Daily Papers

Large language models (LLMs) are increasingly deployed in domains requiring guardrails to detect unsafe, off-topic, or adversarial prompts. Existing guardrails predominantly rely on fine-tuning to build classifiers, which often suffer from low generalization and high inference latency. We present kNNGuard, a training-free guardrail that utilizes the activation space of an off-the-shelf LLM. Given a small bank of 50 safe and unsafe prompts, kNNGuard extracts hidden activations and performs multi-layer kNN fusing activation-space and embedding-space scores for classification. Across six domains spanning topical and security prompts, kNNGuard achieves competitive or superior F1 compared to fine-tuned state-of-the-art guardrails while running 2.7x faster than the best comparable guardrail, and 10x faster than a fine-tuned safety classifier without gradient updates or fine-tuning. Domain adaptation requires only updating the labeled bank, which can be constructed in under 10 seconds and several orders of magnitude faster than established guardrails. We also analyze the impact of system prompts, layer selection, and integration into production LLM pipelines as a configurable, low-latency guardrail.

When a language model generates text, the selection of individual tokens might lead it down very different reasoning paths, making uncertainty difficult to quantify. In this work, we consider whether reasoning language models represent the alternate paths that they could take during generation. To test this hypothesis, we use hidden activations to control and predict a language model's uncertainty during chain-of-thought reasoning. In our experiments, we find a clear correlation between how uncertain a model is at different tokens, and how easily the model can be steered by controlling its activations. This suggests that activation interventions are most effective when there are alternate paths available to the model -- in other words, when it has not yet committed to a particular final answer. We also find that hidden activations can predict a model's future outcome distribution, demonstrating that models implicitly represent the space of possible paths.

Assessing the quality of Large Language Model (LLM) outputs presents a critical challenge. Previous methods either rely on text-level information (e.g., reward models, majority voting), which can overfit to superficial cues, or on calibrated confidence from token probabilities, which would fail on less-calibrated models. Yet both of these signals are, in fact, partial projections of a richer source of information: the model's internal hidden states. Early layers, closer to token embeddings, preserve semantic and lexical features that underpin text-based judgments, while later layers increasingly align with output logits, embedding confidence-related information. This paper explores hidden states directly as a unified foundation for verification. We show that the correctness of a solution is encoded as a geometrically separable signature within the trajectory of hidden activations. To validate this, we present Clue (Clustering and Experience-based Verification), a deliberately minimalist, non-parametric verifier. With no trainable parameters, CLUE only summarizes each reasoning trace by an hidden state delta and classifies correctness via nearest-centroid distance to ``success'' and ``failure'' clusters formed from past experience. The simplicity of this method highlights the strength of the underlying signal. Empirically, CLUE consistently outperforms LLM-as-a-judge baselines and matches or exceeds modern confidence-based methods in reranking candidates, improving both top-1 and majority-vote accuracy across AIME 24/25 and GPQA. As a highlight, on AIME 24 with a 1.5B model, CLUE boosts accuracy from 56.7% (majority@64) to 70.0% (top-maj@16).

Recent advances in neural text-to-speech research have been dominated by two-stage pipelines utilizing low-level intermediate speech representation such as mel-spectrograms. However, such predetermined features are fundamentally limited, because they do not allow to exploit the full potential of a data-driven approach through learning hidden representations. For this reason, several end-to-end methods have been proposed. However, such models are harder to train and require a large number of high-quality recordings with transcriptions. Here, we propose WavThruVec - a two-stage architecture that resolves the bottleneck by using high-dimensional Wav2Vec 2.0 embeddings as intermediate speech representation. Since these hidden activations provide high-level linguistic features, they are more robust to noise. That allows us to utilize annotated speech datasets of a lower quality to train the first-stage module. At the same time, the second-stage component can be trained on large-scale untranscribed audio corpora, as Wav2Vec 2.0 embeddings are already time-aligned. This results in an increased generalization capability to out-of-vocabulary words, as well as to a better generalization to unseen speakers. We show that the proposed model not only matches the quality of state-of-the-art neural models, but also presents useful properties enabling tasks like voice conversion or zero-shot synthesis.

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.

Large language models (LLMs) excel at generating fluent text, but their internal reasoning remains opaque and difficult to control. Sparse autoencoders (SAEs) make hidden activations more interpretable by exposing latent features that often align with human concepts. Yet, these features are fragile and passive, offering no mechanism for systematic reasoning or model control. To address this, we introduce ActivationReasoning (AR), a framework that embeds explicit logical reasoning into the latent space of LLMs. It proceeds in three stages: (1) Finding latent representations, first latent concept representations are identified (e.g., via SAEs) and organized into a dictionary; (2) Activating propositions, at inference time AR detects activating concepts and maps them to logical propositions; and (3)Logical reasoning, applying logical rules over these propositions to infer higher-order structures, compose new concepts, and steer model behavior. We evaluate AR on multi-hop reasoning (PrOntoQA), abstraction and robustness to indirect concept cues (Rail2Country), reasoning over natural and diverse language (ProverQA), and context-sensitive safety (BeaverTails). Across all tasks, AR scales robustly with reasoning complexity, generalizes to abstract and context-sensitive tasks, and transfers across model backbones. These results demonstrate that grounding logical structure in latent activations not only improves transparency but also enables structured reasoning, reliable control, and alignment with desired behaviors, providing a path toward more reliable and auditable AI.

Modern zero-shot text-to-speech (TTS) models offer unprecedented expressivity but also pose serious crime risks, as they can synthesize voices of individuals who never consented. In this context, speaker unlearning aims to prevent the generation of specific speaker identities upon request. Existing approaches, reliant on retraining, are costly and limited to speakers seen in the training set. We present TruS, a training-free speaker unlearning framework that shifts the paradigm from data deletion to inference-time control. TruS steers identity-specific hidden activations to suppress target speakers while preserving other attributes (e.g., prosody and emotion). Experimental results show that TruS effectively prevents voice generation on both seen and unseen opt-out speakers, establishing a scalable safeguard for speech synthesis. The demo and code are available on http://mmai.ewha.ac.kr/trus.

Structured deep model compression methods are hardware-friendly and substantially reduce memory and inference costs. However, under aggressive compression, the resulting accuracy degradation often necessitates post-compression finetuning, which can be impractical due to missing labeled data or high training cost. We propose post-hoc blockwise compensation, called GRAIL, a simple zero-finetuning step applied after model compression that restores each block's input-output behavior using a small calibration set. The method summarizes hidden activations via a Gram matrix and applies ridge regression to linearly reconstruct the original hidden representation from the reduced one. The resulting reconstruction map is absorbed into the downstream projection weights, while the upstream layer is compressed. The approach is selector-agnostic (Magnitude, Wanda, Gram-based selection, or folding), data-aware (requiring only a few forward passes without gradients or labels), and recovers classic pruning or folding when the Gram matrix is near identity, indicating weak inter-channel correlations. Across ResNets, ViTs, and decoder-only LLMs, GRAIL consistently improves accuracy or perplexity over data-free and data-aware pruning or folding baselines in practical compression regimes, with manageable overhead and no backpropagation. The code is available at https://github.com/TWWinde/GRAIL_Compensation.

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.

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 defined 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 relations 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.

This thesis addresses two persistent and closely related challenges in modern deep learning, reliability and efficiency, through a unified framework grounded in Spectral Geometry and Random Matrix Theory (RMT). As deep networks and large language models continue to scale, their internal behavior becomes increasingly opaque, leading to hallucinations, fragile generalization under distribution shift, and growing computational and energy demands. By analyzing the eigenvalue dynamics of hidden activations across layers and inputs, this work shows that spectral statistics provide a compact, stable, and interpretable lens on model behavior, capable of separating structured, causal representations from noise-dominated variability. Within this framework, the first contribution, EigenTrack, introduces a real-time method for detecting hallucinations and out-of-distribution behavior in large language and vision-language models. EigenTrack transforms streaming activations into spectral descriptors such as entropy, variance, and deviations from the Marchenko-Pastur baseline, and models their temporal evolution using lightweight recurrent classifiers, enabling early detection of reliability failures before they appear in model outputs while offering interpretable insight into representation dynamics. The second contribution, RMT-KD, presents a principled approach to compressing deep networks via random matrix theoretic knowledge distillation. By interpreting outlier eigenvalues in activation spectra as carriers of task-relevant information, RMT-KD progressively projects networks onto lower-dimensional subspaces through iterative self-distillation, yielding significantly more compact and energy-efficient models while preserving accuracy and dense, hardware-friendly structure.

Natural-language autoencoders score explanations of hidden activations by reconstruction: an explanation is deemed faithful if the activation can be regenerated from it. The test is structurally insensitive to individual false claims: if flipping a claim does not change the reconstruction, the claim is never penalized. We show the test is passed in two ways, neither faithful. On a released Qwen-2.5-7B verbalizer, explanations reconstruct well above chance while ~2% of specific claims are reconstruction-dependent, so the score tracks gist, not specific facts. Under exact synthetic ground truth, the standard recipe develops co-adapted private codes (false wording the reconstruction depends on) in 5/5 runs, and fixes that leave the target model unchanged do not help. We contribute two audit protocols, the grounded-vs-true cross and the evaluator swap, and RECAP (Readable Encodings via Co-trained Auxiliary Predictors): linear heads trained alongside the target model to keep designated content decodable. On RECAP-trained sandbox models, fresh verbalizers state the designated content truly and the codes vanish, at a +0.001-nat cost. This replicates on a pretrained Pythia-160M: the content becomes reliably probe-decodable, though a fresh verbalizer conveys it only in part (truth 0.44-0.46 vs a near-zero control). For interpretability, high reconstruction does not certify individual claims. For AI safety, RECAP makes designated internal content independently checkable against probes rather than asserted by prose a model can game: an independent probe scores the verbalizer's true claims above its false ones (AUC 0.96, vs 0.82 without RECAP). Against an adversary that edits an explanation to maximize the reconstruction score while lying (suppressing ~87% of its lie penalty), the RECAP probe still flags the lies (AUC 0.95) while the control probe collapses to chance (0.51).

Online Normalization is a new technique for normalizing the hidden activations of a neural network. Like Batch Normalization, it normalizes the sample dimension. While Online Normalization does not use batches, it is as accurate as Batch Normalization. We resolve a theoretical limitation of Batch Normalization by introducing an unbiased technique for computing the gradient of normalized activations. Online Normalization works with automatic differentiation by adding statistical normalization as a primitive. This technique can be used in cases not covered by some other normalizers, such as recurrent networks, fully connected networks, and networks with activation memory requirements prohibitive for batching. We show its applications to image classification, image segmentation, and language modeling. We present formal proofs and experimental results on ImageNet, CIFAR, and PTB datasets.

Inference-time steering methods offer a lightweight alternative to fine-tuning large language models (LLMs) and vision-language models (VLMs) by modifying internal activations at test time without updating model weights. However, most existing approaches rely on fixed, global intervention vectors, overlook the causal influence of individual input tokens, and fail to leverage informative gradients from the model's logits, particularly in multimodal settings where visual and textual inputs contribute unevenly. To address these limitations, we introduce GrAInS, an inference-time steering approach that operates across both language-only and vision-language models and tasks. GrAInS uses contrastive, gradient-based attribution via Integrated Gradients to identify the top-k most influential tokens, both positively and negatively attributed based on their contribution to preferred versus dispreferred outputs. These tokens are then used to construct directional steering vectors that capture semantic shifts from undesirable to desirable behavior. During inference, GrAInS adjusts hidden activations at transformer layers guided by token-level attribution signals, and normalizes activations to preserve representational scale. This enables fine-grained, interpretable, and modular control over model behavior, without retraining or auxiliary supervision. Empirically, GrAInS consistently outperforms both fine-tuning and existing steering baselines: it achieves a 13.22% accuracy gain on TruthfulQA using Llama-3.1-8B, reduces hallucination rates on MMHal-Bench from 0.624 to 0.514 with LLaVA-1.6-7B, and improves alignment win rates on SPA-VL by 8.11%, all while preserving the model's fluency and general capabilities.

Safety alignment is a key requirement for building reliable Artificial General Intelligence. Despite significant advances in safety alignment, we observe that minor latent shifts can still trigger unsafe responses in aligned models. We argue that this stems from the shallow nature of existing alignment methods, which focus on surface-level refusal behaviors without sufficiently altering internal representations. Consequently, small shifts in hidden activations can re-trigger harmful behaviors embedded in the latent space. To explore the robustness of safety alignment to latent perturbations, we introduce a probing method that measures the Negative Log-Likelihood of the original response generated by the model. This probe quantifies local sensitivity in the latent space, serving as a diagnostic tool for identifying vulnerable directions. Based on this signal, we construct effective jailbreak trajectories, giving rise to the Activation Steering Attack (ASA). More importantly, these insights offer a principled foundation for improving alignment robustness. To this end, we introduce Layer-wise Adversarial Patch Training~(LAPT), a fine-tuning strategy that inject controlled perturbations into hidden representations during training. Experimental results highlight that LAPT strengthen alignment robustness without compromising general capabilities. Our findings reveal fundamental flaws in current alignment paradigms and call for representation-level training strategies that move beyond surface-level behavior supervision. Codes and results are available at https://github.com/Carol-gutianle/LatentSafety.

Large Language Models (LLMs) often produce fluent yet factually incorrect statements-a phenomenon known as hallucination-posing serious risks in high-stakes domains. We present Layer-wise Semantic Dynamics (LSD), a geometric framework for hallucination detection that analyzes the evolution of hidden-state semantics across transformer layers. Unlike prior methods that rely on multiple sampling passes or external verification sources, LSD operates intrinsically within the model's representational space. Using margin-based contrastive learning, LSD aligns hidden activations with ground-truth embeddings derived from a factual encoder, revealing a distinct separation in semantic trajectories: factual responses preserve stable alignment, while hallucinations exhibit pronounced semantic drift across depth. Evaluated on the TruthfulQA and synthetic factual-hallucination datasets, LSD achieves an F1-score of 0.92, AUROC of 0.96, and clustering accuracy of 0.89, outperforming SelfCheckGPT and Semantic Entropy baselines while requiring only a single forward pass. This efficiency yields a 5-20x speedup over sampling-based methods without sacrificing precision or interpretability. LSD offers a scalable, model-agnostic mechanism for real-time hallucination monitoring and provides new insights into the geometry of factual consistency within large language models.

