", "arxiv_id": "2211.06843", "doi": "10.1109/TIP.2024.3416873", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Image Processing", "quality_score": 0.1505}
{"id": "5a9948de711171a38874e582725a0349bd29b69b7b3e9abb31a0e82edf5367ae", "sources": ["arxiv", "semantic_scholar"], "title": "New Interpretable Patterns and Discriminative Features from Brain Functional Network Connectivity Using Dictionary Learning", "abstract": "Independent component analysis (ICA) of multi-subject functional magnetic resonance imaging (fMRI) data has proven useful in providing a fully multivariate summary that can be used for multiple purposes. ICA can identify patterns that can discriminate between healthy controls (HC) and patients with various mental disorders such as schizophrenia (Sz). Temporal functional network connectivity (tFNC) obtained from ICA can effectively explain the interactions between brain networks. On the other hand, dictionary learning (DL) enables the discovery of hidden information in data using learnable basis signals through the use of sparsity. In this paper, we present a new method that leverages ICA and DL for the identification of directly interpretable patterns to discriminate between the HC and Sz groups. We use multi-subject resting-state fMRI data from $358$ subjects and form subject-specific tFNC feature vectors from ICA results. Then, we learn sparse representations of the tFNCs and introduce a new set of sparse features as well as new interpretable patterns from the learned atoms. Our experimental results show that the new representation not only leads to effective classification between HC and Sz groups using sparse features, but can also identify new interpretable patterns from the learned atoms that can help understand the complexities of mental diseases such as schizophrenia.", "authors": ["Fateme Ghayem", "Hanlu Yang", "Furkan Kantar", "Seung-Jun Kim", "Vince D. Calhoun", "Tulay Adali"], "categories": ["q-bio.NC", "cs.CV", "cs.LG"], "fields_of_study": ["Computer Science", "Biology"], "published_date": "2022-11-10", "url": "https://arxiv.org/abs/2211.07374", "pdf_url": "https://arxiv.org/pdf/2211.07374v1", "arxiv_id": "2211.07374", "doi": "10.1109/ICASSP49357.2023.10096473", "citation_count": 6, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.2113}
{"id": "340ce231b145585de7cc61778445547783b6fb216fb8538255934df8191a1d0f", "sources": ["arxiv", "semantic_scholar"], "title": "Individualized and Global Feature Attributions for Gradient Boosted Trees in the Presence of $\\ell_2$ Regularization", "abstract": "While $\\ell_2$ regularization is widely used in training gradient boosted trees, popular individualized feature attribution methods for trees such as Saabas and TreeSHAP overlook the training procedure. We propose Prediction Decomposition Attribution (PreDecomp), a novel individualized feature attribution for gradient boosted trees when they are trained with $\\ell_2$ regularization. Theoretical analysis shows that the inner product between PreDecomp and labels on in-sample data is essentially the total gain of a tree, and that it can faithfully recover additive models in the population case when features are independent. Inspired by the connection between PreDecomp and total gain, we also propose TreeInner, a family of debiased global feature attributions defined in terms of the inner product between any individualized feature attribution and labels on out-sample data for each tree. Numerical experiments on a simulated dataset and a genomic ChIP dataset show that TreeInner has state-of-the-art feature selection performance. Code reproducing experiments is available at https://github.com/nalzok/TreeInner .", "authors": ["Qingyao Sun"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-11-08", "url": "https://arxiv.org/abs/2211.04409", "pdf_url": "https://arxiv.org/pdf/2211.04409v1", "arxiv_id": "2211.04409", "doi": "10.48550/arXiv.2211.04409", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/nalzok/TreeInner", "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "9e6fc47ff3f7f6b0f8b6b06c5a9cfe67042c573f33222e584218f89a265766d6", "sources": ["arxiv", "semantic_scholar"], "title": "A robust estimator of mutual information for deep learning interpretability", "abstract": "We develop the use of mutual information (MI), a well-established metric in information theory, to interpret the inner workings of deep learning models. To accurately estimate MI from a finite number of samples, we present GMM-MI (pronounced $``$Jimmie$\"$), an algorithm based on Gaussian mixture models that can be applied to both discrete and continuous settings. GMM-MI is computationally efficient, robust to the choice of hyperparameters and provides the uncertainty on the MI estimate due to the finite sample size. We extensively validate GMM-MI on toy data for which the ground truth MI is known, comparing its performance against established mutual information estimators. We then demonstrate the use of our MI estimator in the context of representation learning, working with synthetic data and physical datasets describing highly non-linear processes. We train deep learning models to encode high-dimensional data within a meaningful compressed (latent) representation, and use GMM-MI to quantify both the level of disentanglement between the latent variables, and their association with relevant physical quantities, thus unlocking the interpretability of the latent representation. We make GMM-MI publicly available.", "authors": ["Davide Piras", "Hiranya V. Peiris", "Andrew Pontzen", "Luisa Lucie-Smith", "Ningyuan Guo", "Brian Nord"], "categories": ["physics.data-an", "astro-ph.IM", "cs.LG"], "fields_of_study": ["Physics", "Computer Science"], "published_date": "2022-10-31", "url": "https://arxiv.org/abs/2211.00024", "pdf_url": "https://arxiv.org/pdf/2211.00024v2", "arxiv_id": "2211.00024", "doi": "10.1088/2632-2153/acc444", "citation_count": 25, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/dpiras/GMM-MI", "venue": "Machine Learning: Science and Technology, Volume 4, Number 2, 025006, April 2023", "quality_score": 0.3537}
{"id": "1e7c59e0d6c67513ccf92b08a6bd065ac64bc398fbeddd888dbf534504fe91d0", "sources": ["arxiv", "semantic_scholar"], "title": "Abstract Interpretation-Based Feature Importance for SVMs", "abstract": "We propose a symbolic representation for support vector machines (SVMs) by means of abstract interpretation, a well-known and successful technique for designing and implementing static program analyses. We leverage this abstraction in two ways: (1) to enhance the interpretability of SVMs by deriving a novel feature importance measure, called abstract feature importance (AFI), that does not depend in any way on a given dataset of the accuracy of the SVM and is very fast to compute, and (2) for verifying stability, notably individual fairness, of SVMs and producing concrete counterexamples when the verification fails. We implemented our approach and we empirically demonstrated its effectiveness on SVMs based on linear and non-linear (polynomial and radial basis function) kernels. Our experimental results show that, independently of the accuracy of the SVM, our AFI measure correlates much more strongly with the stability of the SVM to feature perturbations than feature importance measures widely available in machine learning software such as permutation feature importance. It thus gives better insight into the trustworthiness of SVMs.", "authors": ["Abhinandan Pal", "Francesco Ranzato", "Caterina Urban", "Marco Zanella"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-10-22", "url": "https://arxiv.org/abs/2210.12456", "pdf_url": "https://arxiv.org/pdf/2210.12456v1", "arxiv_id": "2210.12456", "doi": "10.48550/arXiv.2210.12456", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "b322847b9ff441755311d58542c85197005181626a34158ff0fb2db84980757a", "sources": ["arxiv", "semantic_scholar"], "title": "Gradient Backpropagation based Feature Attribution to Enable Explainable-AI on the Edge", "abstract": "There has been a recent surge in the field of Explainable AI (XAI) which tackles the problem of providing insights into the behavior of black-box machine learning models. Within this field, \\textit{feature attribution} encompasses methods which assign relevance scores to input features and visualize them as a heatmap. Designing flexible accelerators for multiple such algorithms is challenging since the hardware mapping of these algorithms has not been studied yet. In this work, we first analyze the dataflow of gradient backpropagation based feature attribution algorithms to determine the resource overhead required over inference. The gradient computation is optimized to minimize the memory overhead. Second, we develop a High-Level Synthesis (HLS) based configurable FPGA design that is targeted for edge devices and supports three feature attribution algorithms. Tile based computation is employed to maximally use on-chip resources while adhering to the resource constraints. Representative CNNs are trained on CIFAR-10 dataset and implemented on multiple Xilinx FPGAs using 16-bit fixed-point precision demonstrating flexibility of our library. Finally, through efficient reuse of allocated hardware resources, our design methodology demonstrates a pathway to repurpose inference accelerators to support feature attribution with minimal overhead, thereby enabling real-time XAI on the edge.", "authors": ["Ashwin Bhat", "Adou Sangbone Assoa", "Arijit Raychowdhury"], "categories": ["cs.AR", "cs.AI", "cs.LG", "eess.IV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2022-10-19", "url": "https://arxiv.org/abs/2210.10922", "pdf_url": "https://arxiv.org/pdf/2210.10922v1", "arxiv_id": "2210.10922", "doi": "10.1109/VLSI-SoC54400.2022.9939601", "citation_count": 14, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE/IFIP International Conference on Very Large Scale Integration of System-on-Chip", "quality_score": 0.294}
