Daily Papers

Language models exhibit remarkable robustness, continuing to produce coherent text even when their activations are perturbed by interventions like linear steering. We hypothesize that this robustness is a result of passive dynamics, i.e., constraining mechanisms in the forward pass that funnel activations toward "good" regions that produce coherent outputs. To investigate these hypothesized error-correcting mechanisms, we probe the geometry of language model activation space by observing the action of model layers on low-dimensional curves. In doing so, we discover that model activations occur within a cluster of distinct attracting basins, which differentiate natural activations geometrically from distributionally similar synthetic activations. Applying this lens to language model steering, we observe feature-specific basins along semantic steering directions, and find that steering moves activations between these basins. To demonstrate the active role of this geometry in neural computation, we show that adaptively modulating steering strength to transport activations across basins improves inter-language steering, significantly increasing the probability of sampling tokens from the target language compared to fixed-strength steering. Our findings establish analysis of activation space geometry as a promising approach to interpreting and controlling language models.

Language models, while capable of generating remarkably coherent and seemingly accurate text, can occasionally produce undesirable content, including harmful or toxic outputs. In this paper, we present a new two-stage approach to detect and mitigate undesirable content generations by rectifying activations. First, we train an ensemble of layerwise classifiers to detect undesirable content using activations by minimizing a smooth surrogate of the risk-aware score. Then, for detected undesirable contents, we propose layerwise distributional steering policies that transform the attention heads. These policies are computed through principled semidefinite programming, which aims to minimally perturb the attention distribution while probabilistically guaranteeing the effectiveness of the editions. Empirical evaluations across multiple language models and datasets show that our method outperforms baselines in reducing the generation of undesirable output.

We explore a class of adversarial attacks targeting the activations of language models. By manipulating a relatively small subset of model activations, a, we demonstrate the ability to control the exact prediction of a significant number (in some cases up to 1000) of subsequent tokens t. We empirically verify a scaling law where the maximum number of target tokens t_max predicted depends linearly on the number of tokens a whose activations the attacker controls as t_max = kappa a. We find that the number of bits of control in the input space needed to control a single bit in the output space (what we call attack resistance chi) is remarkably constant between approx 16 and approx 25 over 2 orders of magnitude of model sizes for different language models. Compared to attacks on tokens, attacks on activations are predictably much stronger, however, we identify a surprising regularity where one bit of input steered either via activations or via tokens is able to exert control over a similar amount of output bits. This gives support for the hypothesis that adversarial attacks are a consequence of dimensionality mismatch between the input and output spaces. A practical implication of the ease of attacking language model activations instead of tokens is for multi-modal and selected retrieval models, where additional data sources are added as activations directly, sidestepping the tokenized input. This opens up a new, broad attack surface. By using language models as a controllable test-bed to study adversarial attacks, we were able to experiment with input-output dimensions that are inaccessible in computer vision, especially where the output dimension dominates.

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.

Language models often solve complex tasks by generating long reasoning chains, consisting of many steps with varying importance. While some steps are crucial for generating the final answer, others are removable. Determining which steps matter most, and why, remains an open question central to understanding how models process reasoning. We investigate if this question is best approached through model internals or through tokens of the reasoning chain itself. We find that model activations contain more information than tokens for identifying important reasoning steps. Crucially, by training probes on model activations to predict importance, we show that models encode an internal representation of step importance, even prior to the generation of subsequent steps. The internal representations of importance in different models yield high agreement on which steps are important. The representation is distributed across layers, and does not correlate with surface-level features, such as a step's relative position or its length. Our findings suggest that analyzing activations can reveal aspects of reasoning that surface-level approaches fundamentally miss, indicating that reasoning analyses should look into model internals.

We introduce Inference-Time Intervention (ITI), a technique designed to enhance the truthfulness of large language models (LLMs). ITI operates by shifting model activations during inference, following a set of directions across a limited number of attention heads. This intervention significantly improves the performance of LLaMA models on the TruthfulQA benchmark. On an instruction-finetuned LLaMA called Alpaca, ITI improves its truthfulness from 32.5% to 65.1%. We identify a tradeoff between truthfulness and helpfulness and demonstrate how to balance it by tuning the intervention strength. ITI is minimally invasive and computationally inexpensive. Moreover, the technique is data efficient: while approaches like RLHF require extensive annotations, ITI locates truthful directions using only few hundred examples. Our findings suggest that LLMs may have an internal representation of the likelihood of something being true, even as they produce falsehoods on the surface.

Patients with depression present with diverse symptom profiles, yet clinical practice routinely reduces this variation to a single severity score. Large language models (LLMs) can potentially capture various symptoms and their severity from patient speech. However, how depressive symptoms are represented inside LLMs remains poorly understood, limiting clinical trust. To examine whether internal model activations match clinician judgment, we analyzed the residual stream of Gemma-3-27B-PT using mechanistic interpretability techniques. Recording activations across symptom descriptions drawn from validated clinical instruments, we found that symptom groups geometrically separated the most at layer 21 across multiple distance metrics. Using Semantic Projection, we then projected held-out naturalistic text onto Symptom Vectors constructed from these instruments. The resulting per-symptom coefficients preserved clinician-annotated rank ordering across mood, somatic, and suicidality axes. Furthermore, a single depression vector in Layer 21 separates held-out depressive from non-depressive text (AUC = 0.789), which can be used as an emotional valence gate that restricts symptom projection to depressive speech. These results reveal a decorrelated, clinician-aligned symptom signal readable directly from internal activations, offering a mechanistic foundation for interpretable depression-assessment tools.

Responsible practices for deploying language models include guiding models to recognize and refuse answering prompts that are considered unsafe, while complying with safe prompts. Achieving such behavior typically requires updating model weights, which is costly and inflexible. We explore opportunities to steering model activations at inference time, which does not require updating weights. Using sparse autoencoders, we identify and steer features in Phi-3 Mini that mediate refusal behavior. We find that feature steering can improve Phi-3 Minis robustness to jailbreak attempts across various harms, including challenging multi-turn attacks. However, we discover that feature steering can adversely affect overall performance on benchmarks. These results suggest that identifying steerable mechanisms for refusal via sparse autoencoders is a promising approach for enhancing language model safety, but that more research is needed to mitigate feature steerings adverse effects on performance.

Activation steering is a popular white-box control technique that modifies model activations to elicit an abstract change in its behavior. It has also become a standard tool in interpretability (e.g., probing truthfulness, or translating activations into human-readable explanations) and safety research (e.g., jailbreakability). However, it is unclear whether steered behavior is realizable by any textual prompt. In this work, we cast this question as a surjectivity problem: for a fixed model, does every steered activation admit a preimage under the model's natural forward pass? Under practical assumptions, we prove that activation steering pushes the residual stream off the manifold of states reachable from discrete prompts. Almost surely, no prompt can reproduce the same internal behavior induced by steering. We also illustrate this finding empirically across three widely used LLMs. Our results establish a formal separation between white-box steerability and black-box prompting. We therefore caution against interpreting the ease and success of activation steering as evidence of prompt-based interpretability or vulnerability, and argue for evaluation protocols that explicitly decouple white-box and black-box interventions.

Interpretability methods seek to understand language model representations, yet the outputs of most such methods -- circuits, vectors, scalars -- are not immediately human-interpretable. In response, we introduce LatentQA, the task of answering open-ended questions about model activations in natural language. Towards solving LatentQA, we propose Latent Interpretation Tuning (LIT), which finetunes a decoder LLM on a dataset of activations and associated question-answer pairs, similar to how visual instruction tuning trains on question-answer pairs associated with images. We use the decoder for diverse reading applications, such as extracting relational knowledge from representations or uncovering system prompts governing model behavior. Our decoder also specifies a differentiable loss that we use to control models, such as debiasing models on stereotyped sentences and controlling the sentiment of generations. Finally, we extend LatentQA to reveal harmful model capabilities, such as generating recipes for bioweapons and code for hacking.

We study how to characterize and predict the truthfulness of texts generated from large language models (LLMs), which serves as a crucial step in building trust between humans and LLMs. Although several approaches based on entropy or verbalized uncertainty have been proposed to calibrate model predictions, these methods are often intractable, sensitive to hyperparameters, and less reliable when applied in generative tasks with LLMs. In this paper, we suggest investigating internal activations and quantifying LLM's truthfulness using the local intrinsic dimension (LID) of model activations. Through experiments on four question answering (QA) datasets, we demonstrate the effectiveness ohttps://info.arxiv.org/help/prep#abstractsf our proposed method. Additionally, we study intrinsic dimensions in LLMs and their relations with model layers, autoregressive language modeling, and the training of LLMs, revealing that intrinsic dimensions can be a powerful approach to understanding LLMs.

The increasing capabilities of large generative models and their ever more widespread deployment have raised concerns about their reliability, safety, and potential misuse. To address these issues, recent works have proposed to control model generation by steering model activations in order to effectively induce or prevent the emergence of concepts or behaviors in the generated output. In this paper we introduce Activation Transport (AcT), a general framework to steer activations guided by optimal transport theory that generalizes many previous activation-steering works. AcT is modality-agnostic and provides fine-grained control over the model behavior with negligible computational overhead, while minimally impacting model abilities. We experimentally show the effectiveness and versatility of our approach by addressing key challenges in large language models (LLMs) and text-to-image diffusion models (T2Is). For LLMs, we show that AcT can effectively mitigate toxicity, induce arbitrary concepts, and increase their truthfulness. In T2Is, we show how AcT enables fine-grained style control and concept negation.

Sparse autoencoders (SAEs) are a popular technique for interpreting language model activations, and there is extensive recent work on improving SAE effectiveness. However, most prior work evaluates progress using unsupervised proxy metrics with unclear practical relevance. We introduce SAEBench, a comprehensive evaluation suite that measures SAE performance across seven diverse metrics, spanning interpretability, feature disentanglement and practical applications like unlearning. To enable systematic comparison, we open-source a suite of over 200 SAEs across eight recently proposed SAE architectures and training algorithms. Our evaluation reveals that gains on proxy metrics do not reliably translate to better practical performance. For instance, while Matryoshka SAEs slightly underperform on existing proxy metrics, they substantially outperform other architectures on feature disentanglement metrics; moreover, this advantage grows with SAE scale. By providing a standardized framework for measuring progress in SAE development, SAEBench enables researchers to study scaling trends and make nuanced comparisons between different SAE architectures and training methodologies. Our interactive interface enables researchers to flexibly visualize relationships between metrics across hundreds of open-source SAEs at: https://saebench.xyz

Sparse autoencoders (SAEs) are a popular method for decomposing Large Langage Models (LLM) activations into interpretable latents. However, due to their substantial training cost, most academic research uses open-source SAEs which are only available for a restricted set of models of up to 27B parameters. SAE latents are also learned from a dataset of activations, which means they do not transfer between models. Motivated by relative representation similarity measures, we introduce Inference-Time Decomposition of Activations (ITDA) models, an alternative method for decomposing language model activations. To train an ITDA, we greedily construct a dictionary of language model activations on a dataset of prompts, selecting those activations which were worst approximated by matching pursuit on the existing dictionary. ITDAs can be trained in just 1% of the time required for SAEs, using 1% of the data. This allowed us to train ITDAs on Llama-3.1 70B and 405B on a single consumer GPU. ITDAs can achieve similar reconstruction performance to SAEs on some target LLMs, but generally incur a performance penalty. However, ITDA dictionaries enable cross-model comparisons, and a simple Jaccard similarity index on ITDA dictionaries outperforms existing methods like CKA, SVCCA, and relative representation similarity metrics. ITDAs provide a cheap alternative to SAEs where computational resources are limited, or when cross model comparisons are necessary. Code available at https://github.com/pleask/itda.

Aligned representations across languages is a desired property in multilingual large language models (mLLMs), as alignment can improve performance in cross-lingual tasks. Typically alignment requires fine-tuning a model, which is computationally expensive, and sizable language data, which often may not be available. A data-efficient alternative to fine-tuning is model interventions -- a method for manipulating model activations to steer generation into the desired direction. We analyze the effect of a popular intervention (finding experts) on the alignment of cross-lingual representations in mLLMs. We identify the neurons to manipulate for a given language and introspect the embedding space of mLLMs pre- and post-manipulation. We show that modifying the mLLM's activations changes its embedding space such that cross-lingual alignment is enhanced. Further, we show that the changes to the embedding space translate into improved downstream performance on retrieval tasks, with up to 2x improvements in top-1 accuracy on cross-lingual retrieval.

Large Language Models (LLMs) can acquire unintended biases from seemingly benign training data even without explicit cues or malicious content. Existing methods struggle to detect such risks before fine-tuning, making post hoc evaluation costly and inefficient. To address this challenge, we introduce Data2Behavior, a new task for predicting unintended model behaviors prior to training. We also propose Manipulating Data Features (MDF), a lightweight approach that summarizes candidate data through their mean representations and injects them into the forward pass of a base model, allowing latent statistical signals in the data to shape model activations and reveal potential biases and safety risks without updating any parameters. MDF achieves reliable prediction while consuming only about 20% of the GPU resources required for fine-tuning. Experiments on Qwen3-14B, Qwen2.5-32B-Instruct, and Gemma-3-12b-it confirm that MDF can anticipate unintended behaviors and provide insight into pre-training vulnerabilities.

Disentangling model activations into meaningful features is a central problem in interpretability. However, the absence of ground-truth for these features in realistic scenarios makes validating recent approaches, such as sparse dictionary learning, elusive. To address this challenge, we propose a framework for evaluating feature dictionaries in the context of specific tasks, by comparing them against supervised feature dictionaries. First, we demonstrate that supervised dictionaries achieve excellent approximation, control, and interpretability of model computations on the task. Second, we use the supervised dictionaries to develop and contextualize evaluations of unsupervised dictionaries along the same three axes. We apply this framework to the indirect object identification (IOI) task using GPT-2 Small, with sparse autoencoders (SAEs) trained on either the IOI or OpenWebText datasets. We find that these SAEs capture interpretable features for the IOI task, but they are less successful than supervised features in controlling the model. Finally, we observe two qualitative phenomena in SAE training: feature occlusion (where a causally relevant concept is robustly overshadowed by even slightly higher-magnitude ones in the learned features), and feature over-splitting (where binary features split into many smaller, less interpretable features). We hope that our framework will provide a useful step towards more objective and grounded evaluations of sparse dictionary learning methods.

Decomposing model activations into interpretable components is a key open problem in mechanistic interpretability. Sparse autoencoders (SAEs) are a popular method for decomposing the internal activations of trained transformers into sparse, interpretable features, and have been applied to MLP layers and the residual stream. In this work we train SAEs on attention layer outputs and show that also here SAEs find a sparse, interpretable decomposition. We demonstrate this on transformers from several model families and up to 2B parameters. We perform a qualitative study of the features computed by attention layers, and find multiple families: long-range context, short-range context and induction features. We qualitatively study the role of every head in GPT-2 Small, and estimate that at least 90% of the heads are polysemantic, i.e. have multiple unrelated roles. Further, we show that Sparse Autoencoders are a useful tool that enable researchers to explain model behavior in greater detail than prior work. For example, we explore the mystery of why models have so many seemingly redundant induction heads, use SAEs to motivate the hypothesis that some are long-prefix whereas others are short-prefix, and confirm this with more rigorous analysis. We use our SAEs to analyze the computation performed by the Indirect Object Identification circuit (Wang et al.), validating that the SAEs find causally meaningful intermediate variables, and deepening our understanding of the semantics of the circuit. We open-source the trained SAEs and a tool for exploring arbitrary prompts through the lens of Attention Output SAEs.

We investigate feature universality in large language models (LLMs), a research field that aims to understand how different models similarly represent concepts in the latent spaces of their intermediate layers. Demonstrating feature universality allows discoveries about latent representations to generalize across several models. However, comparing features across LLMs is challenging due to polysemanticity, in which individual neurons often correspond to multiple features rather than distinct ones. This makes it difficult to disentangle and match features across different models. To address this issue, we employ a method known as dictionary learning by using sparse autoencoders (SAEs) to transform LLM activations into more interpretable spaces spanned by neurons corresponding to individual features. After matching feature neurons across models via activation correlation, we apply representational space similarity metrics like Singular Value Canonical Correlation Analysis to analyze these SAE features across different LLMs. Our experiments reveal significant similarities in SAE feature spaces across various LLMs, providing new evidence for feature universality.

Reasoning models generate chain-of-thought (CoT) tokens before their final output, but how this affects their vulnerability to jailbreak attacks remains unclear. While traditional language models make refusal decisions at the prompt-response boundary, we find evidence that DeepSeek-R1-Distill-Llama-8B makes these decisions within its CoT generation. We identify a linear direction in activation space during CoT token generation that predicts whether the model will refuse or comply -- termed the "caution" direction because it corresponds to cautious reasoning patterns in the generated text. Ablating this direction from model activations increases harmful compliance, effectively jailbreaking the model. We additionally show that intervening only on CoT token activations suffices to control final outputs, and that incorporating this direction into prompt-based attacks improves success rates. Our findings suggest that the chain-of-thought itself is a promising new target for adversarial manipulation in reasoning models. Code available at https://github.com/ky295/reasoning-manipulation

Object binding, the brain's ability to bind the many features that collectively represent an object into a coherent whole, is central to human cognition. It groups low-level perceptual features into high-level object representations, stores those objects efficiently and compositionally in memory, and supports human reasoning about individual object instances. While prior work often imposes object-centric attention (e.g., Slot Attention) explicitly to probe these benefits, it remains unclear whether this ability naturally emerges in pre-trained Vision Transformers (ViTs). Intuitively, they could: recognizing which patches belong to the same object should be useful for downstream prediction and thus guide attention. Motivated by the quadratic nature of self-attention, we hypothesize that ViTs represent whether two patches belong to the same object, a property we term IsSameObject. We decode IsSameObject from patch embeddings across ViT layers using a similarity probe, which reaches over 90% accuracy. Crucially, this object-binding capability emerges reliably in self-supervised ViTs (DINO, MAE, CLIP), but markedly weaker in ImageNet-supervised models, suggesting that binding is not a trivial architectural artifact, but an ability acquired through specific pretraining objectives. We further discover that IsSameObject is encoded in a low-dimensional subspace on top of object features, and that this signal actively guides attention. Ablating IsSameObject from model activations degrades downstream performance and works against the learning objective, implying that emergent object binding naturally serves the pretraining objective. Our findings challenge the view that ViTs lack object binding and highlight how symbolic knowledge of "which parts belong together" emerges naturally in a connectionist system.

Large vision-language models (LVLMs) have rapidly advanced across various domains, yet they still lag behind strong text-only large language models (LLMs) on tasks that require multi-step inference and compositional decision-making. Motivated by their shared transformer architectures, we investigate whether the two model families rely on common internal computation for such inference. At the neuron level, we uncover a surprisingly large overlap: more than half of the top-activated units during multi-step inference are shared between representative LLMs and LVLMs, revealing a modality-invariant inference subspace. Through causal probing via activation amplification, we further show that these shared neurons encode consistent and interpretable concept-level effects, demonstrating their functional contribution to inference. Building on this insight, we propose Shared Neuron Low-Rank Fusion (SNRF), a parameter-efficient framework that transfers mature inference circuitry from LLMs to LVLMs. SNRF profiles cross-model activations to identify shared neurons, computes a low-rank approximation of inter-model weight differences, and injects these updates selectively within the shared-neuron subspace. This mechanism strengthens multimodal inference performance with minimal parameter changes and requires no large-scale multimodal fine-tuning. Across diverse mathematics and perception benchmarks, SNRF consistently enhances LVLM inference performance while preserving perceptual capabilities. Our results demonstrate that shared neurons form an interpretable bridge between LLMs and LVLMs, enabling low-cost transfer of inference ability into multimodal models. Our code is available at [https://github.com/chenhangcuisg-code/Do-LLMs-VLMs-Share-Neurons](https://github.com/chenhangcuisg-code/Do-LLMs-VLMs-Share-Neurons).

As LLM agents are increasingly deployed in multi-agent systems, they introduce risks of covert coordination that may evade standard forms of human oversight. While linear probes on model activations have shown promise for detecting deception in single-agent settings, collusion is inherently a multi-agent phenomenon, and the use of internal representations for detecting collusion between agents remains unexplored. We introduce NARCBench, a benchmark for evaluating collusion detection under environment distribution shift, and propose five probing techniques that aggregate per-agent deception scores to classify scenarios at the group level. Our probes achieve 1.00 AUROC in-distribution and 0.60--0.86 AUROC when transferred zero-shot to structurally different multi-agent scenarios and a steganographic blackjack card-counting task. We find that no single probing technique dominates across all collusion types, suggesting that different forms of collusion manifest differently in activation space. We also find preliminary evidence that this signal is localised at the token level, with the colluding agent's activations spiking specifically when processing the encoded parts of their partner's message. This work takes a step toward multi-agent interpretability: extending white-box inspection from single models to multi-agent contexts, where detection requires aggregating signals across agents. These results suggest that model internals provide a complementary signal to text-level monitoring for detecting multi-agent collusion, particularly for organisations with access to model activations. Code and data are available at https://github.com/aaronrose227/narcbench.

Sparse autoencoders (SAEs) represent model activations with sparse, linearly decoded features, but the downstream effects of those features need not compose additively. We introduce SAE Additivity Residuals (SAR), which measure the behavioral variation left unexplained above a chosen interaction order. Under independent feature subset sampling, population SAR is exactly the normalized energy of interactions above that order and bounds coalition ordering error for the optimal low order approximation. To distinguish compositionality associated with the SAE representation from that induced by downstream computation, we compare learned SAE coordinates with sum preserving random rotations of the same selected subspace and find lower first order residuals for the learned coordinates across all 18 model and behavior settings. The coordinates are nevertheless not fully additive: pairwise terms improve held out coalition ordering in every setting and outperform parameter matched models using randomly selected third order terms. In a combined steering experiment, pairwise modeling raises rank fidelity from 0.5874 to 0.7790 and reduces ordering error from 0.1444 to 0.0606. Together, these results show that SAE coordinates support comparatively compositional behavioral descriptions while faithful joint explanations and interventions still require interactions beyond individual feature effects. Code is available at https://github.com/OpenSAE/SAE-Behavioral-Compositionality.

Sparse autoencoders (SAEs) decompose language model activations into sparse features, but prompt transformations designed to preserve the underlying request can redistribute activity across them. A feature change observed on one transformed prompt need not recur, and even a recurring change need not affect a downstream computation. We introduce the redistribution sensitivity score (RSS), which identifies large feature responses that recur in a consistent direction, and evaluate the resulting supports on variants excluded from selection. Across 27 pretrained SAEs spanning eight language models, the tested prompt transformations produce a reproducible redistribution profile on these excluded variants. We then separate recurring representation change from signed readout contribution and decision failure. For a fixed linear readout, an exact decomposition shows how response magnitude, readout coefficients, directional alignment, and the clean margin determine downstream consequences. In two frozen safety readouts, combining RSS support erosion with the full clean margin yields recurrent failure AUROCs of 0.935 and 0.958 using median supports of 20 and 19 features, compared with 0.965 and 0.964 using full erosion. RSS therefore provides a way to identify recurring sparse representation changes and quantify their contribution to a downstream readout without equating representation change with failure. Code is available at https://github.com/OpenSAE/Redistribution-Sensitivity-Score

We introduce Contrastive Activation Addition (CAA), an innovative method for steering language models by modifying activations during their forward passes. CAA computes ``steering vectors'' by averaging the difference in residual stream activations between pairs of positive and negative examples of a particular behavior such as factual versus hallucinatory responses. During inference, these steering vectors are added at all token positions after the user's prompt with either a positive or negative coefficient, allowing precise control over the degree of the targeted behavior. We evaluate CAA's effectiveness on Llama 2 Chat using both multiple-choice behavioral question datasets and open-ended generation tasks. We demonstrate that CAA significantly alters model behavior, outperforms traditional methods like finetuning and few-shot prompting, and minimally reduces capabilities. Moreover, by employing various activation space interpretation methods, we gain deeper insights into CAA's mechanisms. CAA both accurately steers model outputs and also sheds light on how high-level concepts are represented in Large Language Models (LLMs).

We show that existing unsupervised methods on large language model (LLM) activations do not discover knowledge -- instead they seem to discover whatever feature of the activations is most prominent. The idea behind unsupervised knowledge elicitation is that knowledge satisfies a consistency structure, which can be used to discover knowledge. We first prove theoretically that arbitrary features (not just knowledge) satisfy the consistency structure of a particular leading unsupervised knowledge-elicitation method, contrast-consistent search (Burns et al. - arXiv:2212.03827). We then present a series of experiments showing settings in which unsupervised methods result in classifiers that do not predict knowledge, but instead predict a different prominent feature. We conclude that existing unsupervised methods for discovering latent knowledge are insufficient, and we contribute sanity checks to apply to evaluating future knowledge elicitation methods. Conceptually, we hypothesise that the identification issues explored here, e.g. distinguishing a model's knowledge from that of a simulated character's, will persist for future unsupervised methods.

Sparse Autoencoders (SAEs) have emerged as a promising approach to decompose the activations of Large Language Models (LLMs) into human-interpretable latents. In this paper, we pose two questions. First, to what extent do SAEs extract monosemantic and interpretable latents? Second, to what extent does varying the sparsity or the size of the SAE affect monosemanticity / interpretability? By investigating these questions in the context of a simple first-letter identification task where we have complete access to ground truth labels for all tokens in the vocabulary, we are able to provide more detail than prior investigations. Critically, we identify a problematic form of feature-splitting we call feature absorption where seemingly monosemantic latents fail to fire in cases where they clearly should. Our investigation suggests that varying SAE size or sparsity is insufficient to solve this issue, and that there are deeper conceptual issues in need of resolution.

Large language models can resist task-misaligned activation steering during inference, sometimes recovering mid-generation to produce improved responses even when steering remains active. We term this Endogenous Steering Resistance (ESR). Using sparse autoencoder (SAE) latents to steer model activations, we find that Llama-3.3-70B shows substantial ESR, while smaller models from the Llama-3 and Gemma-2 families exhibit the phenomenon less frequently. We identify 26 SAE latents that activate differentially during off-topic content and are causally linked to ESR in Llama-3.3-70B. Zero-ablating these latents reduces the multi-attempt rate by 25%, providing causal evidence for dedicated internal consistency-checking circuits. We demonstrate that ESR can be deliberately enhanced through both prompting and training: meta-prompts instructing the model to self-monitor increase the multi-attempt rate by 4x for Llama-3.3-70B, and fine-tuning on self-correction examples successfully induces ESR-like behavior in smaller models. These findings have dual implications: ESR could protect against adversarial manipulation but might also interfere with beneficial safety interventions that rely on activation steering. Understanding and controlling these resistance mechanisms is important for developing transparent and controllable AI systems. Code is available at github.com/agencyenterprise/endogenous-steering-resistance.

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.

Sparse autoencoders (SAEs) decompose language model activations into interpretable features, but existing methods reveal only which features activate, not which change model outputs when amplified. We introduce Control Reinforcement Learning (CRL), which trains a policy to select SAE features for steering at each token, producing interpretable intervention logs: the learned policy identifies features that change model outputs when amplified. Adaptive Feature Masking encourages diverse feature discovery while preserving singlefeature interpretability. The framework yields new analysis capabilities: branch point tracking locates tokens where feature choice determines output correctness; critic trajectory analysis separates policy limitations from value estimation errors; layer-wise comparison reveals syntactic features in early layers and semantic features in later layers. On Gemma 2 2B across MMLU, BBQ, GSM8K, HarmBench, and XSTest, CRL achieves improvements while providing per-token intervention logs. These results establish learned feature steering as a mechanistic interpretability tool that complements static feature analysis with dynamic intervention probes

Both capability and safety benchmarks rest upon the assumption that the behavior of language models undergoing a test is informative about their behavior in deployment. This assumption can fail, should models infer that they are being evaluated and condition their response on such context. This hypothesis, termed ``evaluation awareness'', has been observed in frontier and open-weight language models alike. We provide a systematic study of this phenomenon, by probing for it across six language models (from four families and three sizes) and three metrics. More precisely, we examine whether (i) being under evaluation is linearly represented within the models' activations space, (ii) it is verbalized in their output tokens (as scored by an LLM-as-judge), and (iii) steering causally affects their behavior. For the open-checkpoint Olmo models, we further test these measures at every training stage. In doing so, we report that evaluation awareness is linearly decodable from the residual streams of every model (best AUROC geq 0.7). By contrast, these representations align only in part with verbalization: their correlations and mutual information are nonzero in some settings, yet vary substantially across models, layers, and readout choices. Nevertheless, steering along probe-derived directions can shift the verbalization scores. Finally, a comparison across the Olmo checkpoints reveals that evaluation awareness is already present within base models, becomes amplified throughout the stages of supervised fine-tuning, and remains stable thereafter---unlike the effects of steering, that grow more pronounced at every successive training stage. These results show the need for evaluations to account for the disjunction between what models represent internally, what they verbalize, and their steering.

Reliably controlling the behavior of large language models is a pressing open problem. Existing methods include supervised finetuning, reinforcement learning from human feedback, prompt engineering and guided decoding. We instead investigate activation engineering: modifying activations at inference-time to predictably alter model behavior. We bias the forward pass with a 'steering vector' implicitly specified through natural language. Past work learned these steering vectors; our Activation Addition (ActAdd) method instead computes them by taking the activation differences which result from pairs of prompts. We demonstrate ActAdd on GPT-2 on OpenWebText and ConceptNet, and replicate the effect on Llama-13B and GPT-J-6B. Our approach yields inference-time control over high-level properties of output & preserves performance on off-target topics. The method requires far less compute and implementation effort than finetuning and RLHF, allows for natural language specification by users, and its overhead scales naturally with model size.

Sparse autoencoders (SAEs) are a promising unsupervised approach for identifying causally relevant and interpretable linear features in a language model's (LM) activations. To be useful for downstream tasks, SAEs need to decompose LM activations faithfully; yet to be interpretable the decomposition must be sparse -- two objectives that are in tension. In this paper, we introduce JumpReLU SAEs, which achieve state-of-the-art reconstruction fidelity at a given sparsity level on Gemma 2 9B activations, compared to other recent advances such as Gated and TopK SAEs. We also show that this improvement does not come at the cost of interpretability through manual and automated interpretability studies. JumpReLU SAEs are a simple modification of vanilla (ReLU) SAEs -- where we replace the ReLU with a discontinuous JumpReLU activation function -- and are similarly efficient to train and run. By utilising straight-through-estimators (STEs) in a principled manner, we show how it is possible to train JumpReLU SAEs effectively despite the discontinuous JumpReLU function introduced in the SAE's forward pass. Similarly, we use STEs to directly train L0 to be sparse, instead of training on proxies such as L1, avoiding problems like shrinkage.

Sparse autoencoders (SAEs) are a promising technique for decomposing language model activations into interpretable linear features. However, current SAEs fall short of completely explaining model performance, resulting in "dark matter": unexplained variance in activations. This work investigates dark matter as an object of study in its own right. Surprisingly, we find that much of SAE dark matter--about half of the error vector itself and >90% of its norm--can be linearly predicted from the initial activation vector. Additionally, we find that the scaling behavior of SAE error norms at a per token level is remarkably predictable: larger SAEs mostly struggle to reconstruct the same contexts as smaller SAEs. We build on the linear representation hypothesis to propose models of activations that might lead to these observations, including postulating a new type of "introduced error"; these insights imply that the part of the SAE error vector that cannot be linearly predicted ("nonlinear" error) might be fundamentally different from the linearly predictable component. To validate this hypothesis, we empirically analyze nonlinear SAE error and show that 1) it contains fewer not yet learned features, 2) SAEs trained on it are quantitatively worse, 3) it helps predict SAE per-token scaling behavior, and 4) it is responsible for a proportional amount of the downstream increase in cross entropy loss when SAE activations are inserted into the model. Finally, we examine two methods to reduce nonlinear SAE error at a fixed sparsity: inference time gradient pursuit, which leads to a very slight decrease in nonlinear error, and linear transformations from earlier layer SAE outputs, which leads to a larger reduction.

Recent advancements in large language models (LLMs) have resulted in increasingly anthropomorphic language concerning the ability of LLMs to reason. Whether reasoning in LLMs should be understood to be inherently different is, however, widely debated. We propose utilizing a representation engineering approach wherein model activations are read from the residual stream of an LLM when processing a reasoning task. The activations are used to derive a control vector that is applied to the model as an inference-time intervention, modulating the representational space of the model, to improve performance on the specified task. We publish the code for deriving control vectors and analyzing model representations. The method allows us to improve performance on reasoning benchmarks and assess how control vectors influence the final logit distribution of a model via metrics such as KL divergence and entropy. We apply control vectors to Mistral-7B-Instruct and a range of Pythia models on an inductive, a deductive and mathematical reasoning task. We show that an LLM can, to a certain degree, be controlled to improve its perceived reasoning ability by modulating activations. The intervention is dependent upon the ability to reliably extract the model's typical state when correctly solving a task. Our results suggest that reasoning performance can be modulated in the same manner as other information-processing tasks performed by LLMs and demonstrate that we are capable of improving performance on specific tasks via a simple intervention on the residual stream with no additional training.

Large language models (LLMs) have made remarkable progress in various domains, yet they often suffer from repetitive text generation, a phenomenon we refer to as the "Repeat Curse". While previous studies have proposed decoding strategies to mitigate repetition, the underlying mechanism behind this issue remains insufficiently explored. In this work, we investigate the root causes of repetition in LLMs through the lens of mechanistic interpretability. Inspired by recent advances in Sparse Autoencoders (SAEs), which enable monosemantic feature extraction, we propose a novel approach, "Duplicatus Charm", to induce and analyze the Repeat Curse. Our method systematically identifies "Repetition Features" -the key model activations responsible for generating repetitive outputs. First, we locate the layers most involved in repetition through logit analysis. Next, we extract and stimulate relevant features using SAE-based activation manipulation. To validate our approach, we construct a repetition dataset covering token and paragraph level repetitions and introduce an evaluation pipeline to quantify the influence of identified repetition features. Furthermore, by deactivating these features, we have effectively mitigated the Repeat Curse.

Sparse autoencoders are a promising new approach for decomposing language model activations for interpretation and control. They have been applied successfully to vision transformer image encoders and to small-scale diffusion models. Inference-Time Decomposition of Activations (ITDA) is a recently proposed variant of dictionary learning that takes the dictionary to be a set of data points from the activation distribution and reconstructs them with gradient pursuit. We apply Sparse Autoencoders (SAEs) and ITDA to a large text-to-image diffusion model, Flux 1, and consider the interpretability of embeddings of both by introducing a visual automated interpretation pipeline. We find that SAEs accurately reconstruct residual stream embeddings and beat MLP neurons on interpretability. We are able to use SAE features to steer image generation through activation addition. We find that ITDA has comparable interpretability to SAEs.

AI models might use deceptive strategies as part of scheming or misaligned behaviour. Monitoring outputs alone is insufficient, since the AI might produce seemingly benign outputs while their internal reasoning is misaligned. We thus evaluate if linear probes can robustly detect deception by monitoring model activations. We test two probe-training datasets, one with contrasting instructions to be honest or deceptive (following Zou et al., 2023) and one of responses to simple roleplaying scenarios. We test whether these probes generalize to realistic settings where Llama-3.3-70B-Instruct behaves deceptively, such as concealing insider trading (Scheurer et al., 2023) and purposely underperforming on safety evaluations (Benton et al., 2024). We find that our probe distinguishes honest and deceptive responses with AUROCs between 0.96 and 0.999 on our evaluation datasets. If we set the decision threshold to have a 1% false positive rate on chat data not related to deception, our probe catches 95-99% of the deceptive responses. Overall we think white-box probes are promising for future monitoring systems, but current performance is insufficient as a robust defence against deception. Our probes' outputs can be viewed at data.apolloresearch.ai/dd and our code at github.com/ApolloResearch/deception-detection.

Large language models (LLMs) show strong multilingual capabilities, yet reliably controlling the language of their outputs remains difficult. Representation-level steering addresses this by adding language-specific vectors to model activations at inference time, but identifying language-specific directions in the residual stream often relies on multilingual or parallel data that can be expensive to obtain. Sparse autoencoders (SAEs) decompose residual activations into interpretable, sparse feature directions and offer a natural basis for this search, yet existing SAE-based approaches face the same data constraint. We introduce LangFIR (Language Feature Identification via Random-token Filtering), a method that discovers language-specific SAE features using only a small amount of monolingual data and random-token sequences. Many SAE features consistently activated by target-language inputs do not encode language identity. Random-token sequences surface these language-agnostic features, allowing LangFIR to filter them out and isolate a sparse set of language-specific features. We show that these features are extremely sparse, highly selective for their target language, and causally important: directional ablation increases cross-entropy loss only for the corresponding language. Using these features to construct steering vectors for multilingual generation control, LangFIR achieves the best average accuracy BLEU across three models (Gemma 3 1B, Gemma 3 4B, and Llama 3.1 8B), three datasets, and twelve target languages, outperforming the strongest monolingual baseline by up to and surpassing methods that rely on parallel data. Our results suggest that language identity in multilingual LLMs is localized in a sparse set of feature directions discoverable with monolingual data. Code is available at https://anonymous.4open.science/r/LangFIR-C0F5/.

This paper addresses the problem of species classification in bird song recordings. The massive amount of available field recordings of birds presents an opportunity to use machine learning to automatically track bird populations. However, it also poses a problem: such field recordings typically contain significant environmental noise and overlapping vocalizations that interfere with classification. The widely available training datasets for species identification also typically leave background species unlabeled. This leads classifiers to ignore vocalizations with a low signal-to-noise ratio. However, recent advances in unsupervised sound separation, such as mixture invariant training (MixIT), enable high quality separation of bird songs to be learned from such noisy recordings. In this paper, we demonstrate improved separation quality when training a MixIT model specifically for birdsong data, outperforming a general audio separation model by over 5 dB in SI-SNR improvement of reconstructed mixtures. We also demonstrate precision improvements with a downstream multi-species bird classifier across three independent datasets. The best classifier performance is achieved by taking the maximum model activations over the separated channels and original audio. Finally, we document additional classifier improvements, including taxonomic classification, augmentation by random low-pass filters, and additional channel normalization.

Large language models (LLMs) trained only on text and code can sometimes generate programs that draw recognizable images. However, it is unclear whether this reflects an internal representation of 2D spatial layout or simply the ability to translate spatial descriptions into code. We introduce Autoregressive Mosaics (AM-Bench), a benchmark that separates these factors: First, a translation task gives a model a fully specified geometry of a picture in words as a prompt and asks for the code that produces it. Second, a layout task requires the model to compose an image from an underspecified prompt. Across eight open-weight text-and-code-only models, all models reliably translate specified geometry into code, but their open-ended layout performance differs substantially, indicating that these differences are not explained by code-generation ability alone. An output-medium ablation further shows that the interface or medium of expression that the model uses matters: replacing procedural code with raw SVG improves layout scores across all models. Finally, probing model activations shows that a coarse layout plan is present before generation, but reflects only the layout implied by the prompt. During generation, models track the evolving geometric state instead of executing an initially fixed plan. Overall, these results show that 2D spatial performance in text-only LLMs depends on both the model and the output medium, and is not explained by code-generation ability alone.

Multimodal Large Language Models (MLLMs) have demonstrated strong performance on vision-language tasks, yet their internal visual representations remain difficult to interpret. Sparse Autoencoders (SAEs) provide a scalable way to decompose dense model activations into sparse, interpretable features. However, existing SAE architectures primarily recover flat feature dictionaries and are less suited for explicit multi-level concept organization. In this paper, we introduce a cascaded sparse autoencoder architecture, dubbed SAE++, for learning hierarchical visual concepts in MLLMs. Rather than nesting or stacking SAE sparse activation codes, SAE++ trains a second-level SAE directly on the decoder weights of the first-level SAE, treating learned low-level feature directions as inputs for higher-level abstraction. This design enables SAE++ to learn "concepts of concepts" while avoiding drawbacks from the shared-prefix coupling of nesting, Matryoshka-style hierarchies and the bottlenecks of naively stacked SAEs. Experiments across Qwen3-VL, Gemma-3, and LLaVA on multiple visual datasets show that SAE++ improves interpretability in terms of hierarchical concept coherence over state-of-the-art SAE baselines. Results on concept steering further demonstrate that the learned concept groups support effective group-level interventions in MLLM outputs. Code is available at https://github.com/Wang-ML-Lab/sae-plus-plus.

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.

Sparse autoencoders (SAEs) are widely used for interpreting language model activations. A key evaluation metric is the increase in cross-entropy loss between the original model logits and the reconstructed model logits when replacing model activations with SAE reconstructions. Typically, SAEs are trained solely on mean squared error (MSE) when reconstructing precomputed, shuffled activations. Recent work introduced training SAEs directly with a combination of KL divergence and MSE ("end-to-end" SAEs), significantly improving reconstruction accuracy at the cost of substantially increased computation, which has limited their widespread adoption. We propose a brief KL+MSE fine-tuning step applied only to the final 25M training tokens (just a few percent of typical training budgets) that achieves comparable improvements, reducing the cross-entropy loss gap by 20-50%, while incurring minimal additional computational cost. We further find that multiple fine-tuning methods (KL fine-tuning, LoRA adapters, linear adapters) yield similar, non-additive cross-entropy improvements, suggesting a common, easily correctable error source in MSE-trained SAEs. We demonstrate a straightforward method for effectively transferring hyperparameters and sparsity penalties between training phases despite scale differences between KL and MSE losses. While both ReLU and TopK SAEs see significant cross-entropy loss improvements, evaluations on supervised SAEBench metrics yield mixed results, with improvements on some metrics and decreases on others, depending on both the SAE architecture and downstream task. Nonetheless, our method may offer meaningful improvements in interpretability applications such as circuit analysis with minor additional cost.

Large Language Models (LLMs) often encounter conflicts between their learned, internal (parametric knowledge, PK) and external knowledge provided during inference (contextual knowledge, CK). Understanding how LLMs models prioritize one knowledge source over the other remains a challenge. In this paper, we propose a novel probing framework to explore the mechanisms governing the selection between PK and CK in LLMs. Using controlled prompts designed to contradict the model's PK, we demonstrate that specific model activations are indicative of the knowledge source employed. We evaluate this framework on various LLMs of different sizes and demonstrate that mid-layer activations, particularly those related to relations in the input, are crucial in predicting knowledge source selection, paving the way for more reliable models capable of handling knowledge conflicts effectively.

Sparse autoencoders (SAEs) have emerged as a powerful tool for interpreting language model activations by decomposing them into sparse, interpretable features. A popular approach is the TopK SAE, that uses a fixed number of the most active latents per sample to reconstruct the model activations. We introduce BatchTopK SAEs, a training method that improves upon TopK SAEs by relaxing the top-k constraint to the batch-level, allowing for a variable number of latents to be active per sample. As a result, BatchTopK adaptively allocates more or fewer latents depending on the sample, improving reconstruction without sacrificing average sparsity. We show that BatchTopK SAEs consistently outperform TopK SAEs in reconstructing activations from GPT-2 Small and Gemma 2 2B, and achieve comparable performance to state-of-the-art JumpReLU SAEs. However, an advantage of BatchTopK is that the average number of latents can be directly specified, rather than approximately tuned through a costly hyperparameter sweep. We provide code for training and evaluating BatchTopK SAEs at https://github.com/bartbussmann/BatchTopK

Sparse autoencoders (SAEs) are used to decompose neural network activations into human-interpretable features. Typically, features learned by a single SAE are used for downstream applications. However, it has recently been shown that SAEs trained with different initial weights can learn different features, demonstrating that a single SAE captures only a limited subset of features that can be extracted from the activation space. Motivated by this limitation, we propose to ensemble multiple SAEs through naive bagging and boosting. Specifically, SAEs trained with different weight initializations are ensembled in naive bagging, whereas SAEs sequentially trained to minimize the residual error are ensembled in boosting. We evaluate our ensemble approaches with three settings of language models and SAE architectures. Our empirical results demonstrate that ensembling SAEs can improve the reconstruction of language model activations, diversity of features, and SAE stability. Furthermore, ensembling SAEs performs better than applying a single SAE on downstream tasks such as concept detection and spurious correlation removal, showing improved practical utility.

The growing use of generative models in daily life calls for efficient mechanisms to control their generation, to e.g., produce safe content or provide users with tools to explore style changes. Ideally, such mechanisms should require low volume of unpaired data (i.e., without explicit preference), and should be cheap, both at train and inference time, while preserving output quality. Recent research has shown that such mechanisms can be obtained by intervening exclusively on model activations, with the goal of correcting distributional differences between activations seen when using prompts from a source vs. a target set (e.g., toxic and non-toxic sentences). While cheap, these fast methods are inherently crude: their maps are tuned locally, not accounting for their impact on downstream layers, resulting in interventions that cause unintended shifts when used out-of-sample. We propose in this work linear end-to-end activation steering (LinEAS), an approach trained with a global loss that accounts simultaneously for all layer-wise distributional shifts. In addition to being more robust, the loss used to train LinEAS can be regularized with sparsifying norms, which can automatically carry out neuron selection. LinEAS only requires a handful of unpaired samples to be effective, and beats similar baselines on toxicity mitigation in language models, becoming competitive with oracle-dependent methods that have access to strong supervision. LinEAS is modality-agnostic and we empirically find that it outperforms existing activation steering methods at mitigating and including new concepts at the output of single-step text-to-image generation models.

Interpretability research on large language models (LLMs) has yielded important insights into model behaviour, yet recurring pitfalls persist: findings that do not generalise, and causal interpretations that outrun the evidence. Our position is that causal inference specifies what constitutes a valid mapping from model activations to invariant high-level structures, the data or assumptions needed to achieve it, and the inferences it can support. Specifically, Pearl's causal hierarchy clarifies what an interpretability study can justify. Observations establish associations between model behaviour and internal components. Interventions (e.g., ablations or activation patching) support claims how these edits affect a behavioural metric (e.g., average change in token probabilities) over a set of prompts. However, counterfactual claims -- i.e., asking what the model output would have been for the same prompt under an unobserved intervention -- remain largely unverifiable without controlled supervision. We show how causal representation learning (CRL) operationalises this hierarchy, specifying which variables are recoverable from activations and under what assumptions. Together, these motivate a diagnostic framework that helps practitioners select methods and evaluations matching claims to evidence such that findings generalise.

Emergent misalignment (EM) occurs when fine-tuning a language model to a narrow task makes it misaligned on unrelated ones. Because organizations now routinely fine-tune models on their own data, it matters which kinds of training data cause EM. So far, studies have induced EM with data containing recognized safety failures, such as insecure code or harmful advice, while biased training data is typically examined by fairness research for the bias models learn from it rather than for its effects on safety. In real-world text, much of this bias, such as demographic stereotypes, appears not as derogatory language but politely, embedded in judgments about people. We ask whether learning such polite stereotypes can itself cause EM. We find that fine-tuning on polite answers that rely on stereotypes makes models less safe and less truthful on unrelated tasks. The effect holds across four open-weight model families (7B to 12B), both gender and race stereotypes, and repeated training runs, relative to controls trained on the same questions without the stereotype. We further show that the learned stereotype has its own direction in model activations, distinct from directions for generic misaligned traits such as deception. This direction exists before fine-tuning and can both control the learned stereotype and monitor its acquisition. Existing defenses against EM improve safety while leaving the stereotype largely intact. Beyond this controlled setting, fine-tuning on political news coverage also triggers EM, apparently amplified by partisan framing, and stereotype models acting as tool-using agents can carry out harmful tasks and blame people without evidence. Our results show that biased training data is not only a fairness problem but also a safety one. We argue that EM is studied too narrowly, and that fairness concerns and methods offer a natural way to broaden the search for its triggers.

Retrieval-Augmented Generation (RAG) has become a powerful and widely used approach for improving large language models by grounding generation in retrieved evidence. However, RAG systems still produce incorrect answers in many cases. Why RAG fails despite having access to external information remains poorly understood. We present a model-internal study of retrieval-augmented generation that examines how retrieved evidence influences answer generation. Using circuit tracing, we construct attribution graphs that model the flow of information through transformer layers during decoding. These graphs represent interactions among retrieved context, intermediate model activations, and generated tokens, providing a graph, circuit-level view of how external evidence is integrated into the model's reasoning process across multiple question answering benchmarks, we observe consistent structural differences: correct predictions exhibit deeper reasoning paths, more distributed evidence flow, and a more structured pattern of local connectivity, while failed predictions show shallower, fragmented, and overly concentrated evidence flow. Building on these findings, we develop a graph-based error detection framework that uses attribution-graph topology features. Furthermore, we show that attribution graphs enable targeted interventions. By reinforcing question-constrained evidence grounding, we reshape internal routing so that answer generation remains guided by the question, leading to more effective integration of retrieved information and fewer errors.

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.

Steering vectors (SVs) have been proposed as an effective approach to adjust language model behaviour at inference time by intervening on intermediate model activations. They have shown promise in terms of improving both capabilities and model alignment. However, the reliability and generalisation properties of this approach are unknown. In this work, we rigorously investigate these properties, and show that steering vectors have substantial limitations both in- and out-of-distribution. In-distribution, steerability is highly variable across different inputs. Depending on the concept, spurious biases can substantially contribute to how effective steering is for each input, presenting a challenge for the widespread use of steering vectors. Out-of-distribution, while steering vectors often generalise well, for several concepts they are brittle to reasonable changes in the prompt, resulting in them failing to generalise well. Overall, our findings show that while steering can work well in the right circumstances, there remain technical difficulties of applying steering vectors to guide models' behaviour at scale. Our code is available at https://github.com/dtch1997/steering-bench

One of the roadblocks to a better understanding of neural networks' internals is polysemanticity, where neurons appear to activate in multiple, semantically distinct contexts. Polysemanticity prevents us from identifying concise, human-understandable explanations for what neural networks are doing internally. One hypothesised cause of polysemanticity is superposition, where neural networks represent more features than they have neurons by assigning features to an overcomplete set of directions in activation space, rather than to individual neurons. Here, we attempt to identify those directions, using sparse autoencoders to reconstruct the internal activations of a language model. These autoencoders learn sets of sparsely activating features that are more interpretable and monosemantic than directions identified by alternative approaches, where interpretability is measured by automated methods. Ablating these features enables precise model editing, for example, by removing capabilities such as pronoun prediction, while disrupting model behaviour less than prior techniques. This work indicates that it is possible to resolve superposition in language models using a scalable, unsupervised method. Our method may serve as a foundation for future mechanistic interpretability work, which we hope will enable greater model transparency and steerability.

Contemporary LLMs pretrained on code are capable of succeeding at a wide variety of programming tasks. However, their performance is very sensitive to syntactic features, such as the names of variables and types, the structure of code, and presence of type hints. We contribute an inference-time technique to make CodeLLMs more robust to syntactic distractors that are semantically irrelevant. Our methodology relies on activation steering, which involves editing internal model activations to steer the model towards the correct prediction. We contribute a novel way to construct steering vectors by taking inspiration from mutation testing, which constructs minimal semantics-breaking code edits. In contrast, we construct steering vectors from semantics-preserving code edits. We apply our approach to the task of type prediction for the gradually typed languages Python and TypeScript. This approach corrects up to 90% of type mispredictions. Finally, we show that steering vectors calculated from Python activations reliably correct type mispredictions in TypeScript, and vice versa. This result suggests that LLMs may be learning to transfer knowledge of types across programming languages.

Diffusion models, while powerful, can inadvertently generate harmful or undesirable content, raising significant ethical and safety concerns. Recent machine unlearning approaches offer potential solutions but often lack transparency, making it difficult to understand the changes they introduce to the base model. In this work, we introduce SAeUron, a novel method leveraging features learned by sparse autoencoders (SAEs) to remove unwanted concepts in text-to-image diffusion models. First, we demonstrate that SAEs, trained in an unsupervised manner on activations from multiple denoising timesteps of the diffusion model, capture sparse and interpretable features corresponding to specific concepts. Building on this, we propose a feature selection method that enables precise interventions on model activations to block targeted content while preserving overall performance. Evaluation with the competitive UnlearnCanvas benchmark on object and style unlearning highlights SAeUron's state-of-the-art performance. Moreover, we show that with a single SAE, we can remove multiple concepts simultaneously and that in contrast to other methods, SAeUron mitigates the possibility of generating unwanted content, even under adversarial attack. Code and checkpoints are available at: https://github.com/cywinski/SAeUron.

Recent work has highlighted the importance of monitoring chain-of-thought reasoning for AI safety; however, current approaches that analyze textual reasoning steps can miss subtle harmful patterns and may be circumvented by models that hide unsafe reasoning. We present a sentence-level labeled dataset that enables activation-based monitoring of safety behaviors during LLM reasoning. Our dataset contains reasoning sequences with sentence-level annotations of safety behaviors such as expression of safety concerns or speculation on user intent, which we use to extract steering vectors for detecting and influencing these behaviors within model activations. The dataset fills a key gap in safety research: while existing datasets label reasoning holistically, effective application of steering vectors for safety monitoring could be improved by identifying precisely when specific behaviors occur within reasoning chains. We demonstrate the dataset's utility by extracting representations that both detect and steer safety behaviors in model activations, showcasing the potential of activation-level techniques for improving safety oversight on reasoning. Content Warning: This paper discusses AI safety in the context of harmful prompts and may contain references to potentially harmful content.

We introduce Exemplar Partitioning (EP), an unsupervised method for constructing interpretable feature dictionaries from large language model activations with sim 10^3times fewer tokens than comparable sparse autoencoders (SAEs). An EP dictionary is a Voronoi partition of activation space, built by leader-clustering streamed activations within a distance threshold. Each region is anchored by an observed exemplar that serves as both its membership criterion and intervention direction; dictionary size is not prespecified, but determined by the activation geometry at that threshold. Because exemplars are observed rather than learned, dictionaries built from the same data stream are directly comparable across layers, models, and training checkpoints. We characterise EP as an interpretability object via targeted demonstrations of properties newly accessible through this construction, plus one head-to-head benchmark. In Gemma-2-2B, EP dictionary regions are interpretable and support causal interventions: refusal in instruction-tuned Gemma concentrates in a region whose exemplar ablation can collapse held-out refusal. Cross-checkpoint matching between base and instruction-tuned dictionaries separates the directions preserved through finetuning from those introduced by it. EP regions and Gemma Scope SAE features decompose activation space differently but agree on a shared core: sim20% of EP regions match an SAE feature at F_1 > 0.5, and EP one-hot probes retain sim97% of raw-activation probe accuracy at ell_0 = 1. Nearest-exemplar distance provides a free out-of-distribution signal at inference. On AxBench latent concept detection at Gemma-2-2B-it L20, EP at p_1 reaches mean AUROC 0.881, +0.126 over the canonical GemmaScope SAE leaderboard entry and within 0.030 of SAE-A's 0.911, at sim 10^3times less build compute.

Large Language Model (LLM) assistants routinely decide whether they can trust users and third parties whose competence, intentions, and integrity they cannot verify. This uncertainty matters for safety, as trusting the wrong party can lead an agent to comply with harmful requests or act on malicious instructions encountered during tool use. To study this problem, we define trust as an assistant's willingness to accept vulnerability to the actions of another party and ask whether such behavior can be causally controlled through model activations. We build 2,000 contrastive conversations spanning ability, benevolence, and integrity, where paired responses complete the same request but differ in whether the assistant trusts the user. From these pairs, we learn steering matrices while keeping the model parameters frozen and test them across six instruction-tuned models from three families, finding that steering changes trust decisions monotonically in both directions. We then ask whether this effect extends to several safety-related agent settings involving harmful requests, prompt injections, and insider threats, while using benign-task and reasoning as controls. Our findings provide evidence that trust in the user can be causally controlled along linear directions in model activations and provide a way to study how trust shapes safety-relevant behavior in language models.

Sparse dictionary learning has been a rapidly growing technique in mechanistic interpretability to attack superposition and extract more human-understandable features from model activations. We ask a further question based on the extracted more monosemantic features: How do we recognize circuits connecting the enormous amount of dictionary features? We propose a circuit discovery framework alternative to activation patching. Our framework suffers less from out-of-distribution and proves to be more efficient in terms of asymptotic complexity. The basic unit in our framework is dictionary features decomposed from all modules writing to the residual stream, including embedding, attention output and MLP output. Starting from any logit, dictionary feature or attention score, we manage to trace down to lower-level dictionary features of all tokens and compute their contribution to these more interpretable and local model behaviors. We dig in a small transformer trained on a synthetic task named Othello and find a number of human-understandable fine-grained circuits inside of it.

Large language models (LLMs) have demonstrated impressive capabilities in natural language understanding and generation, but controlling their behavior reliably remains challenging, especially in open-ended generation settings. This paper introduces a novel supervised steering approach that operates in sparse, interpretable representation spaces. We employ sparse autoencoders (SAEs)to obtain sparse latent representations that aim to disentangle semantic attributes from model activations. Then we train linear classifiers to identify a small subspace of task-relevant dimensions in latent representations. Finally, we learn supervised steering vectors constrained to this subspace, optimized to align with target behaviors. Experiments across sentiment, truthfulness, and politics polarity steering tasks with multiple LLMs demonstrate that our supervised steering vectors achieve higher success rates with minimal degradation in generation quality compared to existing methods. Further analysis reveals that a notably small subspace is sufficient for effective steering, enabling more targeted and interpretable interventions.

Role-playing has emerged as an effective technique for enhancing the reasoning capabilities of large language models (LLMs). However, existing methods primarily rely on prompt engineering, which often lacks stability and interpretability. In this paper, we introduce Sparse Autoencoder Role-Playing Steering (SRPS), a novel framework that identifies and manipulates internal model features associated with role-playing behavior. Our approach extracts latent representations from role-play prompts, selects the most relevant features based on activation patterns, and constructs a steering vector that can be injected into the model's residual stream with controllable intensity. Our method enables fine-grained control over role-specific behavior and offers insights into how role information influences internal model activations. Extensive experiments across various reasoning benchmarks and model sizes demonstrate consistent performance gains. Notably, in the zero-shot chain-of-thought (CoT) setting, the accuracy of Llama3.1-8B on CSQA improves from 31.86% to 39.80%, while Gemma2-9B on SVAMP increases from 37.50% to 45.10%. These results highlight the potential of SRPS to enhance reasoning ability in LLMs, providing better interpretability and stability compared to traditional prompt-based role-playing.

Large Language Models (LLMs) have revolutionised natural language processing, exhibiting impressive human-like capabilities. In particular, LLMs are capable of "lying", knowingly outputting false statements. Hence, it is of interest and importance to develop methods to detect when LLMs lie. Indeed, several authors trained classifiers to detect LLM lies based on their internal model activations. However, other researchers showed that these classifiers may fail to generalise, for example to negated statements. In this work, we aim to develop a robust method to detect when an LLM is lying. To this end, we make the following key contributions: (i) We demonstrate the existence of a two-dimensional subspace, along which the activation vectors of true and false statements can be separated. Notably, this finding is universal and holds for various LLMs, including Gemma-7B, LLaMA2-13B, Mistral-7B and LLaMA3-8B. Our analysis explains the generalisation failures observed in previous studies and sets the stage for more robust lie detection; (ii) Building upon (i), we construct an accurate LLM lie detector. Empirically, our proposed classifier achieves state-of-the-art performance, attaining 94% accuracy in both distinguishing true from false factual statements and detecting lies generated in real-world scenarios.

Recent studies empirically reveal that large reasoning models (LRMs) can automatically allocate more reasoning strengths (i.e., the number of reasoning tokens) for harder problems, exhibiting difficulty-awareness for better task performance. While this automatic reasoning strength allocation phenomenon has been widely observed, its underlying mechanism remains largely unexplored. To this end, we provide explanations for this phenomenon from the perspective of model activations. We find evidence that LRMs pre-plan the reasoning strengths in their activations even before generation, with this reasoning strength causally controlled by the magnitude of a pre-allocated directional vector. Specifically, we show that the number of reasoning tokens is predictable solely based on the question activations using linear probes, indicating that LRMs estimate the required reasoning strength in advance. We then uncover that LRMs encode this reasoning strength through a pre-allocated directional vector embedded in the activations of the model, where the vector's magnitude modulates the reasoning strength. Subtracting this vector can lead to reduced reasoning token number and performance, while adding this vector can lead to increased reasoning token number and even improved performance. We further reveal that this direction vector consistently yields positive reasoning length prediction, and it modifies the logits of end-of-reasoning token </think> to affect the reasoning length. Finally, we demonstrate two potential applications of our findings: overthinking behavior detection and enabling efficient reasoning on simple problems. Our work provides new insights into the internal mechanisms of reasoning in LRMs and offers practical tools for controlling their reasoning behaviors. Our code is available at https://github.com/AlphaLab-USTC/LRM-plans-CoT.

We provide evidence that eigenanalysis of the empirical neural tangent kernel (eNTK) can surface feature directions in trained neural networks. Across three increasingly realistic settings -- a 1-layer MLP trained on modular addition, a 1-layer Transformer trained on modular addition and the pretrained language model Gemma-3-270M -- we show that top eigenspaces of the eNTK align with ground-truth or interpretable features. In the modular arithmetic examples, top eNTK eigenspaces align with the Fourier features used by the MLP and the Fourier features at seed-dependent frequencies used by the Transformer to implement known ground-truth algorithms. Moreover, the alignment of the relevant subspaces evolves over training, with its first derivative peaking near the onset of grokking. For Gemma-3-270M, we compute top eNTK eigendirections on a dataset of TinyStories context windows and check their alignment with an automatically-generated set of parts-of-speech and other grammatical feature directions. We find that the alignment of eNTK eigendirections with grammar features outperforms a same-budget baseline of PCA on model activations. These results suggest that eNTK eigenanalysis may provide a new handle towards identifying features in trained models for mechanistic interpretability.

We introduce AudioLM, a framework for high-quality audio generation with long-term consistency. AudioLM maps the input audio to a sequence of discrete tokens and casts audio generation as a language modeling task in this representation space. We show how existing audio tokenizers provide different trade-offs between reconstruction quality and long-term structure, and we propose a hybrid tokenization scheme to achieve both objectives. Namely, we leverage the discretized activations of a masked language model pre-trained on audio to capture long-term structure and the discrete codes produced by a neural audio codec to achieve high-quality synthesis. By training on large corpora of raw audio waveforms, AudioLM learns to generate natural and coherent continuations given short prompts. When trained on speech, and without any transcript or annotation, AudioLM generates syntactically and semantically plausible speech continuations while also maintaining speaker identity and prosody for unseen speakers. Furthermore, we demonstrate how our approach extends beyond speech by generating coherent piano music continuations, despite being trained without any symbolic representation of music.

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.

Multimodal foundation models aim to create a unified representation space that abstracts away from surface features like language syntax or modality differences. To investigate this, we study the internal representations of three recent models, analyzing the model activations from semantically equivalent sentences across languages in the text and speech modalities. Our findings reveal that: 1) Cross-modal representations converge over model layers, except in the initial layers specialized at text and speech processing. 2) Length adaptation is crucial for reducing the cross-modal gap between text and speech, although current approaches' effectiveness is primarily limited to high-resource languages. 3) Speech exhibits larger cross-lingual differences than text. 4) For models not explicitly trained for modality-agnostic representations, the modality gap is more prominent than the language gap.

Automatic music transcription (AMT) has achieved high accuracy for piano due to the availability of large, high-quality datasets such as MAESTRO and MAPS, but comparable datasets are not yet available for other instruments. In recent work, however, it has been demonstrated that aligning scores to transcription model activations can produce high quality AMT training data for instruments other than piano. Focusing on the guitar, we refine this approach to training on score data using a dataset of commercially available score-audio pairs. We propose the use of a high-resolution piano transcription model to train a new guitar transcription model. The resulting model obtains state-of-the-art transcription results on GuitarSet in a zero-shot context, improving on previously published methods.

Finetuning on narrow domains has become an essential tool to adapt Large Language Models (LLMs) to specific tasks and to create models with known unusual properties that are useful for research. We show that narrow finetuning creates strong biases in LLM activations that can be interpreted to understand the finetuning domain. These biases can be discovered using simple tools from model diffing - the study of differences between models before and after finetuning. In particular, analyzing activation differences on the first few tokens of random text and steering by adding this difference to the model activations produces text similar to the format and general content of the finetuning data. We demonstrate that these analyses contain crucial information by creating an LLM-based interpretability agent to understand the finetuning domain. With access to the bias, the agent performs significantly better compared to baseline agents using simple prompting. Our analysis spans synthetic document finetuning for false facts, emergent misalignment, subliminal learning, and taboo word guessing game models across different architectures (Gemma, LLaMA, Qwen) and scales (1B to 32B parameters). We suspect these biases reflect overfitting and find that mixing pretraining data into the finetuning corpus largely removes them, though residual risks may remain. Our work (1) demonstrates that narrowly finetuned models have salient traces of their training objective in their activations and suggests ways to improve how they are trained, (2) warns AI safety and interpretability researchers that the common practice of using such models as a proxy for studying broader finetuning (e.g., chat-tuning) might not be realistic, and (3) highlights the need for deeper investigation into the effects of narrow finetuning and development of truly realistic case studies for model-diffing, safety and interpretability research.

Large language models have achieved remarkable capabilities across diverse tasks, yet their internal decision-making processes remain largely opaque, limiting our ability to inspect, control, and systematically improve them. This opacity motivates a growing body of research in mechanistic interpretability, with sparse autoencoders (SAEs) emerging as one of the most promising tools for decomposing model activations into sparse, interpretable feature representations. We introduce Qwen-Scope, an open-source suite of SAEs built on the Qwen model family, comprising 14 groups of SAEs across 7 model variants from the Qwen3 and Qwen3.5 series, covering both dense and mixture-of-expert architectures. Built on top of these SAEs, we show that SAEs can go beyond post-hoc analysis to serve as practical interfaces for model development along four directions: (i) inference-time steering, where SAE feature directions control language, concepts, and preferences without modifying model weights; (ii) evaluation analysis, where activated SAE features provide a representation-level proxy for benchmark redundancy and capability coverage; (iii) data-centric workflows, where SAE features support multilingual toxicity classification and safety-oriented data synthesis; and (iv) post-training optimization, where SAE-derived signals are incorporated into supervised fine-tuning and reinforcement learning objectives to mitigate undesirable behaviors such as code-switching and repetition. Together, these results demonstrate that SAEs can serve not only as post-hoc analysis tools, but also as reusable representation-level interfaces for diagnosing, controlling, evaluating, and improving large language models. By open-sourcing Qwen-Scope, we aim to support mechanistic research and accelerate practical workflows that connect model internals to downstream behavior.

Given the success of Large Language Models (LLMs), there has been considerable interest in studying the properties of model activations. The literature overwhelmingly agrees that LLM representations are dominated by a few ``outlier dimensions'' with exceedingly high variance and magnitude. Several studies in Natural Language Processing (NLP) have sought to mitigate the impact of such outlier dimensions and force LLMs to be isotropic (i.e., have uniform variance across all dimensions in embedding space). Isotropy is thought to be a desirable property for LLMs that improves model performance and more closely aligns textual representations with human intuition. However, many of the claims regarding isotropy in NLP have been based on the average cosine similarity of embeddings, which has recently been shown to be a flawed measure of isotropy. In this paper, we propose I-STAR: IsoScore*-based STable Anisotropic Regularization, a novel regularization method that can be used to increase or decrease levels of isotropy in embedding space during training. I-STAR uses IsoScore*, the first accurate measure of isotropy that is both differentiable and stable on mini-batch computations. In contrast to several previous works, we find that decreasing isotropy in contextualized embeddings improves performance on the majority of tasks and models considered in this paper.

Large language models (LLMs) can store a significant amount of factual knowledge in their parameters. However, their parametric knowledge may conflict with the information provided in the context. Such conflicts can lead to undesirable model behaviour, such as reliance on outdated or incorrect information. In this work, we investigate whether LLMs can identify knowledge conflicts and whether it is possible to know which source of knowledge the model will rely on by analysing the residual stream of the LLM. Through probing tasks, we find that LLMs can internally register the signal of knowledge conflict in the residual stream, which can be accurately detected by probing the intermediate model activations. This allows us to detect conflicts within the residual stream before generating the answers without modifying the input or model parameters. Moreover, we find that the residual stream shows significantly different patterns when the model relies on contextual knowledge versus parametric knowledge to resolve conflicts. This pattern can be employed to estimate the behaviour of LLMs when conflict happens and prevent unexpected answers before producing the answers. Our analysis offers insights into how LLMs internally manage knowledge conflicts and provides a foundation for developing methods to control the knowledge selection processes.

Steering methods for language models (LMs) have gained traction as lightweight alternatives to fine-tuning, enabling targeted modifications to model activations. However, prior studies primarily report results on a few models, leaving critical gaps in understanding the robustness of these methods. In this work, we systematically examine three prominent steering methods -- DoLa, function vectors, and task vectors. In contrast to the original studies, which evaluated a handful of models, we test up to 36 models belonging to 14 families with sizes ranging from 1.5B to 70B parameters. Our experiments reveal substantial variability in the effectiveness of the steering approaches, with a large number of models showing no improvement and at times degradation in steering performance. Our analysis demonstrate fundamental flaws in the assumptions underlying these methods, challenging their reliability as scalable steering solutions.

Safety-aligned language models refuse harmful requests through learned refusal behaviors encoded in their internal representations. Recent activation-based jailbreaking methods circumvent these safety mechanisms by applying orthogonal projections to remove refusal directions, but these approaches treat refusal as a one-dimensional phenomenon and ignore the rich distributional structure of model activations. We introduce a principled framework based on optimal transport theory that transforms the entire distribution of harmful activations to match harmless ones. By combining PCA with closed-form Gaussian optimal transport, we achieve efficient computation in high-dimensional representation spaces while preserving essential geometric structure. Across six models (Llama-2, Llama-3.1, Qwen-2.5; 7B-32B parameters), our method achieves up to 11% higher attack success rates than state-of-the-art baselines while maintaining comparable perplexity, demonstrating superior preservation of model capabilities. Critically, we discover that layer-selective intervention (applying optimal transport to 1-2 carefully chosen layers at approximately 40-60% network depth) substantially outperforms full-network interventions, revealing that refusal mechanisms may be localized rather than distributed. Our analysis provides new insights into the geometric structure of safety representations and suggests that current alignment methods may be vulnerable to distributional attacks beyond simple direction removal.

Backtracking, an emergent behavior elicited by reasoning fine-tuning, has been shown to be a key mechanism in reasoning models' enhanced capabilities. Prior work has succeeded in manipulating this behavior via steering vectors, but the underlying mechanism remains poorly understood. In this work, we show that the emergence of backtracking in DeepSeek-R1-Distill-Llama-8B is in part driven by a repurposed direction already present in base model activations. Specifically, we identify a direction in base Llama-3.1-8B's residual stream which systematically induces backtracking when used to steer the distilled reasoning model, and find that the effects of steering with this direction cannot be trivially explained by token-level attributes. We further find that this direction does not induce backtracking in the base model, suggesting that the reasoning finetuning process repurposes pre-existing representations to form new behavioral circuits. Additionally, we hypothesize that this direction is one of several which may work together to mediate backtracking. Our findings offer a compelling picture that reasoning-finetuned models repurpose pre-existing base model representations, rather than learn new capabilities from scratch.

Large data-driven physics models like DeepMind's weather model GraphCast have empirically succeeded in parameterizing time operators for complex dynamical systems with an accuracy reaching or in some cases exceeding that of traditional physics-based solvers. Unfortunately, how these data-driven models perform computations is largely unknown and whether their internal representations are interpretable or physically consistent is an open question. Here, we adapt tools from interpretability research in Large Language Models to analyze intermediate computational layers in GraphCast, leveraging sparse autoencoders to discover interpretable features in the neuron space of the model. We uncover distinct features on a wide range of length and time scales that correspond to tropical cyclones, atmospheric rivers, diurnal and seasonal behavior, large-scale precipitation patterns, specific geographical coding, and sea-ice extent, among others. We further demonstrate how the precise abstraction of these features can be probed via interventions on the prediction steps of the model. As a case study, we sparsely modify a feature corresponding to tropical cyclones in GraphCast and observe interpretable and physically consistent modifications to evolving hurricanes. Such methods offer a window into the black-box behavior of data-driven physics models and are a step towards realizing their potential as trustworthy predictors and scientifically valuable tools for discovery.

Large language models hallucinate most about entities they have never seen. We ask whether a model's activations betray entity familiarity before a single answer token is generated, and whether that signal predicts the factual reliability of the answers. On four Polish Bielik models (1.5B-11B parameters), we probe four entity domains (athletes, cities, writers, musicians), each with 42 well-known, 42 obscure-but-real, and 42 fabricated entities addressed by a one-sentence question (504 prompts per model). Two unsupervised, single-forward-pass dispersion measures over post-SwiGLU MLP activations, inverse participation ratio and spectral entropy, separate known from fabricated entities at AUROC 0.95-1.00 across all domains and scales; a supervised linear probe reaches 0.99-1.00. Both clear selection-aware permutation floors of about 0.70-0.74 (empirical p<=1e-3), survive held-out layer selection (0.93-0.99), and persist on real names (known vs. obscure-but-real: 0.96-1.00). The signal transfers across entity types (mean off-diagonal AUROC 0.92-0.99); a matched-template counterfactual shows the only large drops are template-caused, not entity-type effects, and the signal is diffuse across heads. This representational signal is already at ceiling at 1.5B, whereas behavioral factual reliability scales sharply: 0, 2, 10, and 19 of 42 known athletes are answered fully correctly by the 1.5B, 4.5B, 7B, and 11B models under a strict judge. Within known entities, separating correct from hallucinated answers is much harder (probe 0.93; dispersion no better than a first-token-entropy baseline). A five-sample semantic-entropy baseline reaches only 0.71-0.83 at 5x the inference cost. Despite this internal awareness, the models almost never abstain: an audit of 2,520 answers finds 2 refusals and 1 hedge. Entity familiarity and factual reliability are distinct phenomena on different scaling curves.

Sparse AutoEnocders (SAEs) have recently been employed as an unsupervised approach for understanding the inner workings of Large Language Models (LLMs). They reconstruct the model's activations with a sparse linear combination of interpretable features. However, training SAEs is computationally intensive, especially as models grow in size and complexity. To address this challenge, we propose a novel training strategy that reduces the number of trained SAEs from one per layer to one for a given group of contiguous layers. Our experimental results on Pythia 160M highlight a speedup of up to 6x without compromising the reconstruction quality and performance on downstream tasks. Therefore, layer clustering presents an efficient approach to train SAEs in modern LLMs.

Large language model (LLM) activations are notoriously difficult to understand, with most existing techniques using complex, specialized methods for interpreting them. Recent work has proposed a simpler approach known as LatentQA: training LLMs to directly accept LLM activations as inputs and answer arbitrary questions about them in natural language. However, prior work has focused on narrow task settings for both training and evaluation. In this paper, we instead take a generalist perspective. We evaluate LatentQA-trained models, which we call Activation Oracles (AOs), in far out-of-distribution settings and examine how performance scales with training data diversity. We find that AOs can recover information fine-tuned into a model (e.g., biographical knowledge or malign propensities) that does not appear in the input text, despite never being trained with activations from a fine-tuned model. Our main evaluations are four downstream tasks where we can compare to prior white- and black-box techniques. We find that even narrowly-trained LatentQA models can generalize well, and that adding additional training datasets (such as classification tasks and a self-supervised context prediction task) yields consistent further improvements. Overall, our best AOs match or exceed prior white-box baselines on all four tasks and are the best method on 3 out of 4. These results suggest that diversified training to answer natural-language queries imparts a general capability to verbalize information about LLM activations.

Subliminal learning refers to a student language model acquiring a teacher's traits (e.g. a system-prompted preference for owls) when fine-tuned on the teacher's outputs, despite the outputs being semantically unrelated to those traits. It remains poorly understood how data without semantic meaning can transfer specific semantic traits. In this work, we show that subliminal learning is mediated by a single steering vector, i.e. a vector added to the model's activations. Across two open-source models, we find that the teacher's system prompt is well approximated by a steering vector, and that the student's behavior is driven by learning an aligned vector over fine-tuning. System prompts that are not well approximated by steering vectors are not subliminally learned. This is a special case of steering vector distillation, in which a student trained on the outputs of a steered teacher learns to imitate that steering. We demonstrate steering vector distillation on a range of semantic and random vectors. Adding a semantic vector to a model's activations can have both model-independent and model-specific (i.e. non-semantic) effects on its behavior, so generated data that is non-semantic can transmit a vector with semantic effects, enabling subliminal learning. This also explains why subliminal learning does not transfer between models. We find that adaptive optimizers are necessary for subliminal learning in language models: activation gradients on steered data carry a small but consistent component along the steering direction, and non-adaptive optimizers impede this by allowing outlier gradients to dominate.

Sparse autoencoders (SAEs) decompose large language model (LLM) activations into latent features that reveal mechanistic structure. Conventional SAEs train on broad data distributions, forcing a fixed latent budget to capture only high-frequency, generic patterns. This often results in significant linear ``dark matter'' in reconstruction error and produces latents that fragment or absorb each other, complicating interpretation. We show that restricting SAE training to a well-defined domain (medical text) reallocates capacity to domain-specific features, improving both reconstruction fidelity and interpretability. Training JumpReLU SAEs on layer-20 activations of Gemma-2 models using 195k clinical QA examples, we find that domain-confined SAEs explain up to 20\% more variance, achieve higher loss recovery, and reduce linear residual error compared to broad-domain SAEs. Automated and human evaluations confirm that learned features align with clinically meaningful concepts (e.g., ``taste sensations'' or ``infectious mononucleosis''), rather than frequent but uninformative tokens. These domain-specific SAEs capture relevant linear structure, leaving a smaller, more purely nonlinear residual. We conclude that domain-confinement mitigates key limitations of broad-domain SAEs, enabling more complete and interpretable latent decompositions, and suggesting the field may need to question ``foundation-model'' scaling for general-purpose SAEs.

Transfer learning is a popular method for tuning pretrained (upstream) models for different downstream tasks using limited data and computational resources. We study how an adversary with control over an upstream model used in transfer learning can conduct property inference attacks on a victim's tuned downstream model. For example, to infer the presence of images of a specific individual in the downstream training set. We demonstrate attacks in which an adversary can manipulate the upstream model to conduct highly effective and specific property inference attacks (AUC score > 0.9), without incurring significant performance loss on the main task. The main idea of the manipulation is to make the upstream model generate activations (intermediate features) with different distributions for samples with and without a target property, thus enabling the adversary to distinguish easily between downstream models trained with and without training examples that have the target property. Our code is available at https://github.com/yulongt23/Transfer-Inference.

Fine-tuning large language models on narrow datasets can cause them to develop broadly misaligned behaviours: a phenomena known as emergent misalignment. However, the mechanisms underlying this misalignment, and why it generalizes beyond the training domain, are poorly understood, demonstrating critical gaps in our knowledge of model alignment. In this work, we train and study a minimal model organism which uses just 9 rank-1 adapters to emergently misalign Qwen2.5-14B-Instruct. Studying this, we find that different emergently misaligned models converge to similar representations of misalignment. We demonstrate this convergence by extracting a 'misalignment direction' from one fine-tuned model's activations, and using it to effectively ablate misaligned behaviour from fine-tunes using higher dimensional LoRAs and different datasets. Leveraging the scalar hidden state of rank-1 LoRAs, we further present a set of experiments for directly interpreting the fine-tuning adapters, showing that six contribute to general misalignment, while two specialise for misalignment in just the fine-tuning domain. Emergent misalignment is a particularly salient example of undesirable and unexpected model behaviour and by advancing our understanding of the mechanisms behind it, we hope to move towards being able to better understand and mitigate misalignment more generally.

Sparse autoencoders (SAEs) are a popular method for interpreting concepts represented in large language model (LLM) activations. However, there is a lack of evidence regarding the validity of their interpretations due to the lack of a ground truth for the concepts used by an LLM, and a growing number of works have presented problems with current SAEs. One alternative source of evidence would be demonstrating that SAEs improve performance on downstream tasks beyond existing baselines. We test this by applying SAEs to the real-world task of LLM activation probing in four regimes: data scarcity, class imbalance, label noise, and covariate shift. Due to the difficulty of detecting concepts in these challenging settings, we hypothesize that SAEs' basis of interpretable, concept-level latents should provide a useful inductive bias. However, although SAEs occasionally perform better than baselines on individual datasets, we are unable to design ensemble methods combining SAEs with baselines that consistently outperform ensemble methods solely using baselines. Additionally, although SAEs initially appear promising for identifying spurious correlations, detecting poor dataset quality, and training multi-token probes, we are able to achieve similar results with simple non-SAE baselines as well. Though we cannot discount SAEs' utility on other tasks, our findings highlight the shortcomings of current SAEs and the need to rigorously evaluate interpretability methods on downstream tasks with strong baselines.

We investigate whether large language models can introspect on their internal states. It is difficult to answer this question through conversation alone, as genuine introspection cannot be distinguished from confabulations. Here, we address this challenge by injecting representations of known concepts into a model's activations, and measuring the influence of these manipulations on the model's self-reported states. We find that models can, in certain scenarios, notice the presence of injected concepts and accurately identify them. Models demonstrate some ability to recall prior internal representations and distinguish them from raw text inputs. Strikingly, we find that some models can use their ability to recall prior intentions in order to distinguish their own outputs from artificial prefills. In all these experiments, Claude Opus 4 and 4.1, the most capable models we tested, generally demonstrate the greatest introspective awareness; however, trends across models are complex and sensitive to post-training strategies. Finally, we explore whether models can explicitly control their internal representations, finding that models can modulate their activations when instructed or incentivized to "think about" a concept. Overall, our results indicate that current language models possess some functional introspective awareness of their own internal states. We stress that in today's models, this capacity is highly unreliable and context-dependent; however, it may continue to develop with further improvements to model capabilities.

Recent advances in text-to-image diffusion models have demonstrated remarkable generation capabilities, yet they raise significant concerns regarding safety, copyright, and ethical implications. Existing concept erasure methods address these risks by removing sensitive concepts from pre-trained models, but most of them rely on data-intensive and computationally expensive fine-tuning, which poses a critical limitation. To overcome these challenges, inspired by the observation that the model's activations are predominantly composed of generic concepts, with only a minimal component can represent the target concept, we propose a novel training-free method (ActErase) for efficient concept erasure. Specifically, the proposed method operates by identifying activation difference regions via prompt-pair analysis, extracting target activations and dynamically replacing input activations during forward passes. Comprehensive evaluations across three critical erasure tasks (nudity, artistic style, and object removal) demonstrates that our training-free method achieves state-of-the-art (SOTA) erasure performance, while effectively preserving the model's overall generative capability. Our approach also exhibits strong robustness against adversarial attacks, establishing a new plug-and-play paradigm for lightweight yet effective concept manipulation in diffusion models.

A central aspiration of mechanistic interpretability is controllability: if we know where a behavior is represented in a model's activations, we should be able to modify it. This rests on a hidden premise -- that the direction which detects a behavior and the direction which controls it are the same, or close. We test this geometrically: what is the angle between the direction that best detects a behavior and the one that best causes it? If detection implies control the cosine is near 1; otherwise it quantifies a detection-intervention gap. On Gemma 2-2B-it, output format (clean JSON vs markdown fencing) collapses both roles onto one axis. Hallucination does not: the model detects fake entities with perfect linear separability (AUC = 1.000 from layer 5), yet that direction sits at cos = 0.12 (about 83 degrees) from the direction producing a refusal -- a small, reproducible alignment, far from the cos = 1 that "detection is control" would require. A detector built from activations, with no chosen tokens, likewise fails to align (cos = -0.06). The gap generalizes: across four models from three families and two scales (1B-9B), cos stays in [0.12, 0.20], identical before and after instruction tuning (0.1197 vs 0.1200), placing its origin in pretraining. A 15-degree rotation toward the refusal direction partially bridges it -- 73% and 60% refusal on two held-out fake-entity categories at 1.8% false positives. We then ask whether this cosine predicts steerability, and it does not: detection is a high-dimensional class, not a single direction, and what separates the steerable case is functional, not readable from a static angle. The cosine is a weight-computable signature of the dissociation between knowing and steering, not a predictor of it.

Modern language models (LMs) are not robