Datasets:

fineset-io
/

mechanistic-interpretability-papers

id string	sources list	title string	abstract string	authors list	categories list	fields_of_study list	published_date timestamp[s]	url string	pdf_url string	arxiv_id string	doi string	citation_count int64	influential_citation_count int64	has_code bool	code_url string	venue string	quality_score float64
32fc1a0f570981f59187f904f8f84166aa7e94ffbeb6ef2424bb0f2342484e28	[ "arxiv" ]	Where Computation Lives Inside TabPFN: Causal Localisation of Attention Head Function	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: on...	[ "Atharva Gupta", "Dhruv Kumar", "Murari Mandal", "Saurabh Deshpande" ]	[ "cs.LG" ]	[]	2026-06-11T00:00:00	https://arxiv.org/abs/2606.12917	https://arxiv.org/pdf/2606.12917v1	2606.12917	null	0	0	false	null	null	0
3e83e52dd83d1669e01dc3a44bb3f6659d7b1fade95c9c363a66f3ce64d4c987	[ "arxiv", "semantic_scholar" ]	Unstable Features, Reproducible Subspaces: Understanding Seed Dependence in Sparse Autoencoders	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 reapp...	[ "Gleb Gerasimov", "Timofei Rusalev", "Nikita Balagansky", "Daniil Laptev", "Vadim Kurochkin", "Daniil Gavrilov" ]	[ "cs.LG", "cs.AI", "cs.CL" ]	[ "Computer Science" ]	2026-06-10T00:00:00	https://arxiv.org/abs/2606.12138	https://arxiv.org/pdf/2606.12138v1	2606.12138	null	0	0	false	null	null	0
c49810ee97d0a51419cf749c4b4fa9f086d8e584ea6e8b149b965b62160da914	[ "arxiv", "semantic_scholar" ]	Interpretable enzyme function prediction via sparse autoencoder features of ESMC across the microbial protein universe	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 ...	[ "Yue Hu", "Wanyu Cheng", "Junqing Wang", "Yingchao Liu" ]	[ "q-bio.QM" ]	[ "Biology" ]	2026-06-10T00:00:00	https://arxiv.org/abs/2606.12209	https://arxiv.org/pdf/2606.12209v1	2606.12209	null	0	0	false	null	null	0
f123ecf491e9d64e5e781b2cc1ab3626d3463e3f27a9c0378fdb000cb42a35ea	[ "arxiv", "semantic_scholar" ]	ICA Lens: Interpreting Language Models Without Training Another Dictionary	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 diction...	[ "Sida Liu", "Feijiang Han" ]	[ "cs.LG", "cs.AI", "cs.CL" ]	[ "Computer Science" ]	2026-06-10T00:00:00	https://arxiv.org/abs/2606.11722	https://arxiv.org/pdf/2606.11722v1	2606.11722	null	0	0	false	null	null	0
33614241ba5e32d77b7510d7e0ad7bb83883c79cda6aa6c76e4c33f8f7ba4446	[ "arxiv", "semantic_scholar" ]	XtrAIn: Training-Guided Occlusion for Feature Attribution	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-dis...	[ "Thodoris Lymperopoulos", "Ioannis Kakogeorgiou", "Denia Kanellopoulou" ]	[ "cs.LG", "cs.CV" ]	[ "Computer Science" ]	2026-06-09T00:00:00	https://arxiv.org/abs/2606.10877	https://arxiv.org/pdf/2606.10877v1	2606.10877	null	0	0	false	null	null	0
c632c084b375be628d688cc2a9f5042e341eea33091fcf42619da57449942862	[ "arxiv", "semantic_scholar" ]	VFUSE: Virulent Feature Understanding with Sparse autoEncoders	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 dif...	[ "Michael Yu", "Matthew L. Olson" ]	[ "cs.LG", "cs.AI", "q-bio.QM" ]	[ "Computer Science", "Biology" ]	2026-06-08T00:00:00	https://arxiv.org/abs/2606.10080	https://arxiv.org/pdf/2606.10080v1	2606.10080	null	0	0	false	null	null	0
b1b7bbbc93e130c61c887b5281f4bb495b4a4c6064f1d340d481df8a6d7ff8c7	[ "arxiv", "semantic_scholar" ]	Interactions Between Crosscoder Features: A Compact Proofs Perspective	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 ...	[ "Dmitry Manning-Coe", "Thomas Read", "Anna Soligo", "Oliver Clive-Griffin", "Chun-Hei Yip", "Rajashree Agrawal", "Jason Gross" ]	[ "cs.LG", "cs.AI" ]	[ "Computer Science" ]	2026-06-08T00:00:00	https://arxiv.org/abs/2606.09940	https://arxiv.org/pdf/2606.09940v1	2606.09940	null	0	0	true	https://github.com/chainik1125/crosscoders-feature-interactions/tree/arxiv	null	0
b27ce66213a23b46bc48ba7028ec9c506a3ececab4e0e5bf0d112011d8e9c48e	[ "arxiv", "semantic_scholar" ]	Interpreting and Steering a Text-to-Speech Language Model with Sparse Autoencoders	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 a...	[ "Nikita Koriagin", "Georgii Aparin", "Nikita Balagansky", "Daniil Gavrilov" ]	[ "cs.LG", "cs.AI", "cs.CL" ]	[ "Computer Science" ]	2026-06-08T00:00:00	https://arxiv.org/abs/2606.10029	https://arxiv.org/pdf/2606.10029v1	2606.10029	null	0	0	false	null	null	0
5f3b38be10e8c262989013a6103634d853b13b5b4dedc9712840ddd2e44ad384	[ "arxiv", "semantic_scholar" ]	Closure-Validated Circuit Discovery in Attention Heads: Co-activation Proposes, Ablation Disposes	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 -- ...	[ "Yongzhong Xu" ]	[ "cs.LG", "cs.AI" ]	[ "Computer Science" ]	2026-06-08T00:00:00	https://arxiv.org/abs/2606.09607	https://arxiv.org/pdf/2606.09607v1	2606.09607	null	0	0	false	null	null	0
38ffcec353c1329dcce337691d2a0f4e055fc20b487a5ae665cf04fa555cad77	[ "arxiv", "semantic_scholar" ]	SAEExplainer: Interpreting SAE Features with Activation-Guided Preference Optimization	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...	[ "Jingyi He", "Haiyan Zhao", "Ruxue Shi", "Yanguang Liu", "Xin Wang", "Fei Sun", "Mengnan Du" ]	[ "cs.CL", "cs.LG" ]	[ "Computer Science" ]	2026-06-07T00:00:00	https://arxiv.org/abs/2606.08496	https://arxiv.org/pdf/2606.08496v1	2606.08496	null	0	0	false	null	null	0
d8c0c74f8cf16ed3af5d0f91eaabd0c45001bfc085afc8a2c5c94ea3ae38c1a8	[ "arxiv", "semantic_scholar" ]	Ablation-Reversible Heads Don't Transfer: A Stress Test for Mechanistic Role Claims in Transformers	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 mod...	[ "Philip Quirke" ]	[ "cs.AI" ]	[ "Computer Science" ]	2026-06-06T00:00:00	https://arxiv.org/abs/2606.08292	https://arxiv.org/pdf/2606.08292v1	2606.08292	null	0	0	false	null	null	0
d3e0248d8265ec0e442bb16f34ea9dffd5dd1eda3f696dfc20fade9a99576518	[ "arxiv", "semantic_scholar" ]	Pre-Intervention Prediction of Sparse Autoencoder Steering Side Effects	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 featu...	[ "Evan Duan" ]	[ "cs.LG", "cs.AI" ]	[ "Computer Science" ]	2026-06-06T00:00:00	https://arxiv.org/abs/2606.08365	https://arxiv.org/pdf/2606.08365v1	2606.08365	null	0	0	false	null	null	0
e2b02670f6bad11af8c09cd0dfa31318da3fe5d47a1454fca1239b03f3ffd994	[ "arxiv", "semantic_scholar" ]	A Geometric View for Understanding Concept Learning and Neuron Interpretation in Sparse Autoencoders	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 unclea...	[ "Chenhao Zhang", "Chris Lin", "Su-In Lee" ]	[ "cs.LG", "cs.AI" ]	[ "Computer Science" ]	2026-06-05T00:00:00	https://arxiv.org/abs/2606.07007	https://arxiv.org/pdf/2606.07007v1	2606.07007	null	0	0	false	null	null	0