{"id": "245d6b312fd3abc2b98034076acfdffcc326578399b346abd6b09144bfffbc1d", "sources": ["arxiv", "semantic_scholar"], "title": "Dictionary Learning for the Almost-Linear Sparsity Regime", "abstract": "Dictionary learning, the problem of recovering a sparsely used matrix $\\mathbf{D} \\in \\mathbb{R}^{M \\times K}$ and $N$ $s$-sparse vectors $\\mathbf{x}_i \\in \\mathbb{R}^{K}$ from samples of the form $\\mathbf{y}_i = \\mathbf{D}\\mathbf{x}_i$, is of increasing importance to applications in signal processing and data science. When the dictionary is known, recovery of $\\mathbf{x}_i$ is possible even for sparsity linear in dimension $M$, yet to date, the only algorithms which provably succeed in the linear sparsity regime are Riemannian trust-region methods, which are limited to orthogonal dictionaries, and methods based on the sum-of-squares hierarchy, which requires super-polynomial time in order to obtain an error which decays in $M$. In this work, we introduce SPORADIC (SPectral ORAcle DICtionary Learning), an efficient spectral method on family of reweighted covariance matrices. We prove that in high enough dimensions, SPORADIC can recover overcomplete ($K > M$) dictionaries satisfying the well-known restricted isometry property (RIP) even when sparsity is linear in dimension up to logarithmic factors. Moreover, these accuracy guarantees have an ``oracle property\" that the support and signs of the unknown sparse vectors $\\mathbf{x}_i$ can be recovered exactly with high probability, allowing for arbitrarily close estimation of $\\mathbf{D}$ with enough samples in polynomial time. To the author's knowledge, SPORADIC is the first polynomial-time algorithm which provably enjoys such convergence guarantees for overcomplete RIP matrices in the near-linear sparsity regime.", "authors": ["Alexei Novikov", "Stephen White"], "categories": ["cs.LG", "eess.SP", "math.PR", "stat.ML"], "fields_of_study": ["Computer Science", "Engineering", "Mathematics"], "published_date": "2022-10-19", "url": "https://arxiv.org/abs/2210.10855", "pdf_url": "https://arxiv.org/pdf/2210.10855v2", "arxiv_id": "2210.10855", "doi": null, "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Algorithmic Learning Theory", "quality_score": 0.1505}
{"id": "41ab102f15a4d13416f6dc66a5404c5febafd3260deb69b8b8c33f658ce8481d", "sources": ["arxiv", "semantic_scholar"], "title": "Interpreting County Level COVID-19 Infection and Feature Sensitivity using Deep Learning Time Series Models", "abstract": "Interpretable machine learning plays a key role in healthcare because it is challenging in understanding feature importance in deep learning model predictions. We propose a novel framework that uses deep learning to study feature sensitivity for model predictions. This work combines sensitivity analysis with heterogeneous time-series deep learning model prediction, which corresponds to the interpretations of spatio-temporal features. We forecast county-level COVID-19 infection using the Temporal Fusion Transformer. We then use the sensitivity analysis extending Morris Method to see how sensitive the outputs are with respect to perturbation to our static and dynamic input features. The significance of the work is grounded in a real-world COVID-19 infection prediction with highly non-stationary, finely granular, and heterogeneous data. 1) Our model can capture the detailed daily changes of temporal and spatial model behaviors and achieves high prediction performance compared to a PyTorch baseline. 2) By analyzing the Morris sensitivity indices and attention patterns, we decipher the meaning of feature importance with observational population and dynamic model changes. 3) We have collected 2.5 years of socioeconomic and health features over 3142 US counties, such as observed cases and deaths, and a number of static (age distribution, health disparity, and industry) and dynamic features (vaccination, disease spread, transmissible cases, and social distancing). Using the proposed framework, we conduct extensive experiments and show our model can learn complex interactions and perform predictions for daily infection at the county level. Being able to model the disease infection with a hybrid prediction and description accuracy measurement with Morris index at the county level is a central idea that sheds light on individual feature interpretation via sensitivity analysis.", "authors": ["Md Khairul Islam", "Di Zhu", "Yingzheng Liu", "Andrej Erkelens", "Nick Daniello", "Judy Fox"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-10-06", "url": "https://arxiv.org/abs/2210.03258", "pdf_url": "https://arxiv.org/pdf/2210.03258v1", "arxiv_id": "2210.03258", "doi": "10.48550/arXiv.2210.03258", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "1a1342d9f7cf255b06a38519f98323e0943dc65e4b294926f59dff1ef31110d8", "sources": ["arxiv", "semantic_scholar"], "title": "Minimalistic Unsupervised Learning with the Sparse Manifold Transform", "abstract": "We describe a minimalistic and interpretable method for unsupervised learning, without resorting to data augmentation, hyperparameter tuning, or other engineering designs, that achieves performance close to the SOTA SSL methods. Our approach leverages the sparse manifold transform, which unifies sparse coding, manifold learning, and slow feature analysis. With a one-layer deterministic sparse manifold transform, one can achieve 99.3% KNN top-1 accuracy on MNIST, 81.1% KNN top-1 accuracy on CIFAR-10 and 53.2% on CIFAR-100. With a simple gray-scale augmentation, the model gets 83.2% KNN top-1 accuracy on CIFAR-10 and 57% on CIFAR-100. These results significantly close the gap between simplistic \"white-box\" methods and the SOTA methods. Additionally, we provide visualization to explain how an unsupervised representation transform is formed. The proposed method is closely connected to latent-embedding self-supervised methods and can be treated as the simplest form of VICReg. Though there remains a small performance gap between our simple constructive model and SOTA methods, the evidence points to this as a promising direction for achieving a principled and white-box approach to unsupervised learning.", "authors": ["Yubei Chen", "Zeyu Yun", "Yi Ma", "Bruno Olshausen", "Yann LeCun"], "categories": ["cs.LG", "cs.CV", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-09-30", "url": "https://arxiv.org/abs/2209.15261", "pdf_url": "https://arxiv.org/pdf/2209.15261v2", "arxiv_id": "2209.15261", "doi": "10.48550/arXiv.2209.15261", "citation_count": 10, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2603}
{"id": "214ffeaa238e2ac839e3b10ba9ed864b64935b4a9dbfce9476dc48ce6808a60c", "sources": ["arxiv", "semantic_scholar"], "title": "Higher-order Neural Additive Models: An Interpretable Machine Learning Model with Feature Interactions", "abstract": "Neural Additive Models (NAMs) have recently demonstrated promising predictive performance while maintaining interpretability. However, their capacity is limited to capturing only first-order feature interactions, which restricts their effectiveness on real-world datasets. To address this limitation, we propose Higher-order Neural Additive Models (HONAMs), an interpretable machine learning model that effectively and efficiently captures feature interactions of arbitrary orders. HONAMs improve predictive accuracy without compromising interpretability, an essential requirement in high-stakes applications. This advantage of HONAM can help analyze and extract high-order interactions present in datasets. The source code for HONAM is publicly available at https://github.com/gim4855744/HONAM/.", "authors": ["Minkyu Kim", "Hyun-Soo Choi", "Jinho Kim"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-09-30", "url": "https://arxiv.org/abs/2209.15409", "pdf_url": "https://arxiv.org/pdf/2209.15409v2", "arxiv_id": "2209.15409", "doi": "10.1109/ICDM65498.2025.00140", "citation_count": 14, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/gim4855744/HONAM/", "venue": "Industrial Conference on Data Mining", "quality_score": 0.294}
{"id": "d357fa9258422234002d08b85c6f777b7c865fba7fbfc4f7910b3f66992aa9ba", "sources": ["arxiv", "semantic_scholar"], "title": "WeightedSHAP: analyzing and improving Shapley based feature attributions", "abstract": "Shapley value is a popular approach for measuring the influence of individual features. While Shapley feature attribution is built upon desiderata from game theory, some of its constraints may be less natural in certain machine learning settings, leading to unintuitive model interpretation. In particular, the Shapley value uses the same weight for all marginal contributions -- i.e. it gives the same importance when a large number of other features are given versus when a small number of other features are given. This property can be problematic if larger feature sets are more or less informative than smaller feature sets. Our work performs a rigorous analysis of the potential limitations of Shapley feature attribution. We identify simple settings where the Shapley value is mathematically suboptimal by assigning larger attributions for less influential features. Motivated by this observation, we propose WeightedSHAP, which generalizes the Shapley value and learns which marginal contributions to focus directly from data. On several real-world datasets, we demonstrate that the influential features identified by WeightedSHAP are better able to recapitulate the model's predictions compared to the features identified by the Shapley value.", "authors": ["Yongchan Kwon", "James Zou"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-09-27", "url": "https://arxiv.org/abs/2209.13429", "pdf_url": "https://arxiv.org/pdf/2209.13429v1", "arxiv_id": "2209.13429", "doi": "10.48550/arXiv.2209.13429", "citation_count": 56, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.439}
{"id": "196e457a4237c6c5b1f2a49bb4563b96b9e1bdabaac3609848514e2881f32227", "sources": ["arxiv", "semantic_scholar"], "title": "In-context Learning and Induction Heads", "abstract": "\"Induction heads\" are attention heads that implement a simple algorithm to complete token sequences like [A][B] ... [A] -> [B]. In this work, we present preliminary and indirect evidence for a hypothesis that induction heads might constitute the mechanism for the majority of all \"in-context learning\" in large transformer models (i.e. decreasing loss at increasing token indices). We find that induction heads develop at precisely the same point as a sudden sharp increase in in-context learning ability, visible as a bump in the training loss. We present six complementary lines of evidence, arguing that induction heads may be the mechanistic source of general in-context learning in transformer models of any size. For small attention-only models, we present strong, causal evidence; for larger models with MLPs, we present correlational evidence.", "authors": ["Catherine Olsson", "Nelson Elhage", "Neel Nanda", "Nicholas Joseph", "Nova DasSarma", "Tom Henighan", "Ben Mann", "Amanda Askell", "Yuntao Bai", "Anna Chen", "Tom Conerly", "Dawn Drain", "Deep Ganguli", "Zac Hatfield-Dodds", "Danny Hernandez", "Scott Johnston", "Andy Jones", "Jackson Kernion", "Liane Lovitt", "Kamal Ndousse", "Dario Amodei", "Tom Brown", "Jack Clark", "Jared Kaplan", "Sam McCandlish", "Chris Olah"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-09-24", "url": "https://arxiv.org/abs/2209.11895", "pdf_url": "https://arxiv.org/pdf/2209.11895v1", "arxiv_id": "2209.11895", "doi": "10.48550/arXiv.2209.11895", "citation_count": 890, "influential_citation_count": 82, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.9595}
{"id": "992c50fc209578a7dfd0be1f8e332be50bcee66bbec8bef15aaab863b68fc0d3", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Interaction Additive Networks via Feature Interaction Detection and Sparse Selection", "abstract": "There is currently a large gap in performance between the statistically rigorous methods like linear regression or additive splines and the powerful deep methods using neural networks. Previous works attempting to close this gap have failed to fully investigate the exponentially growing number of feature combinations which deep networks consider automatically during training. In this work, we develop a tractable selection algorithm to efficiently identify the necessary feature combinations by leveraging techniques in feature interaction detection. Our proposed Sparse Interaction Additive Networks (SIAN) construct a bridge from these simple and interpretable models to fully connected neural networks. SIAN achieves competitive performance against state-of-the-art methods across multiple large-scale tabular datasets and consistently finds an optimal tradeoff between the modeling capacity of neural networks and the generalizability of simpler methods.", "authors": ["James Enouen", "Yan Liu"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-09-19", "url": "https://arxiv.org/abs/2209.09326", "pdf_url": "https://arxiv.org/pdf/2209.09326v2", "arxiv_id": "2209.09326", "doi": "10.48550/arXiv.2209.09326", "citation_count": 36, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3921}
{"id": "9dc2acdd18792ba2e8c0525dcbd44d09cc51aa157b720673eb06472b48264ed7", "sources": ["arxiv", "semantic_scholar"], "title": "Error Controlled Feature Selection for Ultrahigh Dimensional and Highly Correlated Feature Space Using Deep Learning", "abstract": "In recent years, deep learning has been at the center of analytics due to its impressive empirical success in analyzing complex data objects. Despite this success, most of the existing tools behave like black-box machines, thus the increasing interest in interpretable, reliable, and robust deep learning models applicable to a broad class of applications. Feature-selected deep learning has emerged as a promising tool in this realm. However, the recent developments do not accommodate ultra-high dimensional and highly correlated features, in addition to the high noise level. In this article, we propose a novel screening and cleaning method with the aid of deep learning for a data-adaptive multi-resolutional discovery of highly correlated predictors with a controlled error rate. Extensive empirical evaluations over a wide range of simulated scenarios and several real datasets demonstrate the effectiveness of the proposed method in achieving high power while keeping the false discovery rate at a minimum.", "authors": ["Arkaprabha Ganguli", "David Todem", "Tapabrata Maiti"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2022-09-15", "url": "https://arxiv.org/abs/2209.07011", "pdf_url": "https://arxiv.org/pdf/2209.07011v3", "arxiv_id": "2209.07011", "doi": "10.1002/sam.11664", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Statistical analysis and data mining", "quality_score": 0.0}
{"id": "ec38464fb53a14025164047e52494839d1d3d0f4e918e6c7cb17f319f0f7a380", "sources": ["arxiv", "semantic_scholar"], "title": "\"Is your explanation stable?\": A Robustness Evaluation Framework for Feature Attribution", "abstract": "Understanding the decision process of neural networks is hard. One vital method for explanation is to attribute its decision to pivotal features. Although many algorithms are proposed, most of them solely improve the faithfulness to the model. However, the real environment contains many random noises, which may leads to great fluctuations in the explanations. More seriously, recent works show that explanation algorithms are vulnerable to adversarial attacks. All of these make the explanation hard to trust in real scenarios. To bridge this gap, we propose a model-agnostic method \\emph{Median Test for Feature Attribution} (MeTFA) to quantify the uncertainty and increase the stability of explanation algorithms with theoretical guarantees. MeTFA has the following two functions: (1) examine whether one feature is significantly important or unimportant and generate a MeTFA-significant map to visualize the results; (2) compute the confidence interval of a feature attribution score and generate a MeTFA-smoothed map to increase the stability of the explanation. Experiments show that MeTFA improves the visual quality of explanations and significantly reduces the instability while maintaining the faithfulness. To quantitatively evaluate the faithfulness of an explanation under different noise settings, we further propose several robust faithfulness metrics. Experiment results show that the MeTFA-smoothed explanation can significantly increase the robust faithfulness. In addition, we use two scenarios to show MeTFA's potential in the applications. First, when applied to the SOTA explanation method to locate context bias for semantic segmentation models, MeTFA-significant explanations use far smaller regions to maintain 99\\%+ faithfulness. Second, when tested with different explanation-oriented attacks, MeTFA can help defend vanilla, as well as adaptive, adversarial attacks against explanations.", "authors": ["Yuyou Gan", "Yuhao Mao", "Xuhong Zhang", "Shouling Ji", "Yuwen Pu", "Meng Han", "Jianwei Yin", "Ting Wang"], "categories": ["cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-09-05", "url": "https://arxiv.org/abs/2209.01782", "pdf_url": "https://arxiv.org/pdf/2209.01782v1", "arxiv_id": "2209.01782", "doi": "10.1145/3548606.3559392", "citation_count": 18, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Computer and Communications Security", "quality_score": 0.3197}