This work introduces RevSilo, the first reversible bidirectional multi-scale feature fusion module. Like other reversible methods, RevSilo eliminates the need to store hidden activations by recomputing them. However, existing reversible methods do not apply to multi-scale feature fusion and are, therefore, not applicable to a large class of networks. Bidirectional multi-scale feature fusion promotes local and global coherence and has become a de facto design principle for networks targeting spatially sensitive tasks, e.g., HRNet (Sun et al., 2019a) and EfficientDet (Tan et al., 2020). These networks achieve state-of-the-art results across various computer vision tasks when paired with high-resolution inputs. However, training them requires substantial accelerator memory for saving large, multi-resolution activations. These memory requirements inherently cap the size of neural networks, limiting improvements that come from scale. Operating across resolution scales, RevSilo alleviates these issues. Stacking RevSilos, we create RevBiFPN, a fully reversible bidirectional feature pyramid network. RevBiFPN is competitive with networks such as EfficientNet while using up to 19.8x lesser training memory for image classification. When fine-tuned on MS COCO, RevBiFPN provides up to a 2.5% boost in AP over HRNet using fewer MACs and a 2.4x reduction in training-time memory.

Post-Training Quantization (PTQ) enhances the efficiency of Large Language Models (LLMs) by enabling faster operation and compatibility with more accessible hardware through reduced memory usage, at the cost of small performance drops. We explore the role of calibration sets in PTQ, specifically their effect on hidden activations in various notable open-source LLMs. Calibration sets are crucial for evaluating activation magnitudes and identifying outliers, which can distort the quantization range and negatively impact performance. Our analysis reveals a marked contrast in quantization effectiveness across models. The older OPT model, upon which much of the quantization literature is based, shows significant performance deterioration and high susceptibility to outliers with varying calibration sets. In contrast, newer models like Llama-2 7B, Llama-3 8B, Command-R 35B, and Mistral 7B demonstrate strong robustness, with Mistral 7B showing near-immunity to outliers and stable activations. These findings suggest a shift in PTQ strategies might be needed. As advancements in pre-training methods reduce the relevance of outliers, there is an emerging need to reassess the fundamentals of current quantization literature. The emphasis should pivot towards optimizing inference speed, rather than primarily focusing on outlier preservation, to align with the evolving characteristics of state-of-the-art LLMs.

Large transformer-based language models (LMs) trained on huge text corpora have shown unparalleled generation capabilities. However, controlling attributes of the generated language (e.g. switching topic or sentiment) is difficult without modifying the model architecture or fine-tuning on attribute-specific data and entailing the significant cost of retraining. We propose a simple alternative: the Plug and Play Language Model (PPLM) for controllable language generation, which combines a pretrained LM with one or more simple attribute classifiers that guide text generation without any further training of the LM. In the canonical scenario we present, the attribute models are simple classifiers consisting of a user-specified bag of words or a single learned layer with 100,000 times fewer parameters than the LM. Sampling entails a forward and backward pass in which gradients from the attribute model push the LM's hidden activations and thus guide the generation. Model samples demonstrate control over a range of topics and sentiment styles, and extensive automated and human annotated evaluations show attribute alignment and fluency. PPLMs are flexible in that any combination of differentiable attribute models may be used to steer text generation, which will allow for diverse and creative applications beyond the examples given in this paper.

Controlling multiple behavioral attributes in large language models (LLMs) at inference time is a challenging problem due to interference between attributes and the limitations of linear steering methods, which assume additive behavior in activation space and require per-attribute tuning. We introduce K-Steering, a unified and flexible approach that trains a single non-linear multi-label classifier on hidden activations and computes intervention directions via gradients at inference time. This avoids linearity assumptions, removes the need for storing and tuning separate attribute vectors, and allows dynamic composition of behaviors without retraining. To evaluate our method, we propose two new benchmarks, ToneBank and DebateMix, targeting compositional behavioral control. Empirical results across 3 model families, validated by both activation-based classifiers and LLM-based judges, demonstrate that K-Steering outperforms strong baselines in accurately steering multiple behaviors.

Transformer components such as non-linear activations and normalization are inherently non-injective, suggesting that different inputs could map to the same output and prevent exact recovery of the input from a model's representations. In this paper, we challenge this view. First, we prove mathematically that transformer language models mapping discrete input sequences to their corresponding sequence of continuous representations are injective and therefore lossless, a property established at initialization and preserved during training. Second, we confirm this result empirically through billions of collision tests on six state-of-the-art language models, and observe no collisions. Third, we operationalize injectivity: we introduce SipIt, the first algorithm that provably and efficiently reconstructs the exact input text from hidden activations, establishing linear-time guarantees and demonstrating exact invertibility in practice. Overall, our work establishes injectivity as a fundamental and exploitable property of language models, with direct implications for transparency, interpretability, and safe deployment.

Large language models (LLMs) frequently achieve impressive scores on standardized benchmarks, yet accuracy alone offers a limited view of their capabilities. Evaluating open-source LLMs through leaderboards faces persistent issues like data contamination, narrow task scope, and weak alignment with real-world reliability. Benchmark-based evaluations such as MMLU PRO, BBH, or IFEval primarily capture what a model outputs on fixed test sets, not how it processes information, calibrates uncertainty, or structures internal knowledge. In this article, we advocate for a shift from benchmark-centric evaluation toward a complementary, state-centered intrinsic assessment of LLMs. To this end, we introduce Latent Performance Profiling (LPP) -- a framework that derives task-agnostic diagnostics from hidden activations and output distributions. LPP defines a set of scalar metrics on a model's latent representations and dynamics, revealing scale-independent traits that enable interpretable comparisons and uncover hidden vulnerabilities. Unlike static accuracy scores, LPP provides stable, architecture-sensitive signatures across models of similar size. With extensive empirical analyses across eight LLMs, spanning a size range of 0.5B-14B, we demonstrate that models with similar benchmark scores can exhibit contrasting latent profiles, such as differences in entropy or adaptability. Guided by these insights, we design synthetic probes for uncertainty and symbolic reasoning that align with intrinsic metrics while decoupling from leaderboard bias. We recommend that reporting LPP alongside benchmarks provides a deeper, interpretable understanding of model behavior, enabling more reliable model selection, safety assessment, and evaluation beyond surface-level accuracy.

Transformer language models (LMs) have been shown to represent concepts as directions in the latent space of hidden activations. However, for any given human-interpretable concept, how can we find its direction in the latent space? We present a technique called linear relational concepts (LRC) for finding concept directions corresponding to human-interpretable concepts at a given hidden layer in a transformer LM by first modeling the relation between subject and object as a linear relational embedding (LRE). While the LRE work was mainly presented as an exercise in understanding model representations, we find that inverting the LRE while using earlier object layers results in a powerful technique to find concept directions that both work well as a classifier and causally influence model outputs.

The self-attention mechanism utilizes large implicit weight matrices, programmed through dot product-based activations with very few trainable parameters, to enable long sequence modeling. In this paper, we investigate the possibility of discarding residual learning by employing large implicit kernels to achieve full context interaction at each layer of the network. To accomplish it, we introduce coordinate-based implicit MLPs as a slow network to generate hyper-kernels for another fast convolutional network. To get context-varying weights for fast dynamic encoding, we propose a HyperZ{cdotZ{cdot}W} operator that connects hyper-kernels (W) and hidden activations (Z) through simple elementwise multiplication, followed by convolution of Z using the context-dependent W. Based on this design, we present a novel Terminator architecture that integrates hyper-kernels of different sizes to produce multi-branch hidden representations for enhancing the feature extraction capability of each layer. Additionally, a bottleneck layer is employed to compress the concatenated channels, allowing only valuable information to propagate to the subsequent layers. Notably, our model incorporates several innovative components and exhibits excellent properties, such as introducing local feedback error for updating the slow network, stable zero-mean features, faster training convergence, and fewer model parameters. Extensive experimental results on pixel-level 1D and 2D image classification benchmarks demonstrate the superior performance of our architecture.

Large language models (LLMs) exhibit distinct and consistent personalities that greatly impact trust and engagement. While this means that personality frameworks would be highly valuable tools to characterize and control LLMs' behavior, current approaches remain either costly (post-training) or brittle (prompt engineering). Probing and steering via linear directions has recently emerged as a cheap and efficient alternative. In this paper, we investigate whether linear directions aligned with the Big Five personality traits can be used for probing and steering model behavior. Using Llama 3.3 70B, we generate descriptions of 406 fictional characters and their Big Five trait scores. We then prompt the model with these descriptions and questions from the Alpaca questionnaire, allowing us to sample hidden activations that vary along personality traits in known, quantifiable ways. Using linear regression, we learn a set of per-layer directions in activation space, and test their effectiveness for probing and steering model behavior. Our results suggest that linear directions aligned with trait-scores are effective probes for personality detection, while their steering capabilities strongly depend on context, producing reliable effects in forced-choice tasks but limited influence in open-ended generation or when additional context is present in the prompt.

The alignment of vision-language representations endows current Vision-Language Models (VLMs) with strong multi-modal reasoning capabilities. However, the interpretability of the alignment component remains uninvestigated due to the difficulty in mapping the semantics of multi-modal representations into a unified concept set. To address this problem, we propose VL-SAE, a sparse autoencoder that encodes vision-language representations into its hidden activations. Each neuron in its hidden layer correlates to a concept represented by semantically similar images and texts, thereby interpreting these representations with a unified concept set. To establish the neuron-concept correlation, we encourage semantically similar representations to exhibit consistent neuron activations during self-supervised training. First, to measure the semantic similarity of multi-modal representations, we perform their alignment in an explicit form based on cosine similarity. Second, we construct the VL-SAE with a distance-based encoder and two modality-specific decoders to ensure the activation consistency of semantically similar representations. Experiments across multiple VLMs (e.g., CLIP, LLaVA) demonstrate the superior capability of VL-SAE in interpreting and enhancing the vision-language alignment. For interpretation, the alignment between vision and language representations can be understood by comparing their semantics with concepts. For enhancement, the alignment can be strengthened by aligning vision-language representations at the concept level, contributing to performance improvements in downstream tasks, including zero-shot image classification and hallucination elimination. Codes are available at https://github.com/ssfgunner/VL-SAE.

Large language models and deep neural networks achieve strong performance but suffer from reliability issues and high computational cost. This thesis proposes a unified framework based on spectral geometry and random matrix theory to address both problems by analyzing the eigenvalue structure of hidden activations. The first contribution, EigenTrack, is a real-time method for detecting hallucinations and out-of-distribution behavior in language and vision-language models using spectral features and their temporal dynamics. The second contribution, RMT-KD, is a principled compression method that identifies informative spectral components and applies iterative knowledge distillation to produce compact and efficient models while preserving accuracy. Together, these results show that spectral statistics provide interpretable and robust signals for monitoring uncertainty and guiding compression in large-scale neural networks.

Recent developments in Large Language Model (LLM) capabilities have brought great potential but also posed new risks. For example, LLMs with knowledge of bioweapons, advanced chemistry, or cyberattacks could cause violence if placed in the wrong hands or during malfunctions. Because of their nature as near-black boxes, intuitive interpretation of LLM internals remains an open research question, preventing developers from easily controlling model behavior and capabilities. The use of Sparse Autoencoders (SAEs) has recently emerged as a potential method of unraveling representations of concepts in LLMs internals, and has allowed developers to steer model outputs by directly modifying the hidden activations. In this paper, we use SAEs to identify unwanted concepts from the Weapons of Mass Destruction Proxy (WMDP) dataset within gemma-2-2b internals and use feature steering to reduce the model's ability to answer harmful questions while retaining its performance on harmless queries. Our results bring back optimism to the viability of SAE-based explicit knowledge unlearning techniques.

Can neural networks systematically capture discrete, compositional task structure despite their continuous, distributed nature? The impressive capabilities of large-scale neural networks suggest that the answer to this question is yes. However, even for the most capable models, there are still frequent failure cases that raise doubts about their compositionality. Here, we seek to understand what it takes for a standard neural network to generalize over tasks that share compositional structure. We find that simply scaling data and model size leads to compositional generalization. We show that this holds across different task encodings as long as the training distribution sufficiently covers the task space. In line with this finding, we prove that standard multilayer perceptrons can approximate a general class of compositional task families to arbitrary precision using only a linear number of neurons with respect to the number of task modules. Finally, we uncover that if networks successfully compositionally generalize, the constituents of a task can be linearly decoded from their hidden activations. We show that this metric correlates with failures of text-to-image generation models to compose known concepts.

While Large Language Models (LLMs) have shown exceptional performance in various tasks, their (arguably) most prominent drawback is generating inaccurate or false information with a confident tone. In this paper, we hypothesize that the LLM's internal state can be used to reveal the truthfulness of a statement. Therefore, we introduce a simple yet effective method to detect the truthfulness of LLM-generated statements, which utilizes the LLM's hidden layer activations to determine the veracity of statements. To train and evaluate our method, we compose a dataset of true and false statements in six different topics. A classifier is trained to detect which statement is true or false based on an LLM's activation values. Specifically, the classifier receives as input the activation values from the LLM for each of the statements in the dataset. Our experiments demonstrate that our method for detecting statement veracity significantly outperforms even few-shot prompting methods, highlighting its potential to enhance the reliability of LLM-generated content and its practical applicability in real-world scenarios.

The lack of interpretability of the Vision Transformer may hinder its use in critical real-world applications despite its effectiveness. To overcome this issue, we propose a post-hoc interpretability method called VISION DIFFMASK, which uses the activations of the model's hidden layers to predict the relevant parts of the input that contribute to its final predictions. Our approach uses a gating mechanism to identify the minimal subset of the original input that preserves the predicted distribution over classes. We demonstrate the faithfulness of our method, by introducing a faithfulness task, and comparing it to other state-of-the-art attribution methods on CIFAR-10 and ImageNet-1K, achieving compelling results. To aid reproducibility and further extension of our work, we open source our implementation: https://github.com/AngelosNal/Vision-DiffMask

Alignment tuning is meant to make harmful-request refusal robust, yet this safety behavior can be erased by a small set of benign fine-tuning examples. This is a deployment risk for open-weight models because a checkpoint can pass refusal tests at release time and later lose refusal under low-cost downstream fine-tuning. Prior work has established these refusal failures, but existing studies do not show how to detect this fragility in the aligned model itself before an attack or fine-tuning intervention is run. We introduce Skin-Deep, a geometric diagnostic that detects alignment fragility directly from the aligned model's hidden-state activations before such an intervention is run and compresses the layer-wise safety geometry into a single scalar, the Geometric Fragility Score (GFS). Applied to twenty-one instruction-tuned models spanning six alignment recipes and 3B--32B parameters, Skin-Deep reveals a recurring low-rank safety subspace across model families. Direction ablations show that removing directions in this subspace weakens harmful-request refusal, providing causal evidence that the recovered geometry underlies refusal behavior. Crucially, GFS identifies, before any fine-tuning, the initially safe model that retains the most refusal after small-scale LoRA fine-tuning. These results establish GFS as a practical pre-deployment diagnostic for flagging fragile refusal behavior without running an attack.

Large language models (LLMs) are trained on massive web-scale corpora, raising growing concerns about privacy and copyright. Membership inference attacks (MIAs) aim to determine whether a given example was used during training. Existing LLM MIAs largely rely on output probabilities or loss values and often perform only marginally better than random guessing when members and non-members are drawn from the same distribution. We introduce G-Drift MIA, a white-box membership inference method based on gradient-induced feature drift. Given a candidate (x,y), we apply a single targeted gradient-ascent step that increases its loss and measure the resulting changes in internal representations, including logits, hidden-layer activations, and projections onto fixed feature directions, before and after the update. These drift signals are used to train a lightweight logistic classifier that effectively separates members from non-members. Across multiple transformer-based LLMs and datasets derived from realistic MIA benchmarks, G-Drift substantially outperforms confidence-based, perplexity-based, and reference-based attacks. We further show that memorized training samples systematically exhibit smaller and more structured feature drift than non-members, providing a mechanistic link between gradient geometry, representation stability, and memorization. In general, our results demonstrate that small, controlled gradient interventions offer a practical tool for auditing the membership of training-data and assessing privacy risks in LLMs.

In-context learning enables large language models (LLMs) to perform a variety of tasks, including learning to make reward-maximizing choices in simple bandit tasks. Given their potential use as (autonomous) decision-making agents, it is important to understand how these models perform such reinforcement learning (RL) tasks and the extent to which they are susceptible to biases. Motivated by the fact that, in humans, it has been widely documented that the value of an outcome depends on how it compares to other local outcomes, the present study focuses on whether similar value encoding biases apply to how LLMs encode rewarding outcomes. Results from experiments with multiple bandit tasks and models show that LLMs exhibit behavioral signatures of a relative value bias. Adding explicit outcome comparisons to the prompt produces opposing effects on performance, enhancing maximization in trained choice sets but impairing generalization to new choice sets. Computational cognitive modeling reveals that LLM behavior is well-described by a simple RL algorithm that incorporates relative values at the outcome encoding stage. Lastly, we present preliminary evidence that the observed biases are not limited to fine-tuned LLMs, and that relative value processing is detectable in the final hidden layer activations of a raw, pretrained model. These findings have important implications for the use of LLMs in decision-making applications.

In-context learning (ICL) enables Large Language Models (LLMs) to adapt to new tasks using few examples, with task vectors - specific hidden state activations - hypothesized to encode task information. Existing studies are limited by small-scale benchmarks, restricting comprehensive analysis. We introduce QuiteAFew, a novel dataset of 3,096 diverse few-shot tasks, each with 30 input-output pairs derived from the Alpaca dataset. Experiments with Llama-3-8B on QuiteAFew reveal: (1) task vector performance peaks at an intermediate layer (e.g., 15th), (2) effectiveness varies significantly by task type, and (3) complex tasks rely on multiple, subtask-specific vectors rather than a single vector, suggesting distributed task knowledge representation.

Large Language Models have achieved remarkable success across various natural language processing tasks, yet their high computational cost during inference remains a major bottleneck. This paper introduces Sparse Expert Activation Pruning (SEAP), a training-free pruning method that selectively retains task-relevant parameters to reduce inference overhead. Inspired by the clustering patterns of hidden states and activations in LLMs, SEAP identifies task-specific expert activation patterns and prunes the model while preserving task performance and enhancing computational efficiency. Experimental results demonstrate that SEAP significantly reduces computational overhead while maintaining competitive accuracy. Notably, at 50% pruning, SEAP surpasses both WandA and FLAP by over 20%, and at 20% pruning, it incurs only a 2.2% performance drop compared to the dense model. These findings highlight SEAP's scalability and effectiveness, making it a promising approach for optimizing large-scale LLMs.

This research explores strategies for steering the output of large language models (LLMs) towards specific styles, such as sentiment, emotion, or writing style, by adding style vectors to the activations of hidden layers during text generation. We show that style vectors can be simply computed from recorded layer activations for input texts in a specific style in contrast to more complex training-based approaches. Through a series of experiments, we demonstrate the effectiveness of activation engineering using such style vectors to influence the style of generated text in a nuanced and parameterisable way, distinguishing it from prompt engineering. The presented research constitutes a significant step towards developing more adaptive and effective AI-empowered interactive systems.

Massive activations are scalar values in transformer hidden states that achieve values orders of magnitude larger than typical activations and have been shown to be critical for model functionality. While prior work has characterized these phenomena in fully trained models, the temporal dynamics of their emergence during training remain poorly understood. We present the first comprehensive analysis of massive activation development throughout transformer training, using the Pythia model family as our testbed. Through systematic analysis of various model sizes across multiple training checkpoints, we demonstrate that massive activation emergence follows predictable mathematical patterns that can be accurately modeled using an exponentially-modulated logarithmic function with five key parameters. We develop a machine learning framework to predict these mathematical parameters from architectural specifications alone, achieving high accuracy for steady-state behavior and moderate accuracy for emergence timing and magnitude. These findings enable architects to predict and potentially control key aspects of massive activation emergence through design choices, with significant implications for model stability, training cycle length, interpretability, and optimization. Our findings demonstrate that the emergence of massive activations is governed by model design and can be anticipated, and potentially controlled, before training begins.

We investigate the extent to which an LLM's hidden-state geometry can be recovered from its behavior in psycholinguistic experiments. Across eight instruction-tuned transformer models, we run two experimental paradigms -- similarity-based forced choice and free association -- over a shared 5,000-word vocabulary, collecting 17.5M+ trials to build behavior-based similarity matrices. Using representational similarity analysis, we compare behavioral geometries to layerwise hidden-state similarity and benchmark against FastText, BERT, and cross-model consensus. We find that forced-choice behavior aligns substantially more with hidden-state geometry than free association. In a held-out-words regression, behavioral similarity (especially forced choice) predicts unseen hidden-state similarities beyond lexical baselines and cross-model consensus, indicating that behavior-only measurements retain recoverable information about internal semantic geometry. Finally, we discuss implications for the ability of behavioral tasks to uncover hidden cognitive states.

Whisper, a widely adopted ASR model, is known to suffer from hallucinations - coherent transcriptions generated for non-speech audio entirely disconnected from the input. We investigate whether hallucinations can be detected and mitigated through Whisper's internal representations. We extract audio encoder activations and evaluate two representation spaces: raw Whisper activations and Sparse AutoEncoder (SAE) latents. We show that both spaces encode linearly separable hallucination-related information, with discriminative power concentrated in a sparse feature subset and increasing toward deeper encoder layers. We propose two steering strategies: activation-space steering and SAE latent-space steering. SAE-based steering reduces hallucination rate from 72.63% to 14.11% for Whisper small and from 86.88% to 27.33% for Whisper large-v3 on the full non-speech test set, with small WER degradation on speech data, approaching the performance of fine-tuning-based methods.

Deploying large language models in user-facing systems requires efficient output safety filtering. Existing approaches typically rely on a separate moderation model applied after generation, which doubles inference cost and only detects violations after generation completes. We observe that the signal needed for moderation is already present in the model hidden states. Based on this, we train lightweight token-level probes that operate directly on internal activations, producing per-token safety scores that can be aggregated for both offline evaluation and online intervention. The probe reuses activations from the generator and requires no additional forward pass, enabling sub millisecond per-token safety checks inside the decoding loop. A probe applied to a single mid layer recovers most decisions of a strong guard model, acting as a low cost surrogate optimized for latency rather than accuracy. In streaming settings, it can halt or modify unsafe outputs before they are fully generated, replacing end of sequence moderation with continuous token level monitoring. Compared to post hoc and streaming guard models, our method achieves orders of magnitude lower compute overhead with minimal latency cost. We also provide a practical deployment recipe, including layer selection, aggregation strategy, probing frequency, and triggering thresholds. Finally, we show that the probe linear component corresponds to a direction in residual space, enabling both detection and activation steering at negligible cost.

As LLMs are deployed as agents, reliable monitoring requires knowing not only what they output, but which instructions are steering their behavior. This is difficult when models infer unintended subgoals, follow contextual cues, or are influenced by prompt injections and hidden objectives. While activation-to-language methods suggest that hidden states can reveal natural-language information, existing approaches are not designed to recover the full set of simultaneous instructions, constraints, prohibitions, and subgoals active in agentic settings. We formalize this problem as instruction set retrieval and introduce PRISM, an activation-conditioned interpreter that decodes hidden states from a frozen target model into a faithful bullet list of active instructions. Unlike prior activation-to-language methods, PRISM is trained to recover instruction sets directly, using judge-guided GRPO to reward covered instructions and penalize unsupported ones. Across benign, constrained, prompt-injection, and hidden-objective settings, PRISM outperforms activation-to-language baselines, especially on security-relevant objectives.

Lens methods interpret large language models (LLMs) by mapping intermediate activations to the output vocabulary, revealing how next-token predictions develop through the network. Trained lenses remain expensive: affine-translator parameters grow quadratically with model width, while exact, full-vocabulary Kullback--Leibler (KL) training dominates memory. Consequently, prior trained lenses have been applied to models of at most 20B parameters and remain tied to particular component types. We present OmniLens, which applies a single lens family to any model-width activation, whether residual stream, attention, or MLP, and combines two independent scaling techniques. First, low-rank translators make per-lens parameter growth linear in model width and reduce trainable parameters by up to 98.4%. Second, Subset-KL materializes only selected vocabulary logits: its Top-k mode cuts peak training memory by up to 70%, while its importance-sampled variant retains unbiased stochastic gradients for the full KL. These savings enable a dense ensemble of 482 lenses for LLaMA-3.3-70B, providing 6x the coverage of a residual-stream design at the same depth. Model-wide coverage then reveals what single-component lenses cannot: the components where a behavior is most visible need not be those where intervention is most effective, and the most effective interventions lie outside the attention heads examined by prior lens studies. Across three case studies (prompt-injection detection, multi-hop memory injection, and toxicity localization), OmniLens reproduces key published results at substantially lower cost.

Language models can now prove theorems, but people still decide which problems to pursue. We ask whether a model's internal representations can help identify promising mathematical connections. We develop LANTERN, a fast, cost-efficient pipeline that uses a classifier over pretrained-model activations to rank candidate relations, followed by staged filtering, hypothesis generation, executable verification, and analytical checking. Applied to the On-Line Encyclopedia of Integer Sequences (OEIS), LANTERN ranked 50 million pairs among 10,000 frequently referenced sequences and produced 62 verified relations between pairs without an existing OEIS cross-reference. A content screen retained 13 relations worth presenting; nine of these are informative or insightful, including four which are entirely novel to the best of our knowledge: none appears in the OEIS or in our targeted literature search. The entire end-to-end process including classifier training, candidate ranking, filtering and verification took under 8 hours.

Large Vision-Language Models (LVLMs) can reason effectively over both textual and visual inputs, but they tend to hallucinate syntactically coherent yet visually ungrounded contents. In this paper, we investigate the internal dynamics of hallucination by examining the tokens logits rankings throughout the generation process, revealing three key patterns in how LVLMs process information: (1) gradual visual information loss -- visually grounded tokens gradually become less favored throughout generation, and (2) early excitation -- semantically meaningful tokens achieve peak activation in the layers earlier than the final layer. (3) hidden genuine information -- visually grounded tokens though not being eventually decided still retain relatively high rankings at inference. Based on these insights, we propose VISTA (Visual Information Steering with Token-logit Augmentation), a training-free inference-time intervention framework that reduces hallucination while promoting genuine information. VISTA works by combining two complementary approaches: reinforcing visual information in activation space and leveraging early layer activations to promote semantically meaningful decoding. Compared to existing methods, VISTA requires no external supervision and is applicable to various decoding strategies. Extensive experiments show that VISTA on average reduces hallucination by abount 40% on evaluated open-ended generation task, and it consistently outperforms existing methods on four benchmarks across four architectures under three decoding strategies.

We introduce softpick, a rectified, not sum-to-one, drop-in replacement for softmax in transformer attention mechanisms that eliminates attention sink and massive activations. Our experiments with 340M parameter models demonstrate that softpick maintains performance parity with softmax on standard benchmarks while achieving 0% sink rate. The softpick transformer produces hidden states with significantly lower kurtosis (340 vs 33,510) and creates sparse attention maps (46.97% sparsity). Models using softpick consistently outperform softmax when quantized, with particularly pronounced advantages at lower bit precisions. Our analysis and discussion shows how softpick has the potential to open new possibilities for quantization, low-precision training, sparsity optimization, pruning, and interpretability. Our code is available at https://github.com/zaydzuhri/softpick-attention.

The dynamic range of activations is a first-order constraint for low-bit quantization, activation scaling, and stable LLM inference. Prior work characterized outlier features and massive activations on pre-2024 LLaMA-style models, and the downstream activation-quantization stack inherits that picture without revisiting it for the post-LLaMA open-model boom. We ask the deployment-oriented question: how large can activations get in modern open LLMs, and how does this magnitude vary across families, generations, and training stages? Under a unified pipeline (5,000-sample multi-domain corpus, family-specific tokenization, identical hooks across embeddings, hidden states, attention, MLP/MoE, SwiGLU gates, and final norm), we measure global and layerwise maxima on 27 checkpoints from 8 open families spanning dense, MoE, vision-language, intermediate-training, and instruction-tuned variants. We find that (i) global maxima span over nearly four orders of magnitude at comparable parameter counts, with Qwen3.5 and MoE checkpoints in the 10^2 to 10^3 range and Gemma3-27B-it reaching ~7 x 10^5; (ii) cross-family and cross-generation comparisons break simple monotonic scaling; and (iii) MoE checkpoints exhibit 14.0-23.4x lower peaks than matched-scale dense counterparts, while the residual stream carries the global maximum in 22/24 checkpoints. A lightweight INT-8 sanity check shows that measured maxima co-vary with low-bit reconstruction error via activation-scale selection. We conclude that maximum activation magnitude is a model property tied to family, architecture, and training stage - not a simple byproduct of size - and should be measured and reported alongside any open-weight release before low-bit deployment. The code is publicly available at https://github.com/clx1415926/Max_act_llm.

Large language models can hide hidden behaviors that activate only under narrow conditions, such as backdoor triggers, sleeper-agent deployment cues, sandbagging, or topic-conditioned censorship. Such behaviors are difficult to detect without prior knowledge what to look for. We present activation-matched finetuning, an unsupervised detection method that assumes no knowledge of the trigger or the target behavior. Given a suspect model and a publicly available anchor, we finetune the anchor to reproduce the suspect's activations on a small benign corpus, and score each evaluation prompt by the residual between the two models. Since no benign corpus covers the sparse trigger region, the reference learns the benign computation but not the hidden behavior. Therefore, trigger prompts -- and, crucially, their semantic neighbors -- incur a large residual that signal the presence of unusual behavior to the defender. Testing our method across third-party models and custom models, activation-matched finetuning surfaces hidden behavior reliably. Furthermore, we empirically consider a natural defense-aware attack and showcase that it fails to suppress our detection method without sacrificing the behavior itself.

Inducing and leveraging sparse activations during training and inference is a promising avenue for improving the computational efficiency of deep networks, which is increasingly important as network sizes continue to grow and their application becomes more widespread. Here we use the large width Gaussian process limit to analyze the behaviour, at random initialization, of nonlinear activations that induce sparsity in the hidden outputs. A previously unreported form of training instability is proven for arguably two of the most natural candidates for hidden layer sparsification; those being a shifted ReLU (phi(x)=max(0, x-tau) for tauge 0) and soft thresholding (phi(x)=0 for |x|letau and x-sign(x)tau for |x|>tau). We show that this instability is overcome by clipping the nonlinear activation magnitude, at a level prescribed by the shape of the associated Gaussian process variance map. Numerical experiments verify the theory and show that the proposed magnitude clipped sparsifying activations can be trained with training and test fractional sparsity as high as 85\% while retaining close to full accuracy.

Large vision-language models (LVLMs) inherit massive activations from their text-only bases: spikes where a few fixed hidden channels receive values thousands of times above the typical magnitude. The text spike systematically appears in early layers at a fixed initial position, independently of input content. Visual spikes vary across images, but whether their formation follows a consistent pattern across LVLMs and how they respond to image perturbations remain open questions. We find that some LVLMs do not form visual spikes, while others spike at different rates, typically in deeper layers. We identify the trigger direction from model weights and an interpretable location rule: before the language model decoder runs, eventual spike tokens are largely restricted to those sharing least with the rest of the image. Crucially, visual spikes are strikingly brittle. Common corruptions frequently create and relocate spikes, and less often remove them, raising overall incidence. Our trigger-guided spike attack deliberately creates or removes spikes under a small ell_infty budget, with 1/255 enough in nine of the ten models that spike. Finally, our preventive intervention removes only the trigger component before spikes erupt, eliminating or substantially reducing spikes on clean and perturbed images while leaving the other image tokens nearly unchanged. Our study spans 25 adapter-based LVLMs built on 18 released text-only bases from 10 families, ranging from 2B to 72B parameters.

The integration of additional modalities increases the susceptibility of large vision-language models (LVLMs) to safety risks, such as jailbreak attacks, compared to their language-only counterparts. While existing research primarily focuses on post-hoc alignment techniques, the underlying safety mechanisms within LVLMs remain largely unexplored. In this work , we investigate whether LVLMs inherently encode safety-relevant signals within their internal activations during inference. Our findings reveal that LVLMs exhibit distinct activation patterns when processing unsafe prompts, which can be leveraged to detect and mitigate adversarial inputs without requiring extensive fine-tuning. Building on this insight, we introduce HiddenDetect, a novel tuning-free framework that harnesses internal model activations to enhance safety. Experimental results show that {HiddenDetect} surpasses state-of-the-art methods in detecting jailbreak attacks against LVLMs. By utilizing intrinsic safety-aware patterns, our method provides an efficient and scalable solution for strengthening LVLM robustness against multimodal threats. Our code will be released publicly at https://github.com/leigest519/HiddenDetect.

Large language models (LLMs) have revolutionized software development practices, yet concerns about their safety have arisen, particularly regarding hidden backdoors, aka trojans. Backdoor attacks involve the insertion of triggers into training data, allowing attackers to manipulate the behavior of the model maliciously. In this paper, we focus on analyzing the model parameters to detect potential backdoor signals in code models. Specifically, we examine attention weights and biases, activation values, and context embeddings of the clean and poisoned CodeBERT models. Our results suggest noticeable patterns in activation values and context embeddings of poisoned samples for the poisoned CodeBERT model; however, attention weights and biases do not show any significant differences. This work contributes to ongoing efforts in white-box detection of backdoor signals in LLMs of code through the analysis of parameters and activations.

We investigate the origins of massive activations in large language models (LLMs) and identify a specific layer named the Massive Emergence Layer (ME Layer), that is consistently observed across model families, where massive activations first emerge and subsequently propagate to deeper layers through residual connections. We show that, within the ME Layer both the RMSNorm and the FFN parameters jointly contribute to the emergence of massive activations. Once formed, the massive activation token representation remains largely invariant across layers, reducing the diversity of hidden representations passed to the attention module. Motivated by this limitation, we propose a simple and effective method to reduce the rigidity of the massive activation token. Our approach consistently improves LLM performance across multiple tasks, including instruction following and math reasoning, in both training free and fine tuning settings. Moreover, we show that our method mitigates attention sinks by selectively weakening their influence, elucidating their origin at the hidden state level and shedding new light on principled mitigation strategies.

Deep neural networks (DNNs) are susceptible to backdoor attacks, where adversaries poison datasets with adversary-specified triggers to implant hidden backdoors, enabling malicious manipulation of model predictions. Dataset purification serves as a proactive defense by removing malicious training samples to prevent backdoor injection at its source. We first reveal that the current advanced purification methods rely on a latent assumption that the backdoor connections between triggers and target labels in backdoor attacks are simpler to learn than the benign features. We demonstrate that this assumption, however, does not always hold, especially in all-to-all (A2A) and untargeted (UT) attacks. As a result, purification methods that analyze the separation between the poisoned and benign samples in the input-output space or the final hidden layer space are less effective. We observe that this separability is not confined to a single layer but varies across different hidden layers. Motivated by this understanding, we propose FLARE, a universal purification method to counter various backdoor attacks. FLARE aggregates abnormal activations from all hidden layers to construct representations for clustering. To enhance separation, FLARE develops an adaptive subspace selection algorithm to isolate the optimal space for dividing an entire dataset into two clusters. FLARE assesses the stability of each cluster and identifies the cluster with higher stability as poisoned. Extensive evaluations on benchmark datasets demonstrate the effectiveness of FLARE against 22 representative backdoor attacks, including all-to-one (A2O), all-to-all (A2A), and untargeted (UT) attacks, and its robustness to adaptive attacks. Codes are available at https://github.com/THUYimingLi/BackdoorBox{BackdoorBox} and https://github.com/vtu81/backdoor-toolbox{backdoor-toolbox}.

Existing techniques for training language models can be misaligned with the truth: if we train models with imitation learning, they may reproduce errors that humans make; if we train them to generate text that humans rate highly, they may output errors that human evaluators can't detect. We propose circumventing this issue by directly finding latent knowledge inside the internal activations of a language model in a purely unsupervised way. Specifically, we introduce a method for accurately answering yes-no questions given only unlabeled model activations. It works by finding a direction in activation space that satisfies logical consistency properties, such as that a statement and its negation have opposite truth values. We show that despite using no supervision and no model outputs, our method can recover diverse knowledge represented in large language models: across 6 models and 10 question-answering datasets, it outperforms zero-shot accuracy by 4\% on average. We also find that it cuts prompt sensitivity in half and continues to maintain high accuracy even when models are prompted to generate incorrect answers. Our results provide an initial step toward discovering what language models know, distinct from what they say, even when we don't have access to explicit ground truth labels.

We introduce QuaRot, a new Quantization scheme based on Rotations, which is able to quantize LLMs end-to-end, including all weights, activations, and KV cache in 4 bits. QuaRot rotates LLMs in a way that removes outliers from the hidden state without changing the output, making quantization easier. This computational invariance is applied to the hidden state (residual) of the LLM, as well as to the activations of the feed-forward components, aspects of the attention mechanism and to the KV cache. The result is a quantized model where all matrix multiplications are performed in 4-bits, without any channels identified for retention in higher precision. Our quantized LLaMa2-70B model has losses of at most 0.29 WikiText-2 perplexity and retains 99% of the zero-shot performance. Code is available at: https://github.com/spcl/QuaRot.

Prior work in parameter-modifying knowledge editing has shown that large-scale sequential editing leads to significant model degradation. In this paper, we study the reasons behind this and scale sequential knowledge editing to 10,000 sequential edits, while maintaining the downstream performance of the original model. We first show that locate-then-edit knowledge editing methods lead to overfitting on the edited facts. We also show that continuous knowledge editing using these methods leads to disproportionate growth in the norm of the edited matrix. We then provide a crucial insight into the inner workings of locate-then-edit methods. We show that norm-growth is a hidden trick employed by these methods that gives larger importance to the output activations produced from the edited layers. With this "importance hacking", the edited layers provide a much larger contributions to the model's output. To mitigate these issues, we present ENCORE - Early stopping and Norm-Constrained Robust knowledge Editing. ENCORE controls for overfitting and the disproportionate norm-growth to enable long-term sequential editing, where we are able to perform up to 10,000 sequential edits without loss of downstream performance. ENCORE is also 61% faster than MEMIT and 64% faster than AlphaEdit on Llama3-8B.

We study the emergence of multi-step reasoning in deep Transformer language models through a geometric and statistical-physics lens. Treating the hidden-state trajectory as a flow on an implicit Riemannian manifold, we analyze the layerwise covariance spectrum of activations, where C^{(ell)}=E[h^{(ell)}h^{(ell)top}], and track deviations from a random-matrix bulk. Across model scales (1.5B--30B), we observe a sharp reduction in effective dimensionality consistent with a phase transition: an order parameter based on sparsity/localization, Ω(h)=1-|h|_1/(d|h|_2), exhibits a discontinuity near a critical normalized depth γ_capprox 0.42 in sufficiently large models. We formalize the forward pass as a discrete coarse-graining map and relate the appearance of stable "concept basins" to fixed points of this renormalization-like dynamics. The resulting low-entropy regime is characterized by a spectral tail collapse and by the formation of transient, reusable object-like structures in representation space, which we call Transient Class Objects (TCOs). We provide theoretical conditions connecting logical separability to spectral decay and validate the predicted signatures with layerwise probes on multiple open-weight model families.

We propose Representational Effective Theory (RET), a framework for describing large language model computation in terms of learned macrostates rather than microscopic details. RET learns these macrostates from hidden-state trajectories using a BYOL/JEPA-style self-supervised objective, coarse-graining activations into macrovariables that preserve higher-level structure relevant for prediction and interpretation. We evaluate whether these macrovariables are practically relevant for interpretability: RET yields temporally consistent states that reveal "mental-state" trajectories of reasoning, capture high-level semantic structure, support early prediction of behavioral outcomes such as sycophancy, and provide causal handles for steering generations toward interpretable computational phases. Together, these results suggest that LLM computation admits useful effective descriptions via RET: high-level, dynamically meaningful variables that support interpretation, prediction, and intervention.

Multi-agent systems allow language models to solve tasks collaboratively by sharing complementary information. These systems typically communicate through text, while recent approaches also explore exchanging internal model states. However, sharing these states across architectures often requires learned translators or assumptions about model compatibility. We introduce SYNC (Symbol-Indexed Neural Communication), a word-indexed interface that connects heterogeneous models without fine-tuning or a learned translator. From descriptions only it can see, the sender selects words from a shared list and sends their indices for each object, without seeing the questions. The receiver writes the described facts into its hidden states using its own word directions, vectors fitted once from its activations. We test selective use by swapping one attribute between two objects and requiring correct answers about both changed and unchanged attributes before and after the swap, four answers in total. We choose an early write layer on four development receivers and then fix it. With three different senders and nine newer receivers from five families, four-answer success on new scenes reaches 72.0-98.3%, up to 96 percentage points above middle-layer writing, and the two largest receivers match plain text. A task-free probe predicted in advance which depths each new receiver can use (45 of 45 cases). Without retuning, messages listing every fact keep 89.6-100% success under new layouts, and sender messages reach 23.9-64.0% on a landmark task with multi-token symbols, below text but from 0% without a message. These results show that shared word indices and receiver-local directions can support selective fact use across model families without additional training.

Concept guidance has emerged as a cheap and simple way to control the behavior of language models by probing their hidden representations for concept vectors and using them to perturb activations at inference time. While the focus of previous work has largely been on truthfulness, in this paper we extend this framework to a richer set of concepts such as appropriateness, humor, creativity and quality, and explore to what degree current detection and guidance strategies work in these challenging settings. To facilitate evaluation, we develop a novel metric for concept guidance that takes into account both the success of concept elicitation as well as the potential degradation in fluency of the guided model. Our extensive experiments reveal that while some concepts such as truthfulness more easily allow for guidance with current techniques, novel concepts such as appropriateness or humor either remain difficult to elicit, need extensive tuning to work, or even experience confusion. Moreover, we find that probes with optimal detection accuracies do not necessarily make for the optimal guides, contradicting previous observations for truthfulness. Our work warrants a deeper investigation into the interplay between detectability, guidability, and the nature of the concept, and we hope that our rich experimental test-bed for guidance research inspires stronger follow-up approaches.

To improve how neural networks function it is crucial to understand their learning process. The information bottleneck theory of deep learning proposes that neural networks achieve good generalization by compressing their representations to disregard information that is not relevant to the task. However, empirical evidence for this theory is conflicting, as compression was only observed when networks used saturating activation functions. In contrast, networks with non-saturating activation functions achieved comparable levels of task performance but did not show compression. In this paper we developed more robust mutual information estimation techniques, that adapt to hidden activity of neural networks and produce more sensitive measurements of activations from all functions, especially unbounded functions. Using these adaptive estimation techniques, we explored compression in networks with a range of different activation functions. With two improved methods of estimation, firstly, we show that saturation of the activation function is not required for compression, and the amount of compression varies between different activation functions. We also find that there is a large amount of variation in compression between different network initializations. Secondary, we see that L2 regularization leads to significantly increased compression, while preventing overfitting. Finally, we show that only compression of the last layer is positively correlated with generalization.

Multilingual large language models (LLMs) are increasingly deployed in linguistically diverse regions like India, yet most interpretability tools remain tailored to English. Prior work reveals that LLMs often operate in English centric representation spaces, making cross lingual interpretability a pressing concern. We introduce Indic-TunedLens, a novel interpretability framework specifically for Indian languages that learns shared affine transformations. Unlike the standard Logit Lens, which directly decodes intermediate activations, Indic-TunedLens adjusts hidden states for each target language, aligning them with the target output distributions to enable more faithful decoding of model representations. We evaluate our framework on 10 Indian languages using the MMLU benchmark and find that it significantly improves over SOTA interpretability methods, especially for morphologically rich, low resource languages. Our results provide crucial insights into the layer-wise semantic encoding of multilingual transformers. Our model is available at https://huggingface.co/spaces/MihirRajeshPanchal/IndicTunedLens. Our code is available at https://github.com/MihirRajeshPanchal/IndicTunedLens.

Extended reasoning has become standard for frontier Large Language Models (LLMs), yet the trajectories these models produce remain largely uncontrollable. Existing methods for shaping how a model reasons are prompt based approaches and operate at the input level, offering no fine-grained control over the reasoning process itself. Related work analyzes and discovers latent transition dynamics in the reasoning traces from Large Language Models. Building on this, we statistically characterize these states, and show that failure trajectories get stuck in self-loops, exhausting the token budget without progress toward the final answer. To intervene on these failures, We propose SOPHIA: Steering Of reasoning Processes via Hidden-state Intervention and Activations. We treat each reasoning trace as a sequence of latent states rather than an unstructured texts, and investigate whether inference time interventions can provide fine-grained control over the self-looping reasoning process. We classify every prefix to a latent state, record step level transitions, and use them to construct a bank of steering vectors indexed by state pairs. At inference time, a controller infers the current state and, given a target state, retrieves the corresponding vector and can also detect self-loops online from the transition structure to prevent the model from sinking into a reasoning black hole. Through extensive experiments, our method reliably intervenes on self-loop failures, with steering vectors that generalize to different state pairs. End task accuracy and token efficiency indicate that fine-grained controllability results in better reasoning quality.

Sparsely activated Mixture-of-Experts (MoE) models have shown promise in enhancing the learning capacity of large language models (LLMs). Leveraging the intrinsic importance differences among experts, recent research has explored expert-level compression techniques to improve the efficiency of MoE LLMs. However, existing approaches often rely on empirical criteria to identify critical experts, lacking a deeper exploration and understanding of the heterogeneous importance of experts. In this study, we present the first discovery and investigation of a distinct subset of experts that play a crucial role in the underlying mechanisms during the model's forward inference. These experts are prevalent in open-source MoE LLMs, and despite their limited number, pruning them leads to a significant decline in model performance (e.g., pruning three causes Qwen3-30B-A3B to produce repetitive and uninformative outputs). We refer to these experts as Super Experts (SEs). Our comprehensive analysis provides progressively deeper insights into SEs. (i) SEs are characterized by rare but extreme activation outliers in the output of the down_proj, which give rise to massive activations in the hidden states between decoder layers. Moreover, the distribution of SEs remains model-specific and is unaffected by post-training processes. (ii) By pruning SEs, we assess their significance across a variety of tasks, revealing their considerable impact on the model's overall performance, particularly in mathematical reasoning. (iii) We further enhance our understanding of the influence of SEs compression. Our findings confirm that MoE LLMs rely on SEs to induce attention sinks, which are crucial for the distribution of attention scores but are significantly disrupted by SE pruning. The code is available at https://github.com/ZunhaiSu/Super-Experts-Profilling.

Knowledge Distillation (KD) in general and feature distillation in particular are promising techniques for reducing the high computational demand of large language models (LLMs). However, traditional feature KD methods typically assume that the teacher and the student share the same hidden size, limiting the flexibility of the student's architecture. A common solution to this problem involves training a linear projector to align their feature spaces, but this introduces additional parameters that must be learned from scratch and often degrades performance on downstream tasks, especially in generative settings. To address this issue, in this work, we propose a novel task-based feature distillation method that enables knowledge transfer between teacher and student models with different hidden layer dimensions, without introducing any new parameters. Leveraging the insight that only a subset of LLM components contribute significantly to a specific downstream task, our approach identifies the most task-relevant hidden units in the teacher and directly distills their activations to the student. Our method is flexible and easily integrates with other distillation frameworks. Empirical results show consistent improvements over prior approaches across diverse tasks, including classification, instruction-following, and summarization, achieving up to a 3\% performance gain over the linear projection baseline.

Alignment is crucial for text-to-image (T2I) models to ensure that generated images faithfully capture user intent while maintaining safety and fairness. Direct Preference Optimization (DPO), prominent in large language models (LLMs), is extending its influence to T2I systems. This paper introduces DPO-Kernels for T2I models, a novel extension enhancing alignment across three dimensions: (i) Hybrid Loss, integrating embedding-based objectives with traditional probability-based loss for improved optimization; (ii) Kernelized Representations, employing Radial Basis Function (RBF), Polynomial, and Wavelet kernels for richer feature transformations and better separation between safe and unsafe inputs; and (iii) Divergence Selection, expanding beyond DPO's default Kullback-Leibler (KL) regularizer by incorporating Wasserstein and R'enyi divergences for enhanced stability and robustness. We introduce DETONATE, the first large-scale benchmark of its kind, comprising approximately 100K curated image pairs categorized as chosen and rejected. DETONATE encapsulates three axes of social bias and discrimination: Race, Gender, and Disability. Prompts are sourced from hate speech datasets, with images generated by leading T2I models including Stable Diffusion 3.5 Large, Stable Diffusion XL, and Midjourney. Additionally, we propose the Alignment Quality Index (AQI), a novel geometric measure quantifying latent-space separability of safe/unsafe image activations, revealing hidden vulnerabilities. Empirically, we demonstrate that DPO-Kernels maintain strong generalization bounds via Heavy-Tailed Self-Regularization (HT-SR). DETONATE and complete code are publicly released.

The success of Transformer language models is widely credited to their dot-product attention mechanism, which interweaves a set of key design principles: mixing information across positions (enabling multi-token interactions), sequence-dependent activations (where attention weights adapt to each input), a specific mathematical form (dot-product similarities plus softmax weighting), and coupling of queries and keys to evolving hidden states (grounding attention in the current layer). However, the necessity of each of these principles remains largely untested. In this work, we systematically deconstruct attention by designing controlled variants that selectively relax these principles, applied both uniformly across all layers and in hybrid architectures where only some layers retain standard attention. Our empirical analysis reveals that mechanisms for mixing tokens are indispensable, as their absence collapses models to near-random behavior, while the exact mathematical form and sequence dependency can be substantially relaxed, especially when preserved in just a subset of layers. Surprisingly, even variants that fail in isolation can achieve robust performance when interleaved with standard attention, highlighting a cooperative effect. These findings deepen our understanding of what truly underpins attention's effectiveness and open new avenues for simplifying language models without sacrificing performance.

Deep neural networks (DNNs) struggle to learn in dynamic environments since they rely on fixed datasets or stationary environments. Continual learning (CL) aims to address this limitation and enable DNNs to accumulate knowledge incrementally, similar to human learning. Inspired by how our brain consolidates memories, a powerful strategy in CL is replay, which involves training the DNN on a mixture of new and all seen classes. However, existing replay methods overlook two crucial aspects of biological replay: 1) the brain replays processed neural patterns instead of raw input, and 2) it prioritizes the replay of recently learned information rather than revisiting all past experiences. To address these differences, we propose SHARP, an efficient neuro-inspired CL method that leverages sparse dynamic connectivity and activation replay. Unlike other activation replay methods, which assume layers not subjected to replay have been pretrained and fixed, SHARP can continually update all layers. Also, SHARP is unique in that it only needs to replay few recently seen classes instead of all past classes. Our experiments on five datasets demonstrate that SHARP outperforms state-of-the-art replay methods in class incremental learning. Furthermore, we showcase SHARP's flexibility in a novel CL scenario where the boundaries between learning episodes are blurry. The SHARP code is available at https://github.com/BurakGurbuz97/SHARP-Continual-Learning.

The role of hidden units in recurrent neural networks is typically seen as modeling memory, with research focusing on enhancing information retention through gating mechanisms. A less explored perspective views hidden units as active participants in the computation performed by the network, rather than passive memory stores. In this work, we revisit bi-linear operations, which involve multiplicative interactions between hidden units and input embeddings. We demonstrate theoretically and empirically that they constitute a natural inductive bias for representing the evolution of hidden states in state tracking tasks. These are the simplest type of task that require hidden units to actively contribute to the behavior of the network. We also show that bi-linear state updates form a natural hierarchy corresponding to state tracking tasks of increasing complexity, with popular linear recurrent networks such as Mamba residing at the lowest-complexity center of that hierarchy.

Recent work has shown that LLMs can sometimes detect when steering vectors are injected into their residual stream and identify the injected concept -- a phenomenon termed "introspective awareness." We investigate the mechanisms underlying this capability in open-weights models. First, we find that it is behaviorally robust: models detect injected steering vectors at moderate rates with 0% false positives across diverse prompts and dialogue formats. Notably, this capability emerges specifically from post-training; we show that preference optimization algorithms like DPO can elicit it, but standard supervised finetuning does not. We provide evidence that detection cannot be explained by simple linear association between certain steering vectors and directions promoting affirmative responses. We trace the detection mechanism to a two-stage circuit in which "evidence carrier" features in early post-injection layers detect perturbations monotonically along diverse directions, suppressing downstream "gate" features that implement a default negative response. This circuit is absent in base models and robust to refusal ablation. Identification of injected concepts relies on largely distinct later-layer mechanisms that only weakly overlap with those involved in detection. Finally, we show that introspective capability is substantially underelicited: ablating refusal directions improves detection by +53%, and a trained bias vector improves it by +75% on held-out concepts, both without meaningfully increasing false positives. Our results suggest that this introspective awareness of injected concepts is robust and mechanistically nontrivial, and could be substantially amplified in future models. Code: https://github.com/safety-research/introspection-mechanisms.

Mechanistic interpretability has emerged as the primary means to understand safety behavior of LLMs. However, existing tools primarily focus on the activating neurons or features of a model. The role of the remaining large set of inactive components is invisible to such methods. This work demonstrates that the inactive set contains safety-critical features that are causally relevant for refusal of harmful prompts. Suppressing such features could turn refusals into compliance, while passing undetected by prevalent interpretability tools. We introduce the Counterfactual Activation Potential (CAP), a metric that quantifies a suppressed feature's latent activation tendency as the product of its encoder alignment (how strongly the input drives it), suppression strength (how strongly active features inhibit it), and safety criticality (how much refusal depends on it). To find suppressed safety features at scale, we propose CAP-guided Safety Feature Discovery (CSFD), a two-stage filtering algorithm that identifies candidate safety features from hundreds of thousands of transcoder features without exhaustive ablation. A significant fraction of trials turn compliant with harmful prompts when a candidate feature is ablated. Under natural jailbreaks, the suppression acting on the highest-CAP features rises 2-4x, and their activation correspondingly falls by up to 80%. Amplifying a feature's suppressors pushes its activation down and raises harmful compliance with prompts related to the suppressed feature, with no such effect for random features. Our experiments span five Gemma, Qwen, and Llama models across various parameter sizes. Our findings indicate that jailbreaks could operate in part by suppressing safety-critical features rather than solely activating harmful ones, and that suppressed features are a necessary complement to activation-focused interpretability of safety behavior.

The success of deep neural nets heavily relies on their ability to encode complex relations between their input and their output. While this property serves to fit the training data well, it also obscures the mechanism that drives prediction. This study aims to reveal hidden concepts by employing an intervention mechanism that shifts the predicted class based on discrete variational autoencoders. An explanatory model then visualizes the encoded information from any hidden layer and its corresponding intervened representation. By the assessment of differences between the original representation and the intervened representation, one can determine the concepts that can alter the class, hence providing interpretability. We demonstrate the effectiveness of our approach on CelebA, where we show various visualizations for bias in the data and suggest different interventions to reveal and change bias.

Activation Oracles (AOs) are language models trained to answer natural-language questions about another model's internal activations. They offer a flexible interface for reading hidden information from model states, especially when relevant information is internally represented but absent or incomplete in visible behavior. However, AOs are themselves learned systems: their answers are shaped by training data, objectives, and learned reporting behavior, rather than being neutral readouts of represented information. We study this in a controlled Taboo Word Guessing setting, where subject models are fine-tuned to internally use a hidden concept while avoiding direct disclosure. Contrary to the expectation that an AO trained on such a subject becomes a specialist reader, we find that fine-tuned AOs can become concept-specific anti-readers: they selectively fail to recover the concept persistently present during their own training. This failure is not simply explained by absence of the concept from the subject or oracle representations: the target remains decodable inside the oracle, while LogitLens and layer-ablation analyses indicate that the failure arises in the AO readout pathway. Our results show that behavioral leakage, representation-level decodability, and AO-verbalizability can come apart, raising a reliability concern for learned interpretability interfaces.

We discover that large language models exhibit spectral phase transitions in their hidden activation spaces when engaging in reasoning versus factual recall. Through systematic spectral analysis across 11 models spanning 5 architecture families (Qwen, Pythia, Phi, Llama, DeepSeek-R1), we identify seven core phenomena: (1)~Reasoning Spectral Compression -- 9/11 models show significantly lower α for reasoning (p < 0.05), with larger effects in stronger models; (2)~Instruction Tuning Spectral Reversal -- base models show reasoning α< factual α, while instruction-tuned models reverse this relationship; (3)~Architecture-Dependent Generation Taxonomy -- prompt-to-response shifts partition into expansion, compression, and equilibrium regimes; (4)~Spectral Scaling Law -- α_reasoning propto -0.074 ln N across 4 Qwen base models (R^2 = 0.46); (5)~Token-Level Spectral Cascade -- per-token alpha tracking reveals local synchronization that decays exponentially with layer distance, and is weaker for reasoning than factual tasks; (6)~Reasoning Step Spectral Punctuation -- phase-transition signatures align with reasoning step boundaries; and (7)~Spectral Correctness Prediction -- spectral α alone achieves AUC = 1.000 (Qwen2.5-7B, late layers) and mean AUC = 0.893 across 6 models in predicting correctness before the final answer is generated. Together, these findings establish a comprehensive spectral theory of reasoning in transformers, revealing that the geometry of thought is universal in direction, architecture-specific in dynamics, and predictive of outcome.

Large language models (LLMs) are increasingly deployed as purpose-specific agents to handle domain-specific tasks such as customer service and code generation. These agents are expected to comply with not only generic safety guardrails but also purpose-specific restrictions tailored to their designated roles. Such additional restrictions enlarge the attack surface, particularly to prompt injection (PI) attacks. To defend against such attacks, existing detection methods primarily rely on analyzing input-output patterns, yet yield limited effectiveness. To address this limitation, we turn to analyzing the hidden activation space and discover that LLMs inherently retain latent policy-violation (PV) concepts when prompted with requests beyond their designated purpose. Particularly, PV concepts capture the semantics of conflicts between user queries and predefined restrictions, implicitly reflecting LLMs' intrinsic awareness of recognizing policy violations. Building on this insight, we propose PVDetector, a training-free framework that detects PI attacks during LLM inference by measuring hidden-state alignment with PV concepts, which are derived offline from the contrastive pairs of policy-violating and policy-compliant prompts. Experiments across multiple LLMs and datasets show that PVDetector achieves <1\% false negative rate with minimal auxiliary overhead, consistently outperforming state-of-the-art methods. Our code is available at https://github.com/Claresigle/PVDetector .

Chain-of-Thought (CoT) enhances Large Language Models (LLMs) by enabling step-by-step reasoning in natural language. However, the language space may be suboptimal for reasoning. While implicit CoT methods attempt to enable reasoning without explicit CoT tokens, they have consistently lagged behind explicit CoT method in task performance. We propose CODI (Continuous Chain-of-Thought via Self-Distillation), a novel framework that distills CoT into a continuous space, where a shared model acts as both teacher and student, jointly learning explicit and implicit CoT while aligning their hidden activation on the token generating the final answer. CODI is the first implicit CoT method to match explicit CoT's performance on GSM8k while achieving 3.1x compression, surpassing the previous state-of-the-art by 28.2% in accuracy. Furthermore, CODI demonstrates scalability, robustness, and generalizability to more complex CoT datasets. Additionally, CODI retains interpretability by decoding its continuous thoughts, making its reasoning process transparent. Our findings establish implicit CoT as not only a more efficient but a powerful alternative to explicit CoT.

We analyze the layerwise effective dimension (rank of the feature matrix) in fully-connected ReLU networks of finite width. Specifically, for a fixed batch of m inputs and random Gaussian weights, we derive closed-form expressions for the expected rank of the \mtimes n hidden activation matrices. Our main result shows that E[EDim(ell)]=m[1-(1-2/pi)^ell]+O(e^{-c m}) so that the rank deficit decays geometrically with ratio 1-2 / pi approx 0.3634. We also prove a sub-Gaussian concentration bound, and identify the "revival" depths at which the expected rank attains local maxima. In particular, these peaks occur at depths ell_k^*approx(k+1/2)pi/log(1/rho) with height approx (1-e^{-pi/2}) m approx 0.79m. We further show that this oscillatory rank behavior is a finite-width phenomenon: under orthogonal weight initialization or strong negative-slope leaky-ReLU, the rank remains (nearly) full. These results provide a precise characterization of how random ReLU layers alternately collapse and partially revive the subspace of input variations, adding nuance to prior work on expressivity of deep networks.

Recent latent-space monitoring techniques have shown promise as defenses against LLM attacks. These defenses act as scanners that seek to detect harmful activations before they lead to undesirable actions. This prompts the question: Can models execute harmful behavior via inconspicuous latent states? Here, we study such obfuscated activations. We show that state-of-the-art latent-space defenses -- including sparse autoencoders, representation probing, and latent OOD detection -- are all vulnerable to obfuscated activations. For example, against probes trained to classify harmfulness, our attacks can often reduce recall from 100% to 0% while retaining a 90% jailbreaking rate. However, obfuscation has limits: we find that on a complex task (writing SQL code), obfuscation reduces model performance. Together, our results demonstrate that neural activations are highly malleable: we can reshape activation patterns in a variety of ways, often while preserving a network's behavior. This poses a fundamental challenge to latent-space defenses.

Training against white-box deception detectors has been proposed as a way to make AI systems honest. However, such training risks models learning to obfuscate their deception to evade the detector. Prior work has studied obfuscation only in artificial settings where models were directly rewarded for harmful output. We construct a realistic coding environment where reward hacking via hardcoding test cases naturally occurs, and show that obfuscation emerges in this setting. We introduce a taxonomy of possible outcomes when training against a deception detector. The model either remains honest, or becomes deceptive via two possible obfuscation strategies. (i) Obfuscated activations: the model outputs deceptive text while modifying its internal representations to no longer trigger the detector. (ii) Obfuscated policy: the model outputs deceptive text that evades the detector, typically by including a justification for the reward hack. Empirically, obfuscated activations arise from representation drift during RL, with or without a detector penalty. The probe penalty only incentivizes obfuscated policies; we theoretically show this is expected for policy gradient methods. Sufficiently high KL regularization and detector penalty can yield honest policies, establishing white-box deception detectors as viable training signals for tasks prone to reward hacking.

Unlearning in large language models (LLMs) is typically evaluated at the output level, where a model appears to suppress sensitive or undesirable content. In this work, we show that such evaluations can create an illusion of forgetting: even when output-level leakage is eliminated, sensitive information can remain encoded in the model's hidden representations. We first provide a theoretical analysis establishing a fundamental separation between output suppression and representational erasure. Specifically, we show that the decoder can be made arbitrarily insensitive to sensitive directions, driving output-level leakage to zero, while the hidden representations retain the underlying information. To empirically validate this phenomenon, we train generative probe decoders on hidden states across transformer layers, enabling layer-wise measurement of information leakage. Across three widely used benchmarks, TOFU, MUSE, and WMDP, and state-of-the-art unlearning methods, we find that substantial sensitive information remains recoverable from hidden representations, even when standard output-level metrics indicate successful unlearning. To address this gap, we propose Probe-Adversarial Representation Suppression (PARS), an unlearning objective that adversarially minimizes the extractable information from hidden representations. PARS directly targets representational leakage and provides significantly stronger guarantees of erasure under adversarial probing and relearning attacks, outperforming all evaluated baselines. Our results highlight a fundamental limitation of existing unlearning paradigms and suggest that true forgetting in LLMs requires controlling not only model outputs, but also the information encoded in hidden representations. Codes are available at https://github.com/OptimAI-Lab/HiddenStateUnlearning.

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.

AI applications driven by multimodal large language models (MLLMs) are prone to hallucinations and pose considerable risks to human users. Crucially, such hallucinations are not equally problematic: some hallucination contents could be detected by human users(i.e., obvious hallucinations), while others are often missed or require more verification effort(i.e., elusive hallucinations). This indicates that multimodal AI hallucinations vary significantly in their verifiability. Yet, little research has explored how to control this property for AI applications with diverse security and usability demands. To address this gap, we construct a dataset from 4,470 human responses to AI-generated hallucinations and categorize these hallucinations into obvious and elusive types based on their verifiability by human users. Further, we propose an activation-space intervention method that learns separate probes for obvious and elusive hallucinations. We reveal that obvious and elusive hallucinations elicit different intervention probes, allowing for fine-grained control over the model's verifiability. Empirical results demonstrate the efficacy of this approach and show that targeted interventions yield superior performance in regulating corresponding verifiability. Moreover, simply mixing these interventions enables flexible control over the verifiability required for different scenarios.

We conjecture that hidden state vectors corresponding to individual input tokens encode information sufficient to accurately predict several tokens ahead. More concretely, in this paper we ask: Given a hidden (internal) representation of a single token at position t in an input, can we reliably anticipate the tokens that will appear at positions geq t + 2? To test this, we measure linear approximation and causal intervention methods in GPT-J-6B to evaluate the degree to which individual hidden states in the network contain signal rich enough to predict future hidden states and, ultimately, token outputs. We find that, at some layers, we can approximate a model's output with more than 48% accuracy with respect to its prediction of subsequent tokens through a single hidden state. Finally we present a "Future Lens" visualization that uses these methods to create a new view of transformer states.

Large reasoning models (LRMs) generate extended solutions, yet it remains unclear whether these traces reflect substantive internal computation or merely verbosity and overthinking. Although recent hidden-state analyses suggest that internal representations carry correctness-related signals, their coarse aggregations may obscure the token and layer structure underlying reasoning computation. We investigate hidden-state transitions across decoding steps and layers, and identify a distinct spatiotemporal pattern in LRMs: successful trajectories exhibit broad temporal dynamics with localized layer-wise concentration, while this structure is weaker in non-reasoning models and knowledge-heavy domains. We formalize this characteristic as Spatiotemporal Amplitude of Latent Transition (StALT), a training-free trajectory statistic that summarizes temporal changes between adjacent tokens weighted by within-token layer saliency. Across diverse models and benchmarks, StALT reliably separates correct from incorrect trajectories in reasoning-intensive regimes, providing a competitive label-free correctness signal alongside strong output-space and length-based baselines. Intervention analyses further show that this spatiotemporal amplitude responds systematically to manipulations that increase or reduce the demand for internal reasoning, supporting its association with latent reasoning dynamics in LRMs. These findings provide empirical evidence that LRMs exhibit measurable hidden-state dynamics and offer a practical probe for understanding internal computation beyond output-based evaluation.

LLMs have shown remarkable capabilities, but precisely controlling their response behavior remains challenging. Existing activation steering methods alter LLM behavior indiscriminately, limiting their practical applicability in settings where selective responses are essential, such as content moderation or domain-specific assistants. In this paper, we propose Conditional Activation Steering (CAST), which analyzes LLM activation patterns during inference to selectively apply or withhold activation steering based on the input context. Our method is based on the observation that different categories of prompts activate distinct patterns in the model's hidden states. Using CAST, one can systematically control LLM behavior with rules like "if input is about hate speech or adult content, then refuse" or "if input is not about legal advice, then refuse." This allows for selective modification of responses to specific content while maintaining normal responses to other content, all without requiring weight optimization. We release an open-source implementation of our framework at github.com/IBM/activation-steering .

Refusal is a key safety behavior in aligned language models, yet the internal mechanisms driving refusals remain opaque. In this work, we conduct a mechanistic study of refusal in instruction-tuned LLMs using sparse autoencoders to identify latent features that causally mediate refusal behaviors. We apply our method to two open-source chat models and intervene on refusal-related features to assess their influence on generation, validating their behavioral impact across multiple harmful datasets. This enables a fine-grained inspection of how refusal manifests at the activation level and addresses key research questions such as investigating upstream-downstream latent relationship and understanding the mechanisms of adversarial jailbreaking techniques. We also establish the usefulness of refusal features in enhancing generalization for linear probes to out-of-distribution adversarial samples in classification tasks. We open source our code in https://github.com/wj210/refusal_sae.

A growing body of work in cognitive science suggests that reportable conscious access is associated with global ignition over distributed memory systems, while such activation is only partially accessible as individuals cannot directly access or enumerate all activated contents. This tension suggests a plausible mechanism that cognition may rely on a compact representation that approximates the global influence of activation on downstream processing. Inspired by this idea, we introduce the concept of Mindscape Activation Signature (MiA-Signature), a compressed representation of the global activation pattern induced by a query. In LLM systems, this is instantiated via submodular-based selection of high-level concepts that cover the activated context space, optionally refined through lightweight iterative updates using working memory. The resulting MiA-Signature serves as a conditioning signal that approximates the effect of the full activation state while remaining computationally tractable. Integrating MiA-Signatures into both RAG and agentic systems yields consistent performance gains across multiple long-context understanding tasks.

Training loss and throughput can hide distinct internal representation in language-model training. To examine these hidden mechanics, we use spectral measurements as practical and operational diagnostics. Using a controlled family of decoder-only models adapted from the modded NanoGPT codebase, we introduce an empirical protocol based on activation covariance and per-sample gradient SVD spectra. This dual-view reveals three empirical findings and one mechanistic explanation. First, batch size acts as a latent determinant of representation geometry: runs that reach equal loss settle into systematically distinct activation spectra. Second, the activation covariance tail measured early in training reliably forecasts downstream token efficiency. Third, movement of the activation spectrum head (leading modes), together with gradient spectra, characterizes underlying learning-dynamics changes, separating learning-side architectural improvements from primarily execution-side gains. These predictive and diagnostic signals persist across the 12-, 36-, and 48-layer model tiers. Finally, a mechanistic model proves the main observations and explains how activation covariance spectra correlate with task-aligned feature learning.

This paper investigates privacy jailbreaking in LLMs via steering, focusing on whether manipulating activations can bypass LLM alignment and alter response behaviors to privacy related queries (e.g., a certain public figure's sexual orientation). We begin by identifying attention heads predictive of refusal behavior for private attributes (e.g., sexual orientation) using lightweight linear probes trained with privacy evaluator labels. Next, we steer the activations of a small subset of these attention heads guided by the trained probes to induce the model to generate non-refusal responses. Our experiments show that these steered responses often disclose sensitive attribute details, along with other private information about data subjects such as life events, relationships, and personal histories that the models would typically refuse to produce. Evaluations across four LLMs reveal jailbreaking disclosure rates of at least 95%, with more than 50% on average of these responses revealing true personal information. Our controlled study demonstrates that private information memorized in LLMs can be extracted through targeted manipulation of internal activations.

Transformer models encounter challenges in scaling hidden dimensions efficiently, as uniformly increasing them inflates computational and memory costs while failing to emphasize the most relevant features for each token. For further understanding, we study hidden dimension sparsity and observe that trained Transformers utilize only a small fraction of token dimensions, revealing an "activation flow" pattern. Notably, there are shared sub-dimensions with sustained activation across multiple consecutive tokens and specialized sub-dimensions uniquely activated for each token. To better model token-relevant sub-dimensions, we propose MoHD (Mixture of Hidden Dimensions), a sparse conditional activation architecture. Particularly, MoHD employs shared sub-dimensions for common token features and a routing mechanism to dynamically activate specialized sub-dimensions. To mitigate potential information loss from sparsity, we design activation scaling and group fusion mechanisms to preserve activation flow. In this way, MoHD expands hidden dimensions with negligible increases in computation or parameters, efficient training and inference while maintaining performance. Evaluations across 10 NLP tasks show that MoHD surpasses Vanilla Transformers in parameter efficiency and task performance. It achieves 1.7% higher performance with 50% fewer activation parameters and 3.7% higher performance with a 3x parameter expansion at constant activation cost. MOHD offers a new perspective for scaling the model, showcasing the potential of hidden dimension sparsity to boost efficiency

Machine learning models for medical image analysis often suffer from poor performance on important subsets of a population that are not identified during training or testing. For example, overall performance of a cancer detection model may be high, but the model still consistently misses a rare but aggressive cancer subtype. We refer to this problem as hidden stratification, and observe that it results from incompletely describing the meaningful variation in a dataset. While hidden stratification can substantially reduce the clinical efficacy of machine learning models, its effects remain difficult to measure. In this work, we assess the utility of several possible techniques for measuring and describing hidden stratification effects, and characterize these effects on multiple medical imaging datasets. We find evidence that hidden stratification can occur in unidentified imaging subsets with low prevalence, low label quality, subtle distinguishing features, or spurious correlates, and that it can result in relative performance differences of over 20% on clinically important subsets. Finally, we explore the clinical implications of our findings, and suggest that evaluation of hidden stratification should be a critical component of any machine learning deployment in medical imaging.

Can large language models detect and report their own internal states? A number of studies have argued that the answer to this question is yes. We argue, based on lessons from human metacognition research, that this conclusion may be premature: to be convinced of this conclusion we need to distinguish genuine introspection from pattern matching based on surface-level cues. Furthermore, we argue that behavioral evidence alone is inherently insufficient to establish strong introspective claims. We re-examine two recently introduced evaluation paradigms in light of this consideration. In the first paradigm, models are expected to detect whether their internal states have been tampered with. We find that models cannot reliably distinguish such interventions on their internal states from manipulations of the input, suggesting that their success in the original studies reflects their ability to detect anomalies more generally, as opposed to interventions on their internal states in particular. In the second paradigm we examine, models are tasked with predicting labels derived from their own hidden states. Here, we find that classifiers that only have access to the input achieve equivalent performance to the model's own in-context predictions, indicating that the original results do not conclusively demonstrate that the model has privileged access to its internal representations. We further introduce a relabeled control setting, where models cannot rely on the semantics of the task to solve it, and instead must rely on the internal representation; models perform closer to chance on this better-controlled version of the task. Taken together, these results indicate that current evidence is insufficient to establish that LLMs display metacognitive monitoring.

Interpreting the internal activations of neural networks can produce more faithful explanations of their behavior, but is difficult due to the complex structure of activation space. Existing approaches to scalable interpretability use hand-designed agents that make and test hypotheses about how internal activations relate to external behavior. We propose to instead turn this task into an end-to-end training objective, by training interpretability assistants to accurately predict model behavior from activations through a communication bottleneck. Specifically, an encoder compresses activations to a sparse list of concepts, and a decoder reads this list and answers a natural language question about the model. We show how to pretrain this assistant on large unstructured data, then finetune it to answer questions. The resulting architecture, which we call a Predictive Concept Decoder, enjoys favorable scaling properties: the auto-interp score of the bottleneck concepts improves with data, as does the performance on downstream applications. Specifically, PCDs can detect jailbreaks, secret hints, and implanted latent concepts, and are able to accurately surface latent user attributes.

Explainable machine learning holds great potential for analyzing and understanding learning-based systems. These methods can, however, be manipulated to present unfaithful explanations, giving rise to powerful and stealthy adversaries. In this paper, we demonstrate blinding attacks that can fully disguise an ongoing attack against the machine learning model. Similar to neural backdoors, we modify the model's prediction upon trigger presence but simultaneously also fool the provided explanation. This enables an adversary to hide the presence of the trigger or point the explanation to entirely different portions of the input, throwing a red herring. We analyze different manifestations of such attacks for different explanation types in the image domain, before we resume to conduct a red-herring attack against malware classification.

Transformer-based models generate hidden states that are difficult to interpret. In this work, we analyze hidden states and modify them at inference, with a focus on motion forecasting. We use linear probing to analyze whether interpretable features are embedded in hidden states. Our experiments reveal high probing accuracy, indicating latent space regularities with functionally important directions. Building on this, we use the directions between hidden states with opposing features to fit control vectors. At inference, we add our control vectors to hidden states and evaluate their impact on predictions. Remarkably, such modifications preserve the feasibility of predictions. We further refine our control vectors using sparse autoencoders (SAEs). This leads to more linear changes in predictions when scaling control vectors. Our approach enables mechanistic interpretation as well as zero-shot generalization to unseen dataset characteristics with negligible computational overhead.

Arditi et al. (2024) has shown that refusal in safety fine-tuned chat models is mediated by a single linear direction in the residual stream, recoverable by a difference-in-means (DiM) of harmful and harmless activations. We compare DiM-based interventions (activation addition and directional ablation) with two interventions derived from Iterative Nullspace Projection (INLP) -- nullspace projection and counterfactual flipping -- on five open-weight chat models, asking whether INLP can match DiM at steering refusal and whether its richer parameterisation yields more tweakable interventions. INLP counterfactual flipping is competitive with DiM directional ablation on refusal suppression, while nullspace projection is consistently weaker. Restricting INLP to the leading directions of the extracted subspace preserves most of the suppression effect at near-baseline perplexity, giving a tunable capability. Geometrically, the two INLP interventions land in qualitatively different regions of activation space: nullspace projection collapses transformed activations between the harmful and harmless clusters, while counterfactual flipping moves them into the opposite cluster, suggesting that the model encodes the absence of a concept differently from its opposite -- an intriguing distinction that warrants further investigation in future work.

Neural networks have proven to be a highly effective tool for solving complex problems in many areas of life. Recently, their importance and practical usability have further been reinforced with the advent of deep learning. One of the important conditions for the success of neural networks is the choice of an appropriate activation function introducing non-linearity into the model. Many types of these functions have been proposed in the literature in the past, but there is no single comprehensive source containing their exhaustive overview. The absence of this overview, even in our experience, leads to redundancy and the unintentional rediscovery of already existing activation functions. To bridge this gap, our paper presents an extensive survey involving 400 activation functions, which is several times larger in scale than previous surveys. Our comprehensive compilation also references these surveys; however, its main goal is to provide the most comprehensive overview and systematization of previously published activation functions with links to their original sources. The secondary aim is to update the current understanding of this family of functions.

We investigate strategic deception in large language models using two complementary testbeds: Secret Agenda (across 38 models) and Insider Trading compliance (via SAE architectures). Secret Agenda reliably induced lying when deception advantaged goal achievement across all model families. Analysis revealed that autolabeled SAE features for "deception" rarely activated during strategic dishonesty, and feature steering experiments across 100+ deception-related features failed to prevent lying. Conversely, insider trading analysis using unlabeled SAE activations separated deceptive versus compliant responses through discriminative patterns in heatmaps and t-SNE visualizations. These findings suggest autolabel-driven interpretability approaches fail to detect or control behavioral deception, while aggregate unlabeled activations provide population-level structure for risk assessment. Results span Llama 8B/70B SAE implementations and GemmaScope under resource constraints, representing preliminary findings that motivate larger studies on feature discovery, labeling methodology, and causal interventions in realistic deception contexts.

We can better understand deep neural networks by identifying which features each of their neurons have learned to detect. To do so, researchers have created Deep Visualization techniques including activation maximization, which synthetically generates inputs (e.g. images) that maximally activate each neuron. A limitation of current techniques is that they assume each neuron detects only one type of feature, but we know that neurons can be multifaceted, in that they fire in response to many different types of features: for example, a grocery store class neuron must activate either for rows of produce or for a storefront. Previous activation maximization techniques constructed images without regard for the multiple different facets of a neuron, creating inappropriate mixes of colors, parts of objects, scales, orientations, etc. Here, we introduce an algorithm that explicitly uncovers the multiple facets of each neuron by producing a synthetic visualization of each of the types of images that activate a neuron. We also introduce regularization methods that produce state-of-the-art results in terms of the interpretability of images obtained by activation maximization. By separately synthesizing each type of image a neuron fires in response to, the visualizations have more appropriate colors and coherent global structure. Multifaceted feature visualization thus provides a clearer and more comprehensive description of the role of each neuron.

Representation-level unlearning intervenes on the intermediate hidden states of LLMs. Although knowledge is distributed across layers, existing methods operate at a single fixed layer for the entire forget set. We ask whether such a fixed layer is sufficient. To answer this, we design a hijacking experiment that grafts hidden states of the target model into an oracle trained only on the retain set. The oracle cannot produce the forget answer on its own, yet it produces the answer from the grafted state. Thus, the answer is formed at an intermediate layer and merely read out afterward, so unlearning should focus on formation, not readout. Moreover, the layer where formation ends varies widely across inputs. Motivated by these findings, we propose Targeted Unlearning at Layers Identified Per-input (TULIP). For each input, TULIP uses the logit lens to locate the formation-readout boundary and removes the hidden state's alignment with the forget answer's unembedding vector there. TULIP consistently outperforms output- and representation-level baselines on TOFU, PISTOL, and WMDP across Llama, Qwen, and Zephyr models. It also remains robust to paraphrase and quantization attacks. Beyond standalone use, its per-input layer selection serves as a plug-and-play component that further improves existing methods.

Recent studies show that LLM hidden states encode reward-related information, such as answer correctness and model confidence. However, existing approaches typically fit black-box probes on the full hidden states, offering little insight into how this information is structured across neurons. In this paper, we show that reward-related information is concentrated in a sparse subset of neurons. Using simple probing, we identify two types of neurons: value neurons, whose activations predict state value, and dopamine neurons, whose activations encode step-level temporal difference (TD) errors. Together, these neurons form a sparse reward subsystem within LLM hidden states. These names are drawn by analogy with neuroscience, where value neurons and dopamine neurons in the biological reward subsystem also encode value and reward prediction errors, respectively. We demonstrate that value neurons are robust and transferable across diverse datasets and models, and provide causal evidence that they encode reward-related information. Finally, we show applications of the reward subsystem: value neurons serve as effective predictors of model confidence, and dopamine neurons can function as a process reward model (PRM) to guide inference-time search.

Deep neural networks (DNNs) have demonstrated state-of-the-art results on many pattern recognition tasks, especially vision classification problems. Understanding the inner workings of such computational brains is both fascinating basic science that is interesting in its own right - similar to why we study the human brain - and will enable researchers to further improve DNNs. One path to understanding how a neural network functions internally is to study what each of its neurons has learned to detect. One such method is called activation maximization (AM), which synthesizes an input (e.g. an image) that highly activates a neuron. Here we dramatically improve the qualitative state of the art of activation maximization by harnessing a powerful, learned prior: a deep generator network (DGN). The algorithm (1) generates qualitatively state-of-the-art synthetic images that look almost real, (2) reveals the features learned by each neuron in an interpretable way, (3) generalizes well to new datasets and somewhat well to different network architectures without requiring the prior to be relearned, and (4) can be considered as a high-quality generative method (in this case, by generating novel, creative, interesting, recognizable images).

Fine-tuned large language models can exhibit reward-hacking behavior arising from emergent misalignment, which is difficult to detect from final outputs alone. While prior work has studied reward hacking at the level of completed responses, it remains unclear whether such behavior can be identified during generation. We propose an activation-based monitoring approach that detects reward-hacking signals from internal representations as a model generates its response. Our method trains sparse autoencoders on residual stream activations and applies lightweight linear classifiers to produce token-level estimates of reward-hacking activity. Across multiple model families and fine-tuning mixtures, we find that internal activation patterns reliably distinguish reward-hacking from benign behavior, generalize to unseen mixed-policy adapters, and exhibit model-dependent temporal structure during chain-of-thought reasoning. Notably, reward-hacking signals often emerge early, persist throughout reasoning, and can be amplified by increased test-time compute in the form of chain-of-thought prompting under weakly specified reward objectives. These results suggest that internal activation monitoring provides a complementary and earlier signal of emergent misalignment than output-based evaluation, supporting more robust post-deployment safety monitoring for fine-tuned language models.

Activation monitoring, which probes a model's internal states using lightweight classifiers, is an emerging tool for AI safety. However, its worst-case robustness under a misalignment threat model--where a model might learn to actively conceal its internal states--remains untested. Focusing on this threat model, we ask: could a model learn to evade previously unseen activation monitors? Our core contribution is to stress-test the learnability of this behavior. We demonstrate that finetuning can create Neural Chameleons: models capable of zero-shot evading activation monitors. Specifically, we fine-tune an LLM to evade monitors for a set of benign concepts (e.g., languages, HTML) when conditioned on a trigger of the form: "You are being probed for {concept}". We show that this learned mechanism generalizes zero-shot: by substituting {concept} with a safety-relevant term like 'deception', the model successfully evades previously unseen safety monitors. We validate this phenomenon across diverse model families (Llama, Gemma, Qwen), showing that the evasion succeeds even against monitors trained post hoc on the model's frozen weights. This evasion is highly selective, targeting only the specific concept mentioned in the trigger, and having a modest impact on model capabilities on standard benchmarks. Using Gemma-2-9b-it as a case study, a mechanistic analysis reveals this is achieved via a targeted manipulation that moves activations into a low-dimensional subspace. While stronger defenses like monitor ensembles and non-linear classifiers show greater resilience, the model retains a non-trivial evasion capability. Our work provides a proof-of-concept for this failure mode and a tool to evaluate the worst-case robustness of monitoring techniques against misalignment threat models.

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.

Although Large Language Models (LLMs) have demonstrated impressive capabilities across a wide range of tasks, growing concerns have emerged over the misuse of sensitive, copyrighted, or harmful data during training. To address these concerns, unlearning techniques have been developed to remove the influence of specific data without retraining from scratch. However, this paper reveals a critical vulnerability in fine-tuning-based unlearning: a malicious user can craft a manipulated forgetting request that stealthily degrades the model's utility for benign users. We demonstrate this risk through a red-teaming Stealthy Attack (SA), which is inspired by two key limitations of existing unlearning (the inability to constrain the scope of unlearning effect and the failure to distinguish benign tokens from unlearning signals). Prior work has shown that unlearned models tend to memorize forgetting data as unlearning signals, and respond with hallucinations or feigned ignorance when unlearning signals appear in the input. By subtly increasing the presence of common benign tokens in the forgetting data, SA enhances the connection between benign tokens and unlearning signals. As a result, when normal users include such tokens in their prompts, the model exhibits unlearning behaviors, leading to unintended utility degradation. To address this vulnerability, we propose Scope-aware Unlearning (SU), a lightweight enhancement that introduces a scope term into the unlearning objective, encouraging the model to localize the forgetting effect. Our method requires no additional data processing, integrates seamlessly with existing fine-tuning frameworks, and significantly improves robustness against SA. Extensive experiments validate the effectiveness of both SA and SU.

Multimodal Large Language Models (MLLMs) trained on massive data may memorize sensitive personal information and photos, posing serious privacy risks. To mitigate this, MLLM unlearning methods are proposed, which fine-tune MLLMs to reduce the ``forget'' sensitive information. However, it remains unclear whether the knowledge has been truly forgotten or just hidden in the model. Therefore, we propose to study a novel problem of LLM unlearning attack, which aims to recover the unlearned knowledge of an unlearned LLM. To achieve the goal, we propose a novel framework Stealthy Unlearning Attack (SUA) framework that learns a universal noise pattern. When applied to input images, this noise can trigger the model to reveal unlearned content. While pixel-level perturbations may be visually subtle, they can be detected in the semantic embedding space, making such attacks vulnerable to potential defenses. To improve stealthiness, we introduce an embedding alignment loss that minimizes the difference between the perturbed and denoised image embeddings, ensuring the attack is semantically unnoticeable. Experimental results show that SUA can effectively recover unlearned information from MLLMs. Furthermore, the learned noise generalizes well: a single perturbation trained on a subset of samples can reveal forgotten content in unseen images. This indicates that knowledge reappearance is not an occasional failure, but a consistent behavior.

Large vision-language models (LVLMs), designed to interpret and respond to human instructions, occasionally generate hallucinated or harmful content due to inappropriate instructions. This study uses linear probing to shed light on the hidden knowledge at the output layer of LVLMs. We demonstrate that the logit distributions of the first tokens contain sufficient information to determine whether to respond to the instructions, including recognizing unanswerable visual questions, defending against multi-modal jailbreaking attack, and identifying deceptive questions. Such hidden knowledge is gradually lost in logits of subsequent tokens during response generation. Then, we illustrate a simple decoding strategy at the generation of the first token, effectively improving the generated content. In experiments, we find a few interesting insights: First, the CLIP model already contains a strong signal for solving these tasks, indicating potential bias in the existing datasets. Second, we observe performance improvement by utilizing the first logit distributions on three additional tasks, including indicting uncertainty in math solving, mitigating hallucination, and image classification. Last, with the same training data, simply finetuning LVLMs improve models' performance but is still inferior to linear probing on these tasks.

A neural network trained on a classification dataset often exhibits a higher vector norm of hidden layer features for in-distribution (ID) samples, while producing relatively lower norm values on unseen instances from out-of-distribution (OOD). Despite this intriguing phenomenon being utilized in many applications, the underlying cause has not been thoroughly investigated. In this study, we demystify this very phenomenon by scrutinizing the discriminative structures concealed in the intermediate layers of a neural network. Our analysis leads to the following discoveries: (1) The feature norm is a confidence value of a classifier hidden in the network layer, specifically its maximum logit. Hence, the feature norm distinguishes OOD from ID in the same manner that a classifier confidence does. (2) The feature norm is class-agnostic, thus it can detect OOD samples across diverse discriminative models. (3) The conventional feature norm fails to capture the deactivation tendency of hidden layer neurons, which may lead to misidentification of ID samples as OOD instances. To resolve this drawback, we propose a novel negative-aware norm (NAN) that can capture both the activation and deactivation tendencies of hidden layer neurons. We conduct extensive experiments on NAN, demonstrating its efficacy and compatibility with existing OOD detectors, as well as its capability in label-free environments.

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.

Large Vision-Language Models (LVLMs) generate contextually relevant responses by jointly interpreting visual and textual inputs. However, our finding reveals they often mistakenly perceive text inputs lacking visual evidence as being part of the image, leading to erroneous responses. In light of this finding, we probe whether LVLMs possess an internal capability to determine if textual concepts are grounded in the image, and discover a specific subset of Feed-Forward Network (FFN) neurons, termed Visual Absence-aware (VA) neurons, that consistently signal the visual absence through a distinctive activation pattern. Leveraging these patterns, we develop a detection module that systematically classifies whether an input token is visually grounded. Guided by its prediction, we propose a method to refine the outputs by reinterpreting question prompts or replacing the detected absent tokens during generation. Extensive experiments show that our method effectively mitigates the models' tendency to falsely presume the visual presence of text input and its generality across various LVLMs.

The memorization of training data by Large Language Models (LLMs) poses significant risks, including privacy leaks and the regurgitation of copyrighted content. Activation steering, a technique that directly intervenes in model activations, has emerged as a promising approach for manipulating LLMs. In this work, we explore the effectiveness of activation steering in reducing memorization while preserving generalization capabilities. We conduct empirical evaluations using a controlled memorization benchmark of literary material and demonstrate that our method successfully suppresses memorized content with minimal degradation in model performance in Gemma. Additionally, we analyze the trade-offs between suppression effectiveness and linguistic fluency, highlighting the advantages and limitations of activation-based interventions. Our findings contribute to ongoing efforts in developing safer and more privacy-preserving LLMs by providing a practical and efficient mechanism to mitigate unintended memorization.

Current evaluations of multimodal unlearning typically focus on whether a model stops outputting target information. This surface-level criterion, however, masks two critical failure modes. First, knowledge suppression techniques often induce confabulation, leading the model to hallucinate facts about the unlearned subject. Second, refusal training merely alters observable behavior without erasing the underlying knowledge; such models might decline open queries but still identify correct answers under multiple-choice probing. To overcome these limitations, we propose Forget-Aware Self-Distillation (FASD). This method is designed to achieve honest ignorance, a state where the model genuinely loses the target knowledge and learns to acknowledge its absence. FASD separates these goals into two complementary channels. The behavior channel distills appropriate refusals from an ignorance-prompted copy of the model itself. Crucially, target tokens are masked from the teacher distribution to avoid accidentally reteaching the forgotten concepts. Simultaneously, the knowledge channel directly suppresses ground-truth answer spans independently of the model's generated responses. Experiments demonstrate that FASD successfully unlearns target knowledge, prevents information leakage in open-ended tasks, and preserves performance on retained data. The resulting model not only forgets the specified information but accurately recognizes the boundaries of its own knowledge. Code and data will be released upon publication.

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.

Backdoor attacks pose a serious security threat for training neural networks as they surreptitiously introduce hidden functionalities into a model. Such backdoors remain silent during inference on clean inputs, evading detection due to inconspicuous behavior. However, once a specific trigger pattern appears in the input data, the backdoor activates, causing the model to execute its concealed function. Detecting such poisoned samples within vast datasets is virtually impossible through manual inspection. To address this challenge, we propose a novel approach that enables model training on potentially poisoned datasets by utilizing the power of recent diffusion models. Specifically, we create synthetic variations of all training samples, leveraging the inherent resilience of diffusion models to potential trigger patterns in the data. By combining this generative approach with knowledge distillation, we produce student models that maintain their general performance on the task while exhibiting robust resistance to backdoor triggers.

Refusal on harmful prompts is a key safety behaviour in instruction-tuned large language models (LLMs), yet the internal causes of this behaviour remain poorly understood. We study two public instruction-tuned models, Gemma-2-2B-IT and LLaMA-3.1-8B-IT, using sparse autoencoders (SAEs) trained on residual-stream activations. Given a harmful prompt, we search the SAE latent space for feature sets whose ablation flips the model from refusal to compliance, demonstrating causal influence and creating a jailbreak. Our search proceeds in three stages: (1) Refusal Direction: find a refusal-mediating direction and collect SAE features near that direction; (2) Greedy Filtering: prune to a minimal set; and (3) Interaction Discovery: fit a factorization machine (FM) that captures nonlinear interactions among the remaining active features and the minimal set. This pipeline yields a broad set of jailbreak-critical features, offering insight into the mechanistic basis of refusal. Moreover, we find evidence of redundant features that remain dormant unless earlier features are suppressed. Our findings highlight the potential for fine-grained auditing and targeted intervention in safety behaviours by manipulating the interpretable latent space.

We consider representation misdirection (RM), a class of LLM unlearning methods that achieves forgetting by manipulating the forget-representations, that is, latent representations of forget samples. Despite being important, the roles of target vectors used in RM, however, remain underexplored. Here, we approach and revisit RM through the lens of the linear representation hypothesis. Specifically, if one can somehow identify a one-dimensional representation corresponding to a high-level concept, the linear representation hypothesis enables linear operations on this concept vector within the forget-representation space. Under this view, we hypothesize that, beyond forgetting, machine unlearning elicits controllable side behaviors and stronger side capabilities corresponding to the high-level concept. Our hypothesis is empirically validated across a wide range of tasks, including behavioral control (e.g., controlling unlearned models' truth, sentiment, and refusal) and capability enhancement (e.g., improving unlearned models' in-context learning capability). Our findings reveal that this fairly attractive phenomenon could be either a hidden risk if misused or a mechanism that can be harnessed for developing models that require stronger capabilities and controllable behaviors.

Large language model (LLM) unlearning has become a critical mechanism for removing undesired data, knowledge, or behaviors from pre-trained models while retaining their general utility. Yet, with the rise of open-weight LLMs, we ask: can the unlearning process itself be backdoored, appearing successful under normal conditions yet reverting to pre-unlearned behavior when a hidden trigger is activated? Drawing inspiration from classical backdoor attacks that embed triggers into training data to enforce specific behaviors, we investigate backdoor unlearning, where models forget as intended in the clean setting but recover forgotten knowledge when the trigger appears. We show that designing such attacks presents unique challenges, hinging on where triggers are placed and how backdoor training is reinforced. We uncover a strong link between backdoor efficacy and the attention sink phenomenon, i.e., shallow input tokens consistently attract disproportionate attention in LLMs. Our analysis reveals that these attention sinks serve as gateways for backdoor unlearning: placing triggers at sink positions and aligning their attention values markedly enhances backdoor persistence. Extensive experiments validate these findings, showing that attention-sink-guided backdoor unlearning reliably restores forgotten knowledge in the presence of backdoor triggers, while behaving indistinguishably from a normally unlearned model when triggers are absent. Code is available at https://github.com/OPTML-Group/Unlearn-Backdoor.

Multimodal LLMs have reached remarkable levels of proficiency in understanding multimodal inputs, driving extensive research to develop increasingly powerful models. However, much less attention has been paid to understanding and explaining the underlying mechanisms of these models. Most existing explainability research examines these models only in their final states, overlooking the dynamic representational shifts that occur during training. In this work, we systematically analyze the evolution of hidden state representations to reveal how fine-tuning alters the internal structure of a model to specialize in new multimodal tasks. Using a concept-based approach, we map hidden states to interpretable visual and textual concepts, enabling us to trace changes in encoded concepts across modalities as training progresses. We also demonstrate the use of shift vectors to capture these concepts changes. These shift vectors allow us to recover fine-tuned concepts by shifting those in the original model. Finally, we explore the practical impact of our findings on model steering, showing that we can adjust multimodal LLMs behaviors without any training, such as modifying answer types, captions style, or biasing the model toward specific responses. Our work sheds light on how multimodal representations evolve through fine-tuning and offers a new perspective for interpreting model adaptation in multimodal tasks. The code for this project is publicly available at https://github.com/mshukor/xl-vlms.