fbb5b603e1abda844ac2289acd48a85b4f7cc7f5e8a4f56046a2ea4062257dd4	[ "arxiv", "semantic_scholar" ]	When Attribution Patching Lies: Diagnosis and a Second-Order Correction	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 patchi...	[ "Luyang Zhang", "Jialu Wang" ]	[ "cs.LG", "cs.AI" ]	[ "Computer Science" ]	2026-06-05T00:00:00	https://arxiv.org/abs/2606.09899	https://arxiv.org/pdf/2606.09899v1	2606.09899	null	0	0	false	null	null	0
9403246c88389b0a9840abc7cf0aa0706f5902f6541296d944023428b91ba21d	[ "arxiv", "semantic_scholar" ]	Interpreting Brain Responses to Language with Sparse Features from Language Models	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 bo...	[ "Michael A. Lepori", "Kendrick Kay", "Greta Tuckute" ]	[ "cs.CL" ]	[ "Computer Science" ]	2026-06-05T00:00:00	https://arxiv.org/abs/2606.06857	https://arxiv.org/pdf/2606.06857v1	2606.06857	null	0	0	false	null	null	0
824065f19590f04a193cb5db5ab39d46463158c4b9e8b2908972c742e2f031b2	[ "arxiv", "semantic_scholar" ]	Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability	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 fe...	[ "Seyed Arshan Dalili", "Mehrdad Mahdavi" ]	[ "cs.LG", "cs.AI" ]	[ "Computer Science" ]	2026-06-04T00:00:00	https://arxiv.org/abs/2606.06333	https://arxiv.org/pdf/2606.06333v1	2606.06333	null	0	0	false	null	null	0
945b6b585536ce7c5867ab3fa93c805563e57e4b3afb8b153faa65cd309e220d	[ "arxiv", "semantic_scholar" ]	Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Transformers	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 inte...	[ "Tang Li", "Yanlin Chen", "Mengmeng Ma", "Xi Peng" ]	[ "cs.CV", "cs.AI", "cs.LG" ]	[ "Computer Science" ]	2026-06-04T00:00:00	https://arxiv.org/abs/2606.06664	https://arxiv.org/pdf/2606.06664v1	2606.06664	null	0	0	true	https://github.com/deep-real/ViSAE	null	0
2c7efcda5cd5b35cd7c8c3005d3d661033f4a619491a297a712510d2f6e84f6f	[ "arxiv", "semantic_scholar" ]	Mechanistic Insights into Functional Sparsity in Multimodal LLMs via CoRe Heads	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 struct...	[ "Ruoxi Sun", "Quantong Qiu", "Juntao Li", "Zecheng Tang", "Yihang Lou", "Min Zhang" ]	[ "cs.CL", "cs.AI" ]	[ "Computer Science" ]	2026-06-04T00:00:00	https://arxiv.org/abs/2606.05843	https://arxiv.org/pdf/2606.05843v1	2606.05843	null	0	0	false	null	null	0
8f4a6d83e7fcdb219f540ec9a1069105811aca3c35da09f73ee5cc7e63857c79	[ "arxiv", "semantic_scholar" ]	How Optimality Structures Sparse Dictionaries: A Theory for Understanding SAE Representations	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...	[ "William Dorrell" ]	[ "q-bio.NC", "cs.LG" ]	[ "Biology", "Computer Science" ]	2026-06-01T00:00:00	https://arxiv.org/abs/2606.02385	https://arxiv.org/pdf/2606.02385v1	2606.02385	null	0	0	false	null	null	0
0a93715e09480549c60738a0a724a902ee4e05668a30bd92153e7698b12d12a4	[ "arxiv", "semantic_scholar" ]	Sparse Autoencoders for Interpretable Emotion Control in Text-to-Speech	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 un...	[ "Hongfei Du", "Jiacheng Shi", "Sidi Lu", "Gang Zhou", "Ye Gao" ]	[ "cs.CL" ]	[ "Computer Science" ]	2026-05-31T00:00:00	https://arxiv.org/abs/2606.01479	https://arxiv.org/pdf/2606.01479v1	2606.01479	null	0	0	false	null	null	0
3558c4aa5f7bbf1685e652f1839178a604f8a38a66a09d44339faacec4dbb8c1	[ "arxiv", "semantic_scholar" ]	Query Lens: Interpreting Sparse Key-Value Features with Indirect Effects	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 deco...	[ "Hwiyeong Lee", "Ingyu Bang", "Uiji Hwang", "Hyelim Lim", "Taeuk Kim" ]	[ "cs.LG", "cs.AI" ]	[ "Computer Science" ]	2026-05-30T00:00:00	https://arxiv.org/abs/2606.07617	https://arxiv.org/pdf/2606.07617v1	2606.07617	null	0	0	false	null	null	0
ee332dafe64489c528d752ccef91fbb1df438ff3edaf4f165bfdb4e1c5904c05	[ "arxiv", "semantic_scholar" ]	On the Relationship Between Activation Outliers and Feature Death in Sparse Autoencoders	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 AlphaFold...	[ "Elana Simon", "Etowah Adams", "James Zou" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-29T00:00:00	https://arxiv.org/abs/2605.31518	https://arxiv.org/pdf/2605.31518v1	2605.31518	null	0	0	false	null	null	0
6a4e69ec212affbc81dc1380a6e0e38167ce1446eb999b64a57c8c2716c74bca	[ "arxiv", "semantic_scholar" ]	Toward Identifiable Sparse Autoencoders	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 dic...	[ "Walter Nelson", "Theofanis Karaletsos", "Francesco Locatello" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-29T00:00:00	https://arxiv.org/abs/2605.31245	https://arxiv.org/pdf/2605.31245v1	2605.31245	null	0	0	false	null	null	0
8bbd9ecd0b4553e82f1bcdeb89928572dddcc386e3db384bc49fdeddcdb32d6e	[ "arxiv", "semantic_scholar" ]	Latent Space Disentanglement via Activation Steering for Interpretable Attribute Control in Symbolic Music Generation	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...	[ "Ioannis Prokopiou", "Pantelis Vikatos", "Maximos Kaliakatsos-Papakostas", "Theodoros Giannakopoulos", "Themos Stafylakis" ]	[ "cs.SD", "cs.AI", "cs.IR", "cs.LG" ]	[ "Computer Science" ]	2026-05-29T00:00:00	https://arxiv.org/abs/2605.31295	https://arxiv.org/pdf/2605.31295v1	2605.31295	null	0	0	false	null	null	0
dbb352536e57a425fe8adc4666ff6f754d04c8aa4513247aedb2201679849962	[ "arxiv", "semantic_scholar" ]	Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet	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 lay...	[ "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...	[ "cs.AI" ]	[ "Computer Science" ]	2026-05-28T00:00:00	https://arxiv.org/abs/2605.29358	https://arxiv.org/pdf/2605.29358v1	2605.29358	null	539	38	false	null	null	0.6831
5e0dc9be499614a323467889425964aa828471d3402271a4057ed73d7b562058	[ "arxiv", "semantic_scholar" ]	Discovering a Zeta Map Algorithm on Dyck Paths via Mechanistic Interpretability	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...	[ "Xiaoyu Huang", "Blake Jackson", "Kyu-Hwan Lee" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-28T00:00:00	https://arxiv.org/abs/2605.30482	https://arxiv.org/pdf/2605.30482v1	2605.30482	null	0	0	false	null	null	0
4af67a54fd11fec90900209cf74fab0decfba078b5c50844f087d09f023902e5	[ "arxiv", "semantic_scholar" ]	Semantic Optimal Transport for Sparse Autoencoder Feature Matching and Circuit Compression	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 treat...	[ "Tue M. Cao", "Nguyen Do", "My T. Thai" ]	[ "cs.LG", "cs.AI" ]	[ "Computer Science" ]	2026-05-27T00:00:00	https://arxiv.org/abs/2605.28567	https://arxiv.org/pdf/2605.28567v1	2605.28567	null	0	0	false	null	null	0
64751ee4f010157b0f78e670fd8949318a2adfc99d6fc7cd21954659c31d2fbc	[ "arxiv", "semantic_scholar" ]	Feature Geometry of LoRA Adapters: A Sparse Autoencoder Analysis of Representational Divergence in Fine-Tuned Language Models	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 i...	[ "Prasanth K K" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-27T00:00:00	https://arxiv.org/abs/2605.28896	https://arxiv.org/pdf/2605.28896v1	2605.28896	null	0	0	false	null	null	0
809be0f3978148c25965d948f662a80aa093158510a71330a4802da5001beba0	[ "arxiv", "semantic_scholar" ]	Residualized Temporal Sparse Autoencoders for Interpreting Diffusion Models	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...	[ "Calvin Yeung", "Prathyush Poduval", "Ali Zakeri", "Zhuowen Zou", "Mohsen Imani" ]	[ "cs.CV", "cs.AI", "cs.LG" ]	[ "Computer Science" ]	2026-05-27T00:00:00	https://arxiv.org/abs/2605.27813	https://arxiv.org/pdf/2605.27813v1	2605.27813	null	0	0	false	null	null	0
f37857372b005008aa7061a859ade8cabc445cc5d69cc5b0cc3847af91b73f62	[ "arxiv", "semantic_scholar" ]	Sign-Aware Gated Sparse Autoencoders: Modeling Anticorrelated Features with Bi-Jump-ReLU Activations	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 ...	[ "Bartosz Wieciech", "Zmnako Awrahman", "Marcin Czelej", "Victor Hugo Jaramillo Velasquez", "Wioletta Stobieniecka" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-27T00:00:00	https://arxiv.org/abs/2605.28149	https://arxiv.org/pdf/2605.28149v1	2605.28149	null	0	0	false	null	null	0
d1d5d90b35127bb205288e4c8ae0062fef7e45d371876c90b6c8f4766583db7b	[ "arxiv", "semantic_scholar" ]	SAE-FD: Sparse Autoencoder Feature Distillation for Continual Learning of Large Language Models	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 spa...	[ "Mingxu Zhang", "Yuhan Li", "Lujundong Li", "Dazhong Shen", "Hui Xiong", "Ying Sun" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-25T00:00:00	https://arxiv.org/abs/2605.25525	https://arxiv.org/pdf/2605.25525v1	2605.25525	null	0	0	false	null	null	0
58d241e8cd90205465aebe7d65ddd13a2b7645773d4417166bdf26505ef2cf2b	[ "arxiv", "semantic_scholar" ]	MechRL: Reinforcement Learning Agents Perform Circuit Discovery for Mechanistic Interpretability	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 operate...	[ "Barsat Khadka" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-25T00:00:00	https://arxiv.org/abs/2605.26343	https://arxiv.org/pdf/2605.26343v1	2605.26343	null	0	0	false	null	null	0
12d204ba7e01666f4d81d5aaf79c3626ac05e31a83cbc853a13148ffea2100b7	[ "arxiv", "semantic_scholar" ]	Interpretability Transfer from Language to Vision via Sparse Autoencoders	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 tra...	[ "Alexey Kravets", "Da Li", "Chuan Li", "Da Chen", "Vinay P. Namboodiri" ]	[ "cs.CV" ]	[ "Computer Science" ]	2026-05-24T00:00:00	https://arxiv.org/abs/2605.24946	https://arxiv.org/pdf/2605.24946v1	2605.24946	null	0	0	false	null	ICML 2026	0
d7b4281de74ac9c1c9e358e00dc3dc85b180616a4c18dd8201f6b5f9dbd679df	[ "arxiv", "semantic_scholar" ]	Transformer Field Theory: A Response-Theoretic Approach to Mechanistic Interpretability	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...	[ "David N. Olivieri", "Antonio F. Pérez Rodríguez" ]	[ "cs.LG", "cs.AI" ]	[ "Computer Science" ]	2026-05-24T00:00:00	https://arxiv.org/abs/2605.25225	https://arxiv.org/pdf/2605.25225v2	2605.25225	null	0	0	false	null	null	0
93b95aaa67105b6ba6abfc716cdc422ea9ed2421a90afd83e90d2c3ef2a0008f	[ "arxiv", "semantic_scholar" ]	Geometry-Adaptive Explainer for Faithful Dictionary-Based Interpretability under Distribution Shift	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 s...	[ "Sungjun Lim", "Heedong Kim", "Andrew Lee", "Kyungwoo Song" ]	[ "cs.LG", "cs.CL" ]	[ "Computer Science" ]	2026-05-21T00:00:00	https://arxiv.org/abs/2605.21849	https://arxiv.org/pdf/2605.21849v1	2605.21849	null	0	0	false	null	null	0
7ed723e87320133653416543d6561a4ee198ad7eea915677b5cb13e9e0ea2c26	[ "arxiv", "semantic_scholar" ]	The Attribution Contract: Feature Attribution for Generative Language Models	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 pre...	[ "Giang Nguyen" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-21T00:00:00	https://arxiv.org/abs/2605.23080	https://arxiv.org/pdf/2605.23080v2	2605.23080	null	0	0	false	null	null	0
c8f9e51511321d79452b9aee733a923b550a624dd54c86bd4ca29e7d601eb43b	[ "arxiv", "semantic_scholar" ]	Mechanistic origins of catastrophic forgetting: why RL preserves circuits better than SFT?	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...	[ "Jeanmely Rojas Nunez", "Viraj Sawant", "Nathan Allen", "Nomgondalai Amgalanbaatar", "Yannis Zongo", "Vasu Sharma", "Maheep Chaudhary" ]	[ "cs.LG", "cs.AI", "cs.CL" ]	[ "Computer Science" ]	2026-05-21T00:00:00	https://arxiv.org/abs/2605.28860	https://arxiv.org/pdf/2605.28860v2	2605.28860	null	2	0	true	https://github.com/rl-sft-circuit-research/differential-circuit-vulnerability	null	0.1193
01c7d0afd8807a9b89023cdadcbbc618908fe5b20f2cb3b9a8bd70ec51b7b960	[ "arxiv", "semantic_scholar" ]	Reading Task Failure Off the Activations: A Sparse-Feature Audit of GPT-2 Small on Indirect Object Identification	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 ...	[ "Mahdi Nasermoghadasi" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-21T00:00:00	https://arxiv.org/abs/2605.22719	https://arxiv.org/pdf/2605.22719v1	2605.22719	null	0	0	false	null	null	0
839b40fed6d90373fbba8fd32fb83107c2792406d505557e2fba12891cc50cfa	[ "arxiv", "semantic_scholar" ]	SegCompass: Exploring Interpretable Alignment with Sparse Autoencoders for Enhanced Reasoning Segmentation	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 re...	[ "Zhenyu Lu", "Liupeng Li", "Jinpeng Wang", "Haoqian Kang", "Yan Feng", "Ke Chen", "Yaowei Wang" ]	[ "cs.CV", "cs.LG", "cs.MM", "eess.IV" ]	[ "Computer Science", "Engineering" ]	2026-05-21T00:00:00	https://arxiv.org/abs/2605.22658	https://arxiv.org/pdf/2605.22658v1	2605.22658	null	0	0	true	https://github.com/ZhenyuLU-Heliodore/SegCompass	null	0
8831fedd322a19f2ecf2ec7be845e477aabe2301dcf4bf7bb0b8f27a02b00df8	[ "arxiv", "semantic_scholar" ]	From Correlation to Cause: A Five-Stage Methodology for Feature Analysis in Transformer Language Models	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 ...	[ "Caleb Munigety" ]	[ "cs.CL", "cs.AI" ]	[ "Computer Science" ]	2026-05-21T00:00:00	https://arxiv.org/abs/2605.22462	https://arxiv.org/pdf/2605.22462v1	2605.22462	null	0	0	false	null	null	0
135875bd08b4773326c76ce4ef2dcf45a6a84d77d5504adda56b4774c1de6006	[ "arxiv", "semantic_scholar" ]	Steered Generation via Gradient-Based Optimization on Sparse Query Features	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 att...	[ "Sumanta Bhattacharyya", "Pedram Rooshenas" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-21T00:00:00	https://arxiv.org/abs/2605.23040	https://arxiv.org/pdf/2605.23040v1	2605.23040	null	0	0	false	null	null	0
85706e97c6fd8a954edfd547d8a89e778a7df671e590245626801b8d6a971ec7	[ "arxiv", "semantic_scholar" ]	From Circuit Evidence to Mechanistic Theory: An Inductive Logic Approach	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 w...	[ "Nura Aljaafari", "Danilo S. Carvalho", "Andre Freitas" ]	[ "cs.LG", "cs.AI", "cs.LO" ]	[ "Computer Science" ]	2026-05-20T00:00:00	https://arxiv.org/abs/2605.21303	https://arxiv.org/pdf/2605.21303v1	2605.21303	null	0	0	false	null	null	0
83c24af6fcdc8d35668d50f97cc596a0f5c46ec887214ae149458b409072122c	[ "arxiv", "semantic_scholar" ]	Mechanistic Interpretability for Learning Assurance of a Vision-Based Landing System	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 assur...	[ "Romeo Valentin", "Olivia Beyer Bruvik", "Marc R. Schlichting", "Mykel J. Kochenderfer" ]	[ "cs.LG", "cs.CV", "cs.RO" ]	[ "Computer Science" ]	2026-05-20T00:00:00	https://arxiv.org/abs/2605.20607	https://arxiv.org/pdf/2605.20607v1	2605.20607	null	0	0	false	null	null	0
1fe4079d976d0ae11f3e122a2282b1922bcc8989f94ea4b8f1917bb15b0613da	[ "arxiv", "semantic_scholar" ]	Learning fMRI activations dictionaries across individual geometries via optimal transport	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 chall...	[ "Sonia Mazelet", "Rémi Flamary", "Bertrand Thirion" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-20T00:00:00	https://arxiv.org/abs/2605.20883	https://arxiv.org/pdf/2605.20883v1	2605.20883	null	0	0	false	null	null	0
8ecb521e037b70bf9f4b63a32af4b1f2dbcb67d50e152b65d176b96d5c29e298	[ "arxiv", "semantic_scholar" ]	Spectral Integrated Gradients for Coarse-to-Fine Feature Attribution	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 gr...	[ "Soyeon Kim", "Seongwoo Lim", "Kyowoon Lee", "Jaesik Choi" ]	[ "cs.CV", "cs.AI", "cs.LG" ]	[ "Computer Science" ]	2026-05-19T00:00:00	https://arxiv.org/abs/2605.19607	https://arxiv.org/pdf/2605.19607v1	2605.19607	null	1	0	true	https://github.com/leekwoon/sig/	null	0.0753
e901e89fa28c2409767664c350a48bf45af646d560a4862a30691b0d0355c307	[ "arxiv", "semantic_scholar" ]	Aligned Training: A Parameter-Free Method to Improve Feature Quality and Stability of Sparse Autoencoders (SAE)	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 t...	[ "Michał Brzozowski", "Neo Christopher Chung" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-18T00:00:00	https://arxiv.org/abs/2605.18629	https://arxiv.org/pdf/2605.18629v2	2605.18629	null	1	0	false	null	null	0.0753
422e2c8c0ab828bd924bd80d23a0f7e654795d6d6a530dad3337101746d4179b	[ "arxiv", "semantic_scholar" ]	Toy Combinatorial Interpretability Models Reveal Lottery Tickets in Early Feature Space	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...	[ "Alon Bebchuk", "Nir Shavit" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-18T00:00:00	https://arxiv.org/abs/2605.17704	https://arxiv.org/pdf/2605.17704v1	2605.17704	null	0	0	false	null	null	0
ff90da5610b0f278b605bc6ebc3379ca2ebdbe9360d2fe1e1fe2418a82730d22	[ "arxiv", "semantic_scholar" ]	Beyond Linear Superposition: Discovering Climate Features in AI Weather Models with KAN-SAE	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 existi...	[ "Minjong Cheon" ]	[ "cs.LG", "cs.AI", "cs.CV", "physics.ao-ph" ]	[ "Computer Science", "Physics" ]	2026-05-17T00:00:00	https://arxiv.org/abs/2605.17493	https://arxiv.org/pdf/2605.17493v1	2605.17493	null	0	0	false	null	null	0
91cc15889fe1f757e44e24c5e6d48e54c2b3e5b5528368e088c8d94694c760b7	[ "arxiv", "semantic_scholar" ]	A Distributional View for Visual Mechanistic Interpretability: KL-Minimal Soft-Constraint Principle	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 t...	[ "Guancheng Zhou", "Yisi Luo", "Zhengfu He", "Zhenyu Jin", "Xuyang Ge", "Wentao Shu", "Deyu Meng", "Xipeng Qiu" ]	[ "cs.CV", "cs.AI" ]	[ "Computer Science" ]	2026-05-17T00:00:00	https://arxiv.org/abs/2605.17504	https://arxiv.org/pdf/2605.17504v1	2605.17504	null	0	0	false	null	null	0
d6bfa826345bfe64f55ff6dfe1297951e59daa7464e272f61974827f9d6d8c83	[ "arxiv", "semantic_scholar" ]	Mechanistically Interpretable Neural Encoding Reveals Fine-Grained Functional Selectivity in Human Visual Cortex	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, exi...	[ "Idan Daniel Grosbard", "Mor Geva", "Galit Yovel" ]	[ "cs.CV", "cs.AI", "cs.CL", "cs.LG", "q-bio.NC" ]	[ "Computer Science", "Biology" ]	2026-05-15T00:00:00	https://arxiv.org/abs/2605.16468	https://arxiv.org/pdf/2605.16468v1	2605.16468	null	0	0	false	null	null	0
7d6c5fa3b40eb8bd2f049374259c4db0929c1ccb9d9c6fc05451b6a753ae36d2	[ "arxiv", "semantic_scholar" ]	AGOP-IxG: A Gradient Covariance Filter for Local Feature Attribution on Tabular Data, with a Controlled Benchmark	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 rigo...	[ "Raj Kiran Gupta Katakam" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-15T00:00:00	https://arxiv.org/abs/2605.15700	https://arxiv.org/pdf/2605.15700v1	2605.15700	null	0	0	false	null	null	0
b42a28978334555126449a5cb185a04327665b4c153d33e2c5a19c8c59fde04c	[ "arxiv", "semantic_scholar" ]	Sparse Autoencoders enable Robust and Interpretable Fine-tuning of CLIP models	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 re...	[ "Fabian Morelli", "Arnas Uselis", "Ankit Sonthalia", "Seong Joon Oh" ]	[ "cs.CV" ]	[ "Computer Science" ]	2026-05-15T00:00:00	https://arxiv.org/abs/2605.15961	https://arxiv.org/pdf/2605.15961v1	2605.15961	null	0	0	true	https://github.com/Fabian-Mor/sae-ft	null	0
bb7046132ca96f49b3331b7ca72b53b256254370d52fa6de4feb126429d298ae	[ "arxiv", "semantic_scholar" ]	RoSHAP: A Distributional Framework and Robust Metric for Stable Feature Attribution	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 attributi...	[ "Lanxin Xiang", "Liang Shi", "Youhui Ye", "Boyu Jiang", "Dawei Zhou", "Feng Guo" ]	[ "stat.ML", "cs.LG" ]	[ "Mathematics", "Computer Science" ]	2026-05-14T00:00:00	https://arxiv.org/abs/2605.15154	https://arxiv.org/pdf/2605.15154v1	2605.15154	null	0	0	false	null	null	0
84d4041427d9b5e7945a7e0cd99487bd0b7482a8da9a1a4e22fc5e79ffee7e31	[ "arxiv", "semantic_scholar" ]	Exemplar Partitioning for Mechanistic Interpretability	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-clusterin...	[ "Jessica Rumbelow" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-14T00:00:00	https://arxiv.org/abs/2605.14347	https://arxiv.org/pdf/2605.14347v2	2605.14347	null	0	0	true	https://github.com/jessicarumbelow/exemplar-partitioning	null	0
bc43df3dfcc0a57d000f5b7d5438adfa30e10eaa5a26650f4149a55839e47766	[ "arxiv", "semantic_scholar" ]	From Weight Perturbation to Feature Attribution for Explaining Fully Connected Neural Networks	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 approac...	[ "Thodoris Lymperopoulos", "Denia Kanellopoulou" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-14T00:00:00	https://arxiv.org/abs/2605.15328	https://arxiv.org/pdf/2605.15328v1	2605.15328	null	0	0	false	null	null	0
c4960cfe1a28bb618bfd29c2cc55c49446fa803f91040897197ae7270e9550ca	[ "arxiv", "semantic_scholar" ]	Exploring Geographic Relative Space in Large Language Models through Activation Patching	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 ...	[ "Stef De Sabbata", "Rahul Baiju", "Stefano Mizzaro", "Kevin Roitero" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-14T00:00:00	https://arxiv.org/abs/2605.14535	https://arxiv.org/pdf/2605.14535v1	2605.14535	null	0	0	false	null	null	0
400ff709ffa21ffb129ce28e7bfa969c6e9885bc7425b377d5bb370514408c23	[ "arxiv", "semantic_scholar" ]	Descriptive Collision in Sparse Autoencoder Auto-Interpretability: When One Explanation Describes Many Features	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 mea...	[ "Jordan F. McCann" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-13T00:00:00	https://arxiv.org/abs/2605.12874	https://arxiv.org/pdf/2605.12874v1	2605.12874	null	0	0	false	null	null	0
04d70fdec71bfac307e1e5c47ac9e5ed5c7e77c94dd2aa1818a36dc77f8cec5f	[ "arxiv", "semantic_scholar" ]	Mechanistic Interpretability of EEG Foundation Models via Sparse Autoencoders	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 di...	[ "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" ]	[ "cs.LG", "cs.HC", "cs.NE" ]	[ "Computer Science" ]	2026-05-13T00:00:00	https://arxiv.org/abs/2605.13930	https://arxiv.org/pdf/2605.13930v3	2605.13930	null	0	0	false	null	null	0
812fcb46cbb55be2a5c7e196a348f66e92af1dbc27af29abd8fbfb703f689ad7	[ "arxiv", "semantic_scholar" ]	Deep Learning as Neural Low-Degree Filtering: A Spectral Theory of Hierarchical Feature Learning	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 s...	[ "Yatin Dandi", "Matteo Vilucchio", "Luca Arnaboldi", "Hugo Tabanelli", "Florent Krzakala" ]	[ "cs.LG", "cond-mat.dis-nn", "stat.ML" ]	[ "Computer Science", "Physics", "Mathematics" ]	2026-05-13T00:00:00	https://arxiv.org/abs/2605.13612	https://arxiv.org/pdf/2605.13612v1	2605.13612	null	0	0	true	https://github.com/IdePHICS/Neural-LoFi-Theory	null	0
699e0b8a67cbe85deda042783da6ae8316ba5c91f8ef9ec9fc7951c14a6eafb6	[ "arxiv", "semantic_scholar" ]	Feature Visualization Recovers Known Cortical Selectivity from TRIBE v2	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 whet...	[ "Stuart Bladon", "Brinnae Bent" ]	[ "q-bio.NC", "cs.LG" ]	[ "Biology", "Computer Science" ]	2026-05-13T00:00:00	https://arxiv.org/abs/2605.13904	https://arxiv.org/pdf/2605.13904v1	2605.13904	null	0	0	true	https://github.com/recozers/Tribe-V2-Interp	null	0
31ab91bf5f1eb92a226074829452d376e629a26e031e81ad783688a8b82803b5	[ "arxiv", "semantic_scholar" ]	Mechanistic Interpretability of ASR models using Sparse Autoencoders	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 ...	[ "Dan Pluth", "Zachary Nicholas Houghton", "Yu Zhou", "Vijay K. Gurbani" ]	[ "cs.CL" ]	[ "Computer Science" ]	2026-05-12T00:00:00	https://arxiv.org/abs/2605.12225	https://arxiv.org/pdf/2605.12225v1	2605.12225	null	0	0	false	null	null	0
b044e35668144e1dab66863b395c38765c01c54e69063d360fae5ca15a789d57	[ "arxiv", "semantic_scholar" ]	Qwen-Scope: Turning Sparse Features into Development Tools for Large Language Models	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 ...	[ "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" ]	[ "cs.CL", "cs.LG" ]	[ "Computer Science" ]	2026-05-12T00:00:00	https://arxiv.org/abs/2605.11887	https://arxiv.org/pdf/2605.11887v1	2605.11887	null	4	0	true	null	null	0.1747
146a0ff991161933bd1f30281e5adbe6903d978dba324fe47e1561be632318de	[ "arxiv", "semantic_scholar" ]	AGOP as Explanation: From Feature Learning to Per-Sample Attribution in Image Classifiers	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 ...	[ "Raj Kiran Gupta Katakam" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-12T00:00:00	https://arxiv.org/abs/2605.12816	https://arxiv.org/pdf/2605.12816v1	2605.12816	null	1	0	false	null	null	0.0753
1de81299ec55ce013a02d7aa9d9200b5752600dee358cba0ff3cb5cd4bed805a	[ "arxiv", "semantic_scholar" ]	FAME: Feature Activation Map Explanation on Image Classification and Face Recognition	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 proce...	[ "Xinyi Zhang", "Manuel Günther" ]	[ "cs.CV" ]	[ "Computer Science" ]	2026-05-12T00:00:00	https://arxiv.org/abs/2605.12017	https://arxiv.org/pdf/2605.12017v1	2605.12017	null	0	0	true	https://github.com/AIML-IfI/fame.}	null	0
069b1d13bbafbe7aaaf4316b51477b912695f57083bc71481754fc548bd2d5a0	[ "arxiv", "semantic_scholar" ]	Dissecting Jet-Tagger Through Mechanistic Interpretability	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 comp...	[ "Saurabh Rai", "Sanmay Ganguly" ]	[ "hep-ph", "cs.LG", "hep-ex" ]	[ "Physics", "Computer Science" ]	2026-05-11T00:00:00	https://arxiv.org/abs/2605.09881	https://arxiv.org/pdf/2605.09881v1	2605.09881	null	0	0	false	null	null	0
4c742a039792a6687017b1b76f38658bba531b1ff8feaa5f5730805dcaf2f2c3	[ "arxiv", "semantic_scholar" ]	Sequential Feature Selection for Efficient Landslide Segmentation from Multi-Spectral Data	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: redund...	[ "Arsalaan Ahmad", "Oktay Karakus", "Paul L. Rosin" ]	[ "cs.LG", "cs.AI" ]	[ "Computer Science" ]	2026-05-10T00:00:00	https://arxiv.org/abs/2605.09746	https://arxiv.org/pdf/2605.09746v1	2605.09746	null	0	0	false	null	null	0
e09e55798fef823c95cafa25c486411c578bb604ab99f1b3df84fba8d0ad48af	[ "arxiv", "semantic_scholar" ]	From Mechanistic to Compositional Interpretability	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 interpretabilit...	[ "Ward Gauderis", "Thomas Dooms", "Steven T. Holmer", "Kola Ayonrinde", "Geraint A. Wiggins" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-09T00:00:00	https://arxiv.org/abs/2605.08934	https://arxiv.org/pdf/2605.08934v1	2605.08934	null	1	0	false	null	null	0.0753
fd4c7e9f4943ccb409f5fa5f84dc1593891b23fc33b98bf6e6b7aa02516f56cd	[ "arxiv", "semantic_scholar" ]	Tree SAE: Learning Hierarchical Feature Structures in Sparse Autoencoders	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, th...	[ "Tue M. Cao", "Hoang X. Nhat", "Raed Alharbi", "Phi Le Nguyen", "My T. Thai" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-08T00:00:00	https://arxiv.org/abs/2605.07922	https://arxiv.org/pdf/2605.07922v2	2605.07922	null	0	0	false	null	null	0
b2eefefa6727c1ce0ffdee9986b97a46df6ced35cd43754fb08795a159818d53	[ "arxiv", "semantic_scholar" ]	How Much Do Circuits Tell Us? Measuring the Consistency and Specificity of Language Model Circuits	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-s...	[ "Michael Li", "Nishant Subramani" ]	[ "cs.CL" ]	[ "Computer Science" ]	2026-05-08T00:00:00	https://arxiv.org/abs/2605.08348	https://arxiv.org/pdf/2605.08348v1	2605.08348	null	1	0	false	null	null	0.0753
17cd43c5e3b46996b233604343f278be0b80931253731181c0851382eca069e4	[ "arxiv", "semantic_scholar" ]	From Token Lists to Graph Motifs: Weisfeiler-Lehman Analysis of Sparse Autoencoder Features	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 sh...	[ "Ruben Fernandez-Boullon", "Pablo Magariños-Docampo", "Javier Perez-Robles" ]	[ "cs.AI" ]	[ "Computer Science" ]	2026-05-07T00:00:00	https://arxiv.org/abs/2605.06494	https://arxiv.org/pdf/2605.06494v1	2605.06494	null	0	0	false	null	null	0
034b016c1e5acebed5ebe28808cf92a06f07f2e2e475cde248954f7f4eac011c	[ "arxiv", "semantic_scholar" ]	Patch-Effect Graph Kernels for LLM Interpretability	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 propo...	[ "Ruben Fernandez-Boullon", "David N. Olivieri" ]	[ "cs.AI", "cs.CL" ]	[ "Computer Science" ]	2026-05-07T00:00:00	https://arxiv.org/abs/2605.06480	https://arxiv.org/pdf/2605.06480v1	2605.06480	null	0	0	false	null	null	0
a9bd8482ffec855bcf4f605dcbc8a8f6684b5d520a82fdc86a80d75a7df8fea5	[ "arxiv", "semantic_scholar" ]	SoftSAE: Dynamic Top-K Selection for Adaptive Sparse Autoencoders	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 netwo...	[ "Jakub Stępień", "Marcin Mazur", "Jacek Tabor", "Przemysław Spurek" ]	[ "cs.LG", "cs.CV" ]	[ "Computer Science" ]	2026-05-07T00:00:00	https://arxiv.org/abs/2605.06610	https://arxiv.org/pdf/2605.06610v2	2605.06610	null	0	0	true	https://github.com/St0pien/SoftSAE	null	0
89ed71b98a298aa114566612798e63023180ef6f9110a71c1f62ba811894d25e	[ "arxiv", "semantic_scholar" ]	Attributions All the Way Down? The Metagame of Interpretability	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...	[ "Hubert Baniecki", "Przemyslaw Biecek", "Fabian Fumagalli" ]	[ "cs.LG", "cs.AI", "stat.ML" ]	[ "Computer Science", "Mathematics" ]	2026-05-07T00:00:00	https://arxiv.org/abs/2605.06295	https://arxiv.org/pdf/2605.06295v1	2605.06295	null	0	0	false	null	null	0
a45dff8ddca734b1b4f792e41c4e78b95d040527068a14536c5d96b278068aa0	[ "arxiv", "semantic_scholar" ]	Superposition Is Not Necessary: A Mechanistic Interpretability Analysis of Transformer Representations for Time Series Forecasting	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 mech...	[ "Alper Yıldırım" ]	[ "cs.LG", "cs.AI" ]	[ "Computer Science" ]	2026-05-06T00:00:00	https://arxiv.org/abs/2605.05151	https://arxiv.org/pdf/2605.05151v1	2605.05151	null	0	0	false	null	null	0
9576379bffeb1ac7d7f4ae6a1c5615f048d66cc9dd9478bbe7fdc19ed2d552e5	[ "arxiv", "semantic_scholar" ]	Feature Starvation as Geometric Instability in Sparse Autoencoders	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 exp...	[ "Faris Chaudhry", "Keisuke Yano", "Anthea Monod" ]	[ "cs.LG", "cs.AI", "math.OC", "stat.ML" ]	[ "Computer Science", "Mathematics" ]	2026-05-06T00:00:00	https://arxiv.org/abs/2605.05341	https://arxiv.org/pdf/2605.05341v1	2605.05341	null	0	0	false	null	null	0
44da8df7883394416795eef49d1bd498f12a1bc0b284f73c18047f62c4702490	[ "arxiv", "semantic_scholar" ]	Deep Dreams Are Made of This: Visualizing Monosemantic Features in Diffusion Models	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 repr...	[ "Adam Szokalski", "Mateusz Modrzejewski" ]	[ "cs.LG", "cs.CV" ]	[ "Computer Science" ]	2026-05-06T00:00:00	https://arxiv.org/abs/2605.08218	https://arxiv.org/pdf/2605.08218v1	2605.08218	null	0	0	false	null	null	0
6ee65a54024b5dd864b20015ba4235653b0ce078a9c353993bb28e11afcb5956	[ "arxiv", "semantic_scholar" ]	GRAFT: Auditing Graph Neural Networks via Global Feature Attribution	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 ...	[ "Rishi Raj Sahoo", "Subhankar Mishra" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-05T00:00:00	https://arxiv.org/abs/2605.03377	https://arxiv.org/pdf/2605.03377v1	2605.03377	null	0	0	false	null	null	0
e8df5cdd1d6c33bdb95ffb9a9a9b3f9bf8f77d15fef9a6e2a7a2f3cd716433e4	[ "arxiv", "semantic_scholar" ]	Manifold-Aligned Guided Integrated Gradients for Reliable Feature Attribution	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...	[ "Soyeon Kim", "Seongwoo Lim", "Kyowoon Lee", "Jaesik Choi" ]	[ "cs.LG", "cs.AI", "cs.CV" ]	[ "Computer Science" ]	2026-05-04T00:00:00	https://arxiv.org/abs/2605.02167	https://arxiv.org/pdf/2605.02167v3	2605.02167	null	1	0	true	https://github.com/leekwoon/ma-gig/	null	0.0753
d53ea6db57592e317ee46e283e95a31f1e3f5b777b01113722d06073a9bd1c86	[ "arxiv", "semantic_scholar" ]	Pairwise matrices for sparse autoencoders: single-feature inspection mislabels causal axes	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 ...	[ "Michael A. Riegler", "Birk Sebastian Frostelid Torpmann-Hagen" ]	[ "cs.LG" ]	[ "Computer Science" ]	2026-05-04T00:00:00	https://arxiv.org/abs/2605.03160	https://arxiv.org/pdf/2605.03160v1	2605.03160	null	0	0	false	null	null	0
3f6295c9616a7f3ca1fccd23579a8995493e9ecff20290dac075979cc5c5c3b3	[ "arxiv", "semantic_scholar" ]	Feature Rivalry in Sparse Autoencoder Representations: A Mechanistic Study of Uncertainty-Driven Feature Competition in LLMs	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 mo...	[ " Harshavardhan" ]	[ "cs.LG", "cs.CL" ]	[ "Computer Science" ]	2026-05-03T00:00:00	https://arxiv.org/abs/2605.08149	https://arxiv.org/pdf/2605.08149v1	2605.08149	null	0	0	false	null	null	0
5b9380dd736887588b29cd773371765a51bef30729f588143784b177ef4b5103	[ "arxiv", "semantic_scholar" ]	Automated Interpretability and Feature Discovery in Language Models with Agents	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 pro...	[ "Arnau Marin-Llobet", "Javier Ferrando" ]	[ "cs.CL", "cs.AI", "cs.HC" ]	[ "Computer Science" ]	2026-05-02T00:00:00	https://arxiv.org/abs/2605.01555	https://arxiv.org/pdf/2605.01555v1	2605.01555	null	0	0	false	null	null	0
9ce1a8990c12a905b1fa0c35f73c5f1948d87ac2f3980d878794daa87fcd78f2	[ "arxiv", "semantic_scholar" ]	Borrowed Geometry: Cross-Distribution Head-Importance Fingerprints of Frozen Pretrained Gemma 4 31B	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-langua...	[ "Abay Bektursun" ]	[ "cs.LG", "cs.CL" ]	[ "Computer Science" ]	2026-05-01T00:00:00	https://arxiv.org/abs/2605.00333	https://arxiv.org/pdf/2605.00333v2	2605.00333	null	0	0	false	null	null	0
3f2a23ac869ac545381c6b40a77474ab783fcf9958e81b631a3686ce5b92e6fb	[ "arxiv", "semantic_scholar" ]	From Local to Global to Mechanistic: An iERF-Centered Unified Framework for Interpreting Vision Models	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 ...	[ "Yearim Kim", "Sangyu Han", "Nojun Kwak" ]	[ "cs.CV" ]	[ "Computer Science", "Medicine" ]	2026-05-01T00:00:00	https://arxiv.org/abs/2605.00474	https://arxiv.org/pdf/2605.00474v1	2605.00474	10.1109/TPAMI.2026.3688582	0	0	false	null	IEEE Transactions on Pattern Analysis and Machine Intelligence	0
8eb53b1968eddb333c7dc71f4601c93052184d1107cdc52d55444ed2220cd2ef	[ "arxiv", "semantic_scholar" ]	MoRFI: Monotonic Sparse Autoencoder Feature Identification	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...	[ "Dimitris Dimakopoulos", "Shay B. Cohen", "Ioannis Konstas" ]	[ "cs.CL", "cs.LG" ]	[ "Computer Science" ]	2026-04-29T00:00:00	https://arxiv.org/abs/2604.26866	https://arxiv.org/pdf/2604.26866v1	2604.26866	null	0	0	false	null	null	0
4d9ea8333bb3c132855749bc183aed764581ec8075b803e02c97c5f304aa72a1	[ "arxiv", "semantic_scholar" ]	Validating the Clinical Utility of CineECG 3D Reconstructions through Cross-Modal Feature Attribution	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 pa...	[ "Karol Dobiczek", "Maciej Mozolewski", "Szymon Bobek", "Michał Szafarczyk", "Peter van Dam", "Grzegorz J. Nalepa" ]	[ "eess.IV", "cs.LG", "stat.ML" ]	[ "Engineering", "Computer Science", "Mathematics" ]	2026-04-29T00:00:00	https://arxiv.org/abs/2604.27017	https://arxiv.org/pdf/2604.27017v1	2604.27017	null	0	0	false	null	null	0
9c20d095227a499cdf65ee8b3ff8e2716a9a92d2d07e4bb37d4efef9b2305b6b	[ "arxiv", "semantic_scholar" ]	reward-lens: A Mechanistic Interpretability Library for Reward Models	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 sca...	[ "Mohammed Suhail B Nadaf" ]	[ "cs.LG", "cs.AI" ]	[ "Computer Science" ]	2026-04-28T00:00:00	https://arxiv.org/abs/2604.26130	https://arxiv.org/pdf/2604.26130v1	2604.26130	null	1	0	true	https://github.com/suhailnadaf509/reward-lens	null	0.0753
63635130d3555443a70bf84e7d495b873b0aedf642003d4b874ffb8150cfe3ba	[ "arxiv", "semantic_scholar" ]	SaliencyDecor: Enhancing Neural Network Interpretability through Feature Decorrelation	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 ...	[ "Ali Karkehabadi", "Jamshid Hassanpour", "Houman Homayoun", "Avesta Sasan" ]	[ "cs.CV" ]	[ "Computer Science" ]	2026-04-28T00:00:00	https://arxiv.org/abs/2604.25315	https://arxiv.org/pdf/2604.25315v1	2604.25315	null	0	0	false	null	null	0
c279c3def8e4f783458d5a6cb34044def5256c9dcf1be5be79a167c1d33a73dd	[ "arxiv", "semantic_scholar" ]	Why Does Reinforcement Learning Generalize? A Feature-Level Mechanistic Study of Post-Training in Large Language Models	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...	[ "Dan Shi", "Zhuowen Han", "Simon Ostermann", "Renren Jin", "Josef van Genabith", "Deyi Xiong" ]	[ "cs.CL" ]	[ "Computer Science" ]	2026-04-27T00:00:00	https://arxiv.org/abs/2604.25011	https://arxiv.org/pdf/2604.25011v1	2604.25011	null	1	0	true	https://github.com/danshi777/RL-generalization	null	0.0753
bd97d46f8e0e7aa606f3144d5b8e63f9eade29983e9efdca45cc242332b40fcc	[ "arxiv", "semantic_scholar" ]	Domain-Filtered Knowledge Graphs from Sparse Autoencoder Features	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 relationsh...	[ "John Winnicki", "Abeynaya Gnanasekaran", "Eric Darve" ]	[ "cs.AI" ]	[ "Computer Science" ]	2026-04-26T00:00:00	https://arxiv.org/abs/2604.23829	https://arxiv.org/pdf/2604.23829v2	2604.23829	null	0	0	false	null	null	0
44cc1ad7f40f9bb77d1b4173c0a14ed3ee0a2596f1bdecf4fbcf33d4d3f6e434	[ "arxiv", "semantic_scholar" ]	AIPsy-Affect: A Keyword-Free Clinical Stimulus Battery for Mechanistic Interpretability of Emotion in Language Models	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 un...	[ "Michael Keeman" ]	[ "cs.CL", "cs.AI" ]	[ "Computer Science" ]	2026-04-26T00:00:00	https://arxiv.org/abs/2604.23719	https://arxiv.org/pdf/2604.23719v2	2604.23719	null	0	0	false	null	null	0
e112e28705a76d60e410b4c930d0dad8c5d217da549c2c229b1cc240f351b14a	[ "arxiv", "semantic_scholar" ]	Preference Heads in Large Language Models: A Mechanistic Framework for Interpretable Personalization	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 Pre...	[ "Weixu Zhang", "Ye Yuan", "Changjiang Han", "Yuxing Tian", "Zipeng Sun", "Linfeng Du", "Jikun Kang", "Hong Kang", "Xue Liu", "Haolun Wu" ]	[ "cs.CL" ]	[ "Computer Science" ]	2026-04-24T00:00:00	https://arxiv.org/abs/2604.22345	https://arxiv.org/pdf/2604.22345v1	2604.22345	null	2	0	false	null	null	0.1193
42cf1e8701e9a68dcb7e21b0e89a8943ac615215baf4af6efc98cb2405ba169c	[ "arxiv", "semantic_scholar" ]	On the Properties of Feature Attribution for Supervised Contrastive Learning	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 pro...	[ "Leonardo Arrighi", "Julia Eva Belloni", "Aurélie Gallet", "Ivan Gentile", "Matteo Lippi", "Marco Zullich" ]	[ "cs.LG", "cs.AI" ]	[ "Computer Science" ]	2026-04-24T00:00:00	https://arxiv.org/abs/2604.22540	https://arxiv.org/pdf/2604.22540v1	2604.22540	null	0	0	false	null	null	0
bdb368fea779342584afc71cc7f6b1b5329f38d41b14cc775ec899c269a57013	[ "arxiv", "semantic_scholar" ]	Locating acts of mechanistic reasoning in student team conversations with mechanistic machine learning	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 ...	[ "Kaitlin Gili", "Mainak Nistala", "Kristen Wendell", "Michael C. Hughes" ]	[ "physics.ed-ph", "cs.LG" ]	[ "Physics", "Computer Science" ]	2026-04-23T00:00:00	https://arxiv.org/abs/2604.21870	https://arxiv.org/pdf/2604.21870v1	2604.21870	null	0	0	false	null	null	0

End of preview. Expand in Data Studio

Mechanistic Interpretability Papers — FineSet

A research-paper dataset on Mechanistic Interpretability Papers, assembled, deduplicated, and quality-scored by FineSet from arXiv and Semantic Scholar.

📸 This is a dated snapshot — generated 2026-06-12. It is not auto-updated. Research on Mechanistic Interpretability Papers moves fast — new papers land on arXiv every week. Want this same dataset refreshed daily, on a topic you choose? See the bottom. ↓

Why this dataset

Quality-scored: quality_score float (0–1), citation-normalized — filter out the noise
Papers with code: 133 flagged via has_code — find reproducible work fast
Deduplicated: arXiv + Semantic Scholar cross-referenced, duplicate records merged
Clean JSONL: 748 records, one per line, normalized fields — no encoding garbage

Dataset details

Records: 748
Date range: 2022–2026
Snapshot date: 2026-06-12 (frozen — see note above)
Sources: arXiv, Semantic Scholar (cross-referenced, duplicates merged)
arXiv categories: cs.LG, cs.AI
Quality scoring: citation-normalized, 0–1 (p50=0.119, p90=0.355)
Format: JSONL, one record per line

Fields

Field	Type	Description
id	string	Deterministic SHA256 record id
sources	list	Which sources contributed (`arxiv`, `semantic_scholar`)
title	string	Paper title
abstract	string	Full abstract
authors	list	Author names
categories	list	arXiv category codes
fields_of_study	list	Semantic Scholar field tags
published_date	string	ISO 8601 date
url	string	arXiv abstract URL
pdf_url	string\|null	Open-access PDF if available
arxiv_id	string\|null	arXiv identifier
doi	string\|null	DOI if available
citation_count	int	Citation count (Semantic Scholar)
influential_citation_count	int	Influential citations (Semantic Scholar)
has_code	bool	Code repo detected in the arXiv comment
code_url	string\|null	GitHub URL if detected
venue	string\|null	Publication venue
quality_score	float	0–1, citation-normalized

Quality score methodology

quality_score = min(1.0, log10(citation_count + 1) / 4)

A citation-normalized heuristic: 0 for uncited papers, ~0.5 at 100 citations, ~0.75 at 1,000, 1.0 at 10,000+. Useful for filtering training data by impact.

👉 Want this on YOUR topic, updated daily?

This snapshot is frozen at 2026-06-12. The live FineSet pipeline keeps a dataset like this refreshed every day on whatever topic you describe — new papers in, dedup and quality scoring automatic, export as JSONL/Parquet or push straight to the Hub.

Tell me the topic you'd want and I'll run the pipeline on it — open a discussion on this dataset, it's free and it's how I decide what to build next.

→ fineset.io — describe what you want to train on, get a dataset. Early-access waitlist open (referral skip available).

Downloads last month: 68