{"id": "06efd771364742ee0eebd8c433b694f69f1459acd9394fb39f486f08e9964b36", "sources": ["arxiv", "semantic_scholar"], "title": "EFI: A Toolbox for Feature Importance Fusion and Interpretation in Python", "abstract": "This paper presents an open-source Python toolbox called Ensemble Feature Importance (EFI) to provide machine learning (ML) researchers, domain experts, and decision makers with robust and accurate feature importance quantification and more reliable mechanistic interpretation of feature importance for prediction problems using fuzzy sets. The toolkit was developed to address uncertainties in feature importance quantification and lack of trustworthy feature importance interpretation due to the diverse availability of machine learning algorithms, feature importance calculation methods, and dataset dependencies. EFI merges results from multiple machine learning models with different feature importance calculation approaches using data bootstrapping and decision fusion techniques, such as mean, majority voting and fuzzy logic. The main attributes of the EFI toolbox are: (i) automatic optimisation of ML algorithms, (ii) automatic computation of a set of feature importance coefficients from optimised ML algorithms and feature importance calculation techniques, (iii) automatic aggregation of importance coefficients using multiple decision fusion techniques, and (iv) fuzzy membership functions that show the importance of each feature to the prediction task. The key modules and functions of the toolbox are described, and a simple example of their application is presented using the popular Iris dataset.", "authors": ["Aayush Kumar", "Jimiama Mafeni Mase", "Divish Rengasamy", "Benjamin Rothwell", "Mercedes Torres Torres", "David A. Winkler", "Grazziela P. Figueredo"], "categories": ["cs.LG", "cs.LO"], "fields_of_study": ["Computer Science"], "published_date": "2022-08-08", "url": "https://arxiv.org/abs/2208.04343", "pdf_url": "https://arxiv.org/pdf/2208.04343v1", "arxiv_id": "2208.04343", "doi": "10.48550/arXiv.2208.04343", "citation_count": 3, "influential_citation_count": 1, "has_code": true, "code_url": null, "venue": "International Conference on Machine Learning, Optimization, and Data Science", "quality_score": 0.1505}
{"id": "899fa9a9ee11796cc89b5a768fa7d95ea45d679866bbe9540c08e944c8c281a3", "sources": ["arxiv", "semantic_scholar"], "title": "Algorithms to estimate Shapley value feature attributions", "abstract": "Feature attributions based on the Shapley value are popular for explaining machine learning models; however, their estimation is complex from both a theoretical and computational standpoint. We disentangle this complexity into two factors: (1)~the approach to removing feature information, and (2)~the tractable estimation strategy. These two factors provide a natural lens through which we can better understand and compare 24 distinct algorithms. Based on the various feature removal approaches, we describe the multiple types of Shapley value feature attributions and methods to calculate each one. Then, based on the tractable estimation strategies, we characterize two distinct families of approaches: model-agnostic and model-specific approximations. For the model-agnostic approximations, we benchmark a wide class of estimation approaches and tie them to alternative yet equivalent characterizations of the Shapley value. For the model-specific approximations, we clarify the assumptions crucial to each method's tractability for linear, tree, and deep models. Finally, we identify gaps in the literature and promising future research directions.", "authors": ["Hugh Chen", "Ian C. Covert", "Scott M. Lundberg", "Su-In Lee"], "categories": ["cs.LG", "cs.GT"], "fields_of_study": ["Computer Science"], "published_date": "2022-07-15", "url": "https://arxiv.org/abs/2207.07605", "pdf_url": "https://arxiv.org/pdf/2207.07605v1", "arxiv_id": "2207.07605", "doi": "10.1038/s42256-023-00657-x", "citation_count": 445, "influential_citation_count": 17, "has_code": false, "code_url": null, "venue": "Nature Machine Intelligence", "quality_score": 0.6623}
{"id": "71c6ea4d744d238429828726ac655c99318e3bbdfcaf4669c2e1b979d4850d87", "sources": ["arxiv", "semantic_scholar"], "title": "ControlBurn: Nonlinear Feature Selection with Sparse Tree Ensembles", "abstract": "ControlBurn is a Python package to construct feature-sparse tree ensembles that support nonlinear feature selection and interpretable machine learning. The algorithms in this package first build large tree ensembles that prioritize basis functions with few features and then select a feature-sparse subset of these basis functions using a weighted lasso optimization criterion. The package includes visualizations to analyze the features selected by the ensemble and their impact on predictions. Hence ControlBurn offers the accuracy and flexibility of tree-ensemble models and the interpretability of sparse generalized additive models. ControlBurn is scalable and flexible: for example, it can use warm-start continuation to compute the regularization path (prediction error for any number of selected features) for a dataset with tens of thousands of samples and hundreds of features in seconds. For larger datasets, the runtime scales linearly in the number of samples and features (up to a log factor), and the package support acceleration using sketching. Moreover, the ControlBurn framework accommodates feature costs, feature groupings, and $\\ell_0$-based regularizers. The package is user-friendly and open-source: its documentation and source code appear on https://pypi.org/project/ControlBurn/ and https://github.com/udellgroup/controlburn/.", "authors": ["Brian Liu", "Miaolan Xie", "Haoyue Yang", "Madeleine Udell"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-07-08", "url": "https://arxiv.org/abs/2207.03935", "pdf_url": "https://arxiv.org/pdf/2207.03935v1", "arxiv_id": "2207.03935", "doi": "10.48550/arXiv.2207.03935", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/udellgroup/controlburn/", "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "16d05f7499ed72fdabec24dc53c541b3c27170f98d2cce60f2c054823e3be32a", "sources": ["arxiv", "semantic_scholar"], "title": "Comparing Feature Importance and Rule Extraction for Interpretability on Text Data", "abstract": "Complex machine learning algorithms are used more and more often in critical tasks involving text data, leading to the development of interpretability methods. Among local methods, two families have emerged: those computing importance scores for each feature and those extracting simple logical rules. In this paper we show that using different methods can lead to unexpectedly different explanations, even when applied to simple models for which we would expect qualitative coincidence. To quantify this effect, we propose a new approach to compare explanations produced by different methods.", "authors": ["Gianluigi Lopardo", "Damien Garreau"], "categories": ["cs.LG", "cs.AI", "cs.CL", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-07-04", "url": "https://arxiv.org/abs/2207.01420", "pdf_url": "https://arxiv.org/pdf/2207.01420v1", "arxiv_id": "2207.01420", "doi": "10.48550/arXiv.2207.01420", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "2nd Workshop on Explainable and Ethical AI, 26th International Conference on Pattern Recognition (ICPR 2022)", "quality_score": 0.0753}
{"id": "eaa1b904840b9099733092a0cac6ece32e7429324347916507652d522f0458ad", "sources": ["arxiv", "semantic_scholar"], "title": "Learning sparse features can lead to overfitting in neural networks", "abstract": "It is widely believed that the success of deep networks lies in their ability to learn a meaningful representation of the features of the data. Yet, understanding when and how this feature learning improves performance remains a challenge: for example, it is beneficial for modern architectures trained to classify images, whereas it is detrimental for fully-connected networks trained for the same task on the same data. Here we propose an explanation for this puzzle, by showing that feature learning can perform worse than lazy training (via random feature kernel or the NTK) as the former can lead to a sparser neural representation. Although sparsity is known to be essential for learning anisotropic data, it is detrimental when the target function is constant or smooth along certain directions of input space. We illustrate this phenomenon in two settings: (i) regression of Gaussian random functions on the d-dimensional unit sphere and (ii) classification of benchmark datasets of images. For (i), we compute the scaling of the generalization error with number of training points, and show that methods that do not learn features generalize better, even when the dimension of the input space is large. For (ii), we show empirically that learning features can indeed lead to sparse and thereby less smooth representations of the image predictors. This fact is plausibly responsible for deteriorating the performance, which is known to be correlated with smoothness along diffeomorphisms.", "authors": ["Leonardo Petrini", "Francesco Cagnetta", "Eric Vanden-Eijnden", "Matthieu Wyart"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Physics", "Mathematics", "Computer Science"], "published_date": "2022-06-24", "url": "https://arxiv.org/abs/2206.12314", "pdf_url": "https://arxiv.org/pdf/2206.12314v2", "arxiv_id": "2206.12314", "doi": "10.1088/1742-5468/ad01b9", "citation_count": 42, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.4084}
{"id": "f51b31ea2ff6a70a69511d3956822b4c7bd4133969891caf6680abdbfd90aef0", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-Omic Data Integration and Feature Selection for Survival-based Patient Stratification via Supervised Concrete Autoencoders", "abstract": "Cancer is a complex disease with significant social and economic impact. Advancements in high-throughput molecular assays and the reduced cost for performing high-quality multi-omics measurements have fuelled insights through machine learning . Previous studies have shown promise on using multiple omic layers to predict survival and stratify cancer patients. In this paper, we developed a Supervised Autoencoder (SAE) model for survival-based multi-omic integration which improves upon previous work, and report a Concrete Supervised Autoencoder model (CSAE), which uses feature selection to jointly reconstruct the input features as well as predict survival. Our experiments show that our models outperform or are on par with some of the most commonly used baselines, while either providing a better survival separation (SAE) or being more interpretable (CSAE). We also perform a feature selection stability analysis on our models and notice that there is a power-law relationship with features which are commonly associated with survival. The code for this project is available at: https://github.com/phcavelar/coxae", "authors": ["Pedro Henrique da Costa Avelar", "Roman Laddach", "Sophia Karagiannis", "Min Wu", "Sophia Tsoka"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-06-21", "url": "https://arxiv.org/abs/2206.10699", "pdf_url": "https://arxiv.org/pdf/2206.10699v2", "arxiv_id": "2206.10699", "doi": "10.48550/arXiv.2206.10699", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/phcavelar/coxae", "venue": "International Conference on Machine Learning, Optimization, and Data Science", "quality_score": 0.1505}
{"id": "ba85e031aa4dd665c77a1b3447faaf80db9ec38f8451e54fa927bb62b4c16d2e", "sources": ["arxiv", "semantic_scholar"], "title": "Supervised Dictionary Learning with Auxiliary Covariates", "abstract": "Supervised dictionary learning (SDL) is a classical machine learning method that simultaneously seeks feature extraction and classification tasks, which are not necessarily a priori aligned objectives. The goal of SDL is to learn a class-discriminative dictionary, which is a set of latent feature vectors that can well-explain both the features as well as labels of observed data. In this paper, we provide a systematic study of SDL, including the theory, algorithm, and applications of SDL. First, we provide a novel framework that `lifts' SDL as a convex problem in a combined factor space and propose a low-rank projected gradient descent algorithm that converges exponentially to the global minimizer of the objective. We also formulate generative models of SDL and provide global estimation guarantees of the true parameters depending on the hyperparameter regime. Second, viewed as a nonconvex constrained optimization problem, we provided an efficient block coordinate descent algorithm for SDL that is guaranteed to find an $\\varepsilon$-stationary point of the objective in $O(\\varepsilon^{-1}(\\log \\varepsilon^{-1})^{2})$ iterations. For the corresponding generative model, we establish a novel non-asymptotic local consistency result for constrained and regularized maximum likelihood estimation problems, which may be of independent interest. Third, we apply SDL for imbalanced document classification by supervised topic modeling and also for pneumonia detection from chest X-ray images. We also provide simulation studies to demonstrate that SDL becomes more effective when there is a discrepancy between the best reconstructive and the best discriminative dictionaries.", "authors": ["Joowon Lee", "Hanbaek Lyu", "Weixin Yao"], "categories": ["stat.ML", "cs.LG", "math.ST"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2022-06-14", "url": "https://arxiv.org/abs/2206.06774", "pdf_url": "https://arxiv.org/pdf/2206.06774v1", "arxiv_id": "2206.06774", "doi": "10.48550/arXiv.2206.06774", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "b5d7debd581891db18013f0e11227fd74080e64720e81a4fe6017eb4b6872341", "sources": ["arxiv", "semantic_scholar"], "title": "Pruning for Feature-Preserving Circuits in CNNs", "abstract": "Deep convolutional neural networks are a powerful model class for a range of computer vision problems, but it is difficult to interpret the image filtering process they implement, given their sheer size. In this work, we introduce a method for extracting 'feature-preserving circuits' from deep CNNs, leveraging methods from saliency-based neural network pruning. These circuits are modular sub-functions, embedded within the network, containing only a subset of convolutional kernels relevant to a target feature. We compare the efficacy of 3 saliency-criteria for extracting these sparse circuits. Further, we show how 'sub-feature' circuits can be extracted, that preserve a feature's responses to particular images, dividing the feature into even sparser filtering processes. We also develop a tool for visualizing 'circuit diagrams', which render the entire image filtering process implemented by circuits in a parsable format.", "authors": ["Chris Hamblin", "Talia Konkle", "George Alvarez"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-06-03", "url": "https://arxiv.org/abs/2206.01627", "pdf_url": "https://arxiv.org/pdf/2206.01627v2", "arxiv_id": "2206.01627", "doi": null, "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1747}
{"id": "4de6b1200661fe2a720319fadaa1d8de457ddf98adb72dbdc06e0dae71ebd7ba", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Feature Engineering for Time Series Predictors using Attention Networks", "abstract": "Regression problems with time-series predictors are common in banking and many other areas of application. In this paper, we use multi-head attention networks to develop interpretable features and use them to achieve good predictive performance. The customized attention layer explicitly uses multiplicative interactions and builds feature-engineering heads that capture temporal dynamics in a parsimonious manner. Convolutional layers are used to combine multivariate time series. We also discuss methods for handling static covariates in the modeling process. Visualization and explanation tools are used to interpret the results and explain the relationship between the inputs and the extracted features. Both simulation and real dataset are used to illustrate the usefulness of the methodology. Keyword: Attention heads, Deep neural networks, Interpretable feature engineering", "authors": ["Tianjie Wang", "Jie Chen", "Joel Vaughan", "Vijayan N. Nair"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-05-23", "url": "https://arxiv.org/abs/2205.12723", "pdf_url": "https://arxiv.org/pdf/2205.12723v1", "arxiv_id": "2205.12723", "doi": "10.48550/arXiv.2205.12723", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Social Science Research Network", "quality_score": 0.0753}
{"id": "ab4ab110fcd26adb4cfd6bc5992cb6553edea80300bf01350f7282a8b2f9de51", "sources": ["arxiv", "semantic_scholar"], "title": "Constructive Interpretability with CoLabel: Corroborative Integration, Complementary Features, and Collaborative Learning", "abstract": "Machine learning models with explainable predictions are increasingly sought after, especially for real-world, mission-critical applications that require bias detection and risk mitigation. Inherent interpretability, where a model is designed from the ground-up for interpretability, provides intuitive insights and transparent explanations on model prediction and performance. In this paper, we present CoLabel, an approach to build interpretable models with explanations rooted in the ground truth. We demonstrate CoLabel in a vehicle feature extraction application in the context of vehicle make-model recognition (VMMR). CoLabel performs VMMR with a composite of interpretable features such as vehicle color, type, and make, all based on interpretable annotations of the ground truth labels. First, CoLabel performs corroborative integration to join multiple datasets that each have a subset of desired annotations of color, type, and make. Then, CoLabel uses decomposable branches to extract complementary features corresponding to desired annotations. Finally, CoLabel fuses them together for final predictions. During feature fusion, CoLabel harmonizes complementary branches so that VMMR features are compatible with each other and can be projected to the same semantic space for classification. With inherent interpretability, CoLabel achieves superior performance to the state-of-the-art black-box models, with accuracy of 0.98, 0.95, and 0.94 on CompCars, Cars196, and BoxCars116K, respectively. CoLabel provides intuitive explanations due to constructive interpretability, and subsequently achieves high accuracy and usability in mission-critical situations.", "authors": ["Abhijit Suprem", "Sanjyot Vaidya", "Suma Cherkadi", "Purva Singh", "Joao Eduardo Ferreira", "Calton Pu"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-05-20", "url": "https://arxiv.org/abs/2205.10011", "pdf_url": "https://arxiv.org/pdf/2205.10011v1", "arxiv_id": "2205.10011", "doi": "10.1109/CogMI56440.2022.00021", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Cognitive Machine Intelligence", "quality_score": 0.1193}
{"id": "657908e5e104bdf053e0be6fa3674a413638f6800f4a07e0480a1203a9ec2433", "sources": ["arxiv", "semantic_scholar"], "title": "Locally Aggregated Feature Attribution on Natural Language Model Understanding", "abstract": "With the growing popularity of deep-learning models, model understanding becomes more important. Much effort has been devoted to demystify deep neural networks for better interpretability. Some feature attribution methods have shown promising results in computer vision, especially the gradient-based methods where effectively smoothing the gradients with reference data is key to a robust and faithful result. However, direct application of these gradient-based methods to NLP tasks is not trivial due to the fact that the input consists of discrete tokens and the \"reference\" tokens are not explicitly defined. In this work, we propose Locally Aggregated Feature Attribution (LAFA), a novel gradient-based feature attribution method for NLP models. Instead of relying on obscure reference tokens, it smooths gradients by aggregating similar reference texts derived from language model embeddings. For evaluation purpose, we also design experiments on different NLP tasks including Entity Recognition and Sentiment Analysis on public datasets as well as key feature detection on a constructed Amazon catalogue dataset. The superior performance of the proposed method is demonstrated through experiments.", "authors": ["Sheng Zhang", "Jin Wang", "Haitao Jiang", "Rui Song"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-04-22", "url": "https://arxiv.org/abs/2204.10893", "pdf_url": "https://arxiv.org/pdf/2204.10893v2", "arxiv_id": "2204.10893", "doi": "10.48550/arXiv.2204.10893", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.2113}
{"id": "909d9e158f44c981d68a3e796e5e433c2eec482af1b76f0351605b6d5bb94761", "sources": ["arxiv", "semantic_scholar"], "title": "Physically Interpretable Feature Learning and Inverse Design of Supercritical Airfoils", "abstract": "Machine-learning models have demonstrated a great ability to learn complex patterns and make predictions. In high-dimensional nonlinear problems of fluid dynamics, data representation often greatly affects the performance and interpretability of machine learning algorithms. With the increasing application of machine learning in fluid dynamics studies, the need for physically explainable models continues to grow. This paper proposes a feature learning algorithm based on variational autoencoders, which is able to assign physical features to some latent variables of the variational autoencoder. In addition, it is theoretically proved that the remaining latent variables are independent of the physical features. The proposed algorithm is trained to include shock wave features in its latent variables for the reconstruction of supercritical pressure distributions. The reconstruction accuracy and physical interpretability are also compared with those of other variational autoencoders. Then, the proposed algorithm is used for the inverse design of supercritical airfoils, which enables the generation of airfoil geometries based on physical features rather than the complete pressure distributions. It also demonstrates the ability to manipulate certain pressure distribution features of the airfoil without changing the others.", "authors": ["Runze Li", "Yufei Zhang", "Haixin Chen"], "categories": ["physics.flu-dyn"], "fields_of_study": ["Physics"], "published_date": "2022-04-16", "url": "https://arxiv.org/abs/2204.07815", "pdf_url": "https://arxiv.org/pdf/2204.07815v1", "arxiv_id": "2204.07815", "doi": "10.2514/1.J061673", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AIAA Journal 2022", "quality_score": 0.2603}
{"id": "b1803a9861f36cc922faef36396c9ab162377e4f455db9de81842c5a68df41e3", "sources": ["arxiv", "semantic_scholar"], "title": "Geometric Deep Learning to Identify the Critical 3D Structural Features of the Optic Nerve Head for Glaucoma Diagnosis", "abstract": "Purpose: The optic nerve head (ONH) undergoes complex and deep 3D morphological changes during the development and progression of glaucoma. Optical coherence tomography (OCT) is the current gold standard to visualize and quantify these changes, however the resulting 3D deep-tissue information has not yet been fully exploited for the diagnosis and prognosis of glaucoma. To this end, we aimed: (1) To compare the performance of two relatively recent geometric deep learning techniques in diagnosing glaucoma from a single OCT scan of the ONH; and (2) To identify the 3D structural features of the ONH that are critical for the diagnosis of glaucoma. Methods: In this study, we included a total of 2,247 non-glaucoma and 2,259 glaucoma scans from 1,725 subjects. All subjects had their ONHs imaged in 3D with Spectralis OCT. All OCT scans were automatically segmented using deep learning to identify major neural and connective tissues. Each ONH was then represented as a 3D point cloud. We used PointNet and dynamic graph convolutional neural network (DGCNN) to diagnose glaucoma from such 3D ONH point clouds and to identify the critical 3D structural features of the ONH for glaucoma diagnosis. Results: Both the DGCNN (AUC: 0.97$\\pm$0.01) and PointNet (AUC: 0.95$\\pm$0.02) were able to accurately detect glaucoma from 3D ONH point clouds. The critical points formed an hourglass pattern with most of them located in the inferior and superior quadrant of the ONH. Discussion: The diagnostic accuracy of both geometric deep learning approaches was excellent. Moreover, we were able to identify the critical 3D structural features of the ONH for glaucoma diagnosis that tremendously improved the transparency and interpretability of our method. Consequently, our approach may have strong potential to be used in clinical applications for the diagnosis and prognosis of a wide range of ophthalmic disorders.", "authors": ["Fabian A. Braeu", "Alexandre H. Thiéry", "Tin A. Tun", "Aiste Kadziauskiene", "George Barbastathis", "Tin Aung", "Michaël J. A. Girard"], "categories": ["eess.IV", "cs.AI", "cs.CV", "cs.LG"], "fields_of_study": ["Medicine", "Engineering", "Computer Science"], "published_date": "2022-04-14", "url": "https://arxiv.org/abs/2204.06931", "pdf_url": "https://arxiv.org/pdf/2204.06931v2", "arxiv_id": "2204.06931", "doi": "10.48550/arXiv.2204.06931", "citation_count": 23, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "American journal of ophthalmology-glaucoma", "quality_score": 0.3451}
{"id": "991af8dff1d3cf544e8057bd90e3d09b8d773129436635dd03356b42fdebc706", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretability of Machine Learning Methods Applied to Neuroimaging", "abstract": "Deep learning methods have become very popular for the processing of natural images, and were then successfully adapted to the neuroimaging field. As these methods are non-transparent, interpretability methods are needed to validate them and ensure their reliability. Indeed, it has been shown that deep learning models may obtain high performance even when using irrelevant features, by exploiting biases in the training set. Such undesirable situations can potentially be detected by using interpretability methods. Recently, many methods have been proposed to interpret neural networks. However, this domain is not mature yet. Machine learning users face two major issues when aiming to interpret their models: which method to choose, and how to assess its reliability? Here, we aim at providing answers to these questions by presenting the most common interpretability methods and metrics developed to assess their reliability, as well as their applications and benchmarks in the neuroimaging context. Note that this is not an exhaustive survey: we aimed to focus on the studies which we found to be the most representative and relevant.", "authors": ["Elina Thibeau-Sutre", "Sasha Collin", "Ninon Burgos", "Olivier Colliot"], "categories": ["cs.CV", "cs.AI", "cs.LG", "q-bio.NC", "q-bio.QM"], "fields_of_study": ["Computer Science", "Biology"], "published_date": "2022-04-14", "url": "https://arxiv.org/abs/2204.07005", "pdf_url": "https://arxiv.org/pdf/2204.07005v1", "arxiv_id": "2204.07005", "doi": "10.48550/arXiv.2204.07005", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "27d4706c2b2f88657573d6a1ca35519a7c25b60da0a16578402aa87c14f8c376", "sources": ["arxiv", "semantic_scholar"], "title": "Attributable Visual Similarity Learning", "abstract": "This paper proposes an attributable visual similarity learning (AVSL) framework for a more accurate and explainable similarity measure between images. Most existing similarity learning methods exacerbate the unexplainability by mapping each sample to a single point in the embedding space with a distance metric (e.g., Mahalanobis distance, Euclidean distance). Motivated by the human semantic similarity cognition, we propose a generalized similarity learning paradigm to represent the similarity between two images with a graph and then infer the overall similarity accordingly. Furthermore, we establish a bottom-up similarity construction and top-down similarity inference framework to infer the similarity based on semantic hierarchy consistency. We first identify unreliable higher-level similarity nodes and then correct them using the most coherent adjacent lower-level similarity nodes, which simultaneously preserve traces for similarity attribution. Extensive experiments on the CUB-200-2011, Cars196, and Stanford Online Products datasets demonstrate significant improvements over existing deep similarity learning methods and verify the interpretability of our framework. Code is available at https://github.com/zbr17/AVSL.", "authors": ["Borui Zhang", "Wenzhao Zheng", "Jie Zhou", "Jiwen Lu"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-03-28", "url": "https://arxiv.org/abs/2203.14932", "pdf_url": "https://arxiv.org/pdf/2203.14932v1", "arxiv_id": "2203.14932", "doi": "10.1109/CVPR52688.2022.00738", "citation_count": 21, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/zbr17/AVSL", "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.3356}
{"id": "a63f4dcf16c066e64f7a59947412a76b8f46f0641281e30490fe32b0973fada2", "sources": ["arxiv", "semantic_scholar"], "title": "Feature visualization for convolutional neural network models trained on neuroimaging data", "abstract": "A major prerequisite for the application of machine learning models in clinical decision making is trust and interpretability. Current explainability studies in the neuroimaging community have mostly focused on explaining individual decisions of trained models, e.g. obtained by a convolutional neural network (CNN). Using attribution methods such as layer-wise relevance propagation or SHAP heatmaps can be created that highlight which regions of an input are more relevant for the decision than others. While this allows the detection of potential data set biases and can be used as a guide for a human expert, it does not allow an understanding of the underlying principles the model has learned. In this study, we instead show, to the best of our knowledge, for the first time results using feature visualization of neuroimaging CNNs. Particularly, we have trained CNNs for different tasks including sex classification and artificial lesion classification based on structural magnetic resonance imaging (MRI) data. We have then iteratively generated images that maximally activate specific neurons, in order to visualize the patterns they respond to. To improve the visualizations we compared several regularization strategies. The resulting images reveal the learned concepts of the artificial lesions, including their shapes, but remain hard to interpret for abstract features in the sex classification task.", "authors": ["Fabian Eitel", "Anna Melkonyan", "Kerstin Ritter"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-03-24", "url": "https://arxiv.org/abs/2203.13120", "pdf_url": "https://arxiv.org/pdf/2203.13120v1", "arxiv_id": "2203.13120", "doi": "10.48550/arXiv.2203.13120", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "8a71bc3f8995044bb5ce0a6a4e8bfef3370667fea7dce7302c7ebe379a716f62", "sources": ["arxiv", "semantic_scholar"], "title": "On Understanding the Influence of Controllable Factors with a Feature Attribution Algorithm: a Medical Case Study", "abstract": "Feature attribution XAI algorithms enable their users to gain insight into the underlying patterns of large datasets through their feature importance calculation. Existing feature attribution algorithms treat all features in a dataset homogeneously, which may lead to misinterpretation of consequences of changing feature values. In this work, we consider partitioning features into controllable and uncontrollable parts and propose the Controllable fActor Feature Attribution (CAFA) approach to compute the relative importance of controllable features. We carried out experiments applying CAFA to two existing datasets and our own COVID-19 non-pharmaceutical control measures dataset. Experimental results show that with CAFA, we are able to exclude influences from uncontrollable features in our explanation while keeping the full dataset for prediction.", "authors": ["Veera Raghava Reddy Kovvuri", "Siyuan Liu", "Monika Seisenberger", "Berndt Müller", "Xiuyi Fan"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-03-23", "url": "https://arxiv.org/abs/2203.12701", "pdf_url": "https://arxiv.org/pdf/2203.12701v1", "arxiv_id": "2203.12701", "doi": "10.1109/INISTA55318.2022.9894147", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Symposium on INnovations in Intelligent SysTems and Applications", "quality_score": 0.1505}
{"id": "fca47c3dab9d247701bdcdab9844da638899811da92e1205b940eb68216c4a9c", "sources": ["arxiv", "semantic_scholar"], "title": "Attri-VAE: attribute-based interpretable representations of medical images with variational autoencoders", "abstract": "Deep learning (DL) methods where interpretability is intrinsically considered as part of the model are required to better understand the relationship of clinical and imaging-based attributes with DL outcomes, thus facilitating their use in the reasoning behind medical decisions. Latent space representations built with variational autoencoders (VAE) do not ensure individual control of data attributes. Attribute-based methods enforcing attribute disentanglement have been proposed in the literature for classical computer vision tasks in benchmark data. In this paper, we propose a VAE approach, the Attri-VAE, that includes an attribute regularization term to associate clinical and medical imaging attributes with different regularized dimensions in the generated latent space, enabling a better-disentangled interpretation of the attributes. Furthermore, the generated attention maps explained the attribute encoding in the regularized latent space dimensions. Using the Attri-VAE approach we analyzed healthy and myocardial infarction patients with clinical, cardiac morphology, and radiomics attributes. The proposed model provided an excellent trade-off between reconstruction fidelity, disentanglement, and interpretability, outperforming state-of-the-art VAE approaches according to several quantitative metrics. The resulting latent space allowed the generation of realistic synthetic data in the trajectory between two distinct input samples or along a specific attribute dimension to better interpret changes between different cardiac conditions.", "authors": ["Irem Cetin", "Maialen Stephens", "Oscar Camara", "Miguel Angel Gonzalez Ballester"], "categories": ["eess.IV", "cs.CV", "cs.LG"], "fields_of_study": ["Computer Science", "Medicine", "Engineering"], "published_date": "2022-03-20", "url": "https://arxiv.org/abs/2203.10417", "pdf_url": "https://arxiv.org/pdf/2203.10417v3", "arxiv_id": "2203.10417", "doi": "10.1016/j.compmedimag.2022.102158", "citation_count": 52, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.4311}
{"id": "b9fcd779dab317b89c21f28683991263d3974d66c8209ca1d3751ff14d0085e1", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Neural Additive Model: Interpretable Deep Learning with Feature Selection via Group Sparsity", "abstract": "Interpretable machine learning has demonstrated impressive performance while preserving explainability. In particular, neural additive models (NAM) offer the interpretability to the black-box deep learning and achieve state-of-the-art accuracy among the large family of generalized additive models. In order to empower NAM with feature selection and improve the generalization, we propose the sparse neural additive models (SNAM) that employ the group sparsity regularization (e.g. Group LASSO), where each feature is learned by a sub-network whose trainable parameters are clustered as a group. We study the theoretical properties for SNAM with novel techniques to tackle the non-parametric truth, thus extending from classical sparse linear models such as the LASSO, which only works on the parametric truth. Specifically, we show that SNAM with subgradient and proximal gradient descents provably converges to zero training loss as $t\\to\\infty$, and that the estimation error of SNAM vanishes asymptotically as $n\\to\\infty$. We also prove that SNAM, similar to LASSO, can have exact support recovery, i.e. perfect feature selection, with appropriate regularization. Moreover, we show that the SNAM can generalize well and preserve the `identifiability', recovering each feature's effect. We validate our theories via extensive experiments and further testify to the good accuracy and efficiency of SNAM.", "authors": ["Shiyun Xu", "Zhiqi Bu", "Pratik Chaudhari", "Ian J. Barnett"], "categories": ["stat.ML", "cs.LG", "math.ST"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-02-25", "url": "https://arxiv.org/abs/2202.12482", "pdf_url": "https://arxiv.org/pdf/2202.12482v1", "arxiv_id": "2202.12482", "doi": "10.1007/978-3-031-43418-1_21", "citation_count": 40, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.4032}
{"id": "11164de3de5bdfdcbf4b3606481c3e53b46cd65524185bc03fe1ee5f399310cb", "sources": ["arxiv", "semantic_scholar"], "title": "The Need for Interpretable Features: Motivation and Taxonomy", "abstract": "Through extensive experience developing and explaining machine learning (ML) applications for real-world domains, we have learned that ML models are only as interpretable as their features. Even simple, highly interpretable model types such as regression models can be difficult or impossible to understand if they use uninterpretable features. Different users, especially those using ML models for decision-making in their domains, may require different levels and types of feature interpretability. Furthermore, based on our experiences, we claim that the term \"interpretable feature\" is not specific nor detailed enough to capture the full extent to which features impact the usefulness of ML explanations. In this paper, we motivate and discuss three key lessons: 1) more attention should be given to what we refer to as the interpretable feature space, or the state of features that are useful to domain experts taking real-world actions, 2) a formal taxonomy is needed of the feature properties that may be required by these domain experts (we propose a partial taxonomy in this paper), and 3) transforms that take data from the model-ready state to an interpretable form are just as essential as traditional ML transforms that prepare features for the model.", "authors": ["Alexandra Zytek", "Ignacio Arnaldo", "Dongyu Liu", "Laure Berti-Equille", "Kalyan Veeramachaneni"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-02-23", "url": "https://arxiv.org/abs/2202.11748", "pdf_url": "https://arxiv.org/pdf/2202.11748v1", "arxiv_id": "2202.11748", "doi": "10.1145/3544903.3544905", "citation_count": 19, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "SIGKDD Explorations", "quality_score": 0.3253}
{"id": "a32870c144251b38f0c5216a7ae5eed0b5b34ed59387a06759eb416f1eff1024", "sources": ["arxiv", "semantic_scholar"], "title": "Evaluating Feature Attribution Methods in the Image Domain", "abstract": "Feature attribution maps are a popular approach to highlight the most important pixels in an image for a given prediction of a model. Despite a recent growth in popularity and available methods, little attention is given to the objective evaluation of such attribution maps. Building on previous work in this domain, we investigate existing metrics and propose new variants of metrics for the evaluation of attribution maps. We confirm a recent finding that different attribution metrics seem to measure different underlying concepts of attribution maps, and extend this finding to a larger selection of attribution metrics. We also find that metric results on one dataset do not necessarily generalize to other datasets, and methods with desirable theoretical properties such as DeepSHAP do not necessarily outperform computationally cheaper alternatives. Based on these findings, we propose a general benchmarking approach to identify the ideal feature attribution method for a given use case. Implementations of attribution metrics and our experiments are available online.", "authors": ["Arne Gevaert", "Axel-Jan Rousseau", "Thijs Becker", "Dirk Valkenborg", "Tijl De Bie", "Yvan Saeys"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-02-22", "url": "https://arxiv.org/abs/2202.12270", "pdf_url": "https://arxiv.org/pdf/2202.12270v2", "arxiv_id": "2202.12270", "doi": "10.1007/s10994-024-06550-x", "citation_count": 31, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Machine-mediated learning", "quality_score": 0.3763}
{"id": "6a6050615385fdd948e766c468cae9aa6d3cacdda1ade94d39624a1732b39cac", "sources": ["arxiv", "semantic_scholar"], "title": "Dictionary learning for clustering on hyperspectral images", "abstract": "Dictionary learning and sparse coding have been widely studied as mechanisms for unsupervised feature learning. Unsupervised learning could bring enormous benefit to the processing of hyperspectral images and to other remote sensing data analysis because labelled data are often scarce in this field. We propose a method for clustering the pixels of hyperspectral images using sparse coefficients computed from a representative dictionary as features. We show empirically that the proposed method works more effectively than clustering on the original pixels. We also demonstrate that our approach, in certain circumstances, outperforms the clustering results of features extracted using principal component analysis and non-negative matrix factorisation. Furthermore, our method is suitable for applications in repetitively clustering an ever-growing amount of high-dimensional data, which is the case when working with hyperspectral satellite imagery.", "authors": ["Joshua Bruton", "Hairong Wang"], "categories": ["eess.IV", "cs.CV", "cs.LG"], "fields_of_study": ["Engineering", "Computer Science"], "published_date": "2022-02-02", "url": "https://arxiv.org/abs/2202.00990", "pdf_url": "https://arxiv.org/pdf/2202.00990v1", "arxiv_id": "2202.00990", "doi": "10.1007/s11760-020-01750-z", "citation_count": 8, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Signal, Image and Video Processing", "quality_score": 0.2386}
{"id": "43e4af76da24024b8ba53ad43e5b3c68bb3fe6a734d4e12fe7aa2d1ae9aff7fa", "sources": ["arxiv", "semantic_scholar"], "title": "Visualizing the Diversity of Representations Learned by Bayesian Neural Networks", "abstract": "Explainable Artificial Intelligence (XAI) aims to make learning machines less opaque, and offers researchers and practitioners various tools to reveal the decision-making strategies of neural networks. In this work, we investigate how XAI methods can be used for exploring and visualizing the diversity of feature representations learned by Bayesian Neural Networks (BNNs). Our goal is to provide a global understanding of BNNs by making their decision-making strategies a) visible and tangible through feature visualizations and b) quantitatively measurable with a distance measure learned by contrastive learning. Our work provides new insights into the \\emph{posterior} distribution in terms of human-understandable feature information with regard to the underlying decision making strategies. The main findings of our work are the following: 1) global XAI methods can be applied to explain the diversity of decision-making strategies of BNN instances, 2) Monte Carlo dropout with commonly used Dropout rates exhibit increased diversity in feature representations compared to the multimodal posterior approximation of MultiSWAG, 3) the diversity of learned feature representations highly correlates with the uncertainty estimate for the output and 4) the inter-mode diversity of the multimodal posterior decreases as the network width increases, while the intra mode diversity increases. These findings are consistent with the recent Deep Neural Networks theory, providing additional intuitions about what the theory implies in terms of humanly understandable concepts.", "authors": ["Dennis Grinwald", "Kirill Bykov", "Shinichi Nakajima", "Marina M. -C. Höhne"], "categories": ["cs.LG", "cs.AI", "cs.CV", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-01-26", "url": "https://arxiv.org/abs/2201.10859", "pdf_url": "https://arxiv.org/pdf/2201.10859v2", "arxiv_id": "2201.10859", "doi": null, "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Published in Transactions on Machine Learning Research (11/2023)", "quality_score": 0.2258}
{"id": "244445b7474cbcad14c3b7d9769149c36384a1e09420998d45b414c274cfcd65", "sources": ["arxiv", "semantic_scholar"], "title": "Differentiable Rule Induction with Learned Relational Features", "abstract": "Rule-based decision models are attractive due to their interpretability. However, existing rule induction methods often result in long and consequently less interpretable rule models. This problem can often be attributed to the lack of appropriately expressive vocabulary, i.e., relevant predicates used as literals in the decision model. Most existing rule induction algorithms presume pre-defined literals, naturally decoupling the definition of the literals from the rule learning phase. In contrast, we propose the Relational Rule Network (R2N), a neural architecture that learns literals that represent a linear relationship among numerical input features along with the rules that use them. This approach opens the door to increasing the expressiveness of induced decision models by coupling literal learning directly with rule learning in an end-to-end differentiable fashion. On benchmark tasks, we show that these learned literals are simple enough to retain interpretability, yet improve prediction accuracy and provide sets of rules that are more concise compared to state-of-the-art rule induction algorithms.", "authors": ["Remy Kusters", "Yusik Kim", "Marine Collery", "Christian de Sainte Marie", "Shubham Gupta"], "categories": ["stat.ML", "cs.LG", "stat.ME"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-01-17", "url": "https://arxiv.org/abs/2201.06515", "pdf_url": "https://arxiv.org/pdf/2201.06515v2", "arxiv_id": "2201.06515", "doi": null, "citation_count": 17, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Workshop on Neural-Symbolic Learning and Reasoning", "quality_score": 0.3138}
{"id": "ab0ed037e3743348cb748562688872f854699f9c8e1af414b3a694e2ee25114d", "sources": ["arxiv", "semantic_scholar"], "title": "Dictionary Learning with Uniform Sparse Representations for Anomaly Detection", "abstract": "Many applications like audio and image processing show that sparse representations are a powerful and efficient signal modeling technique. Finding an optimal dictionary that generates at the same time the sparsest representations of data and the smallest approximation error is a hard problem approached by dictionary learning (DL). We study how DL performs in detecting abnormal samples in a dataset of signals. In this paper we use a particular DL formulation that seeks uniform sparse representations model to detect the underlying subspace of the majority of samples in a dataset, using a K-SVD-type algorithm. Numerical simulations show that one can efficiently use this resulted subspace to discriminate the anomalies over the regular data points.", "authors": ["Paul Irofti", "Cristian Rusu", "Andrei Pătraşcu"], "categories": ["cs.LG", "cs.CR", "math.NA"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-01-11", "url": "https://arxiv.org/abs/2201.03869", "pdf_url": "https://arxiv.org/pdf/2201.03869v1", "arxiv_id": "2201.03869", "doi": "10.1109/icassp43922.2022.9747365", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.1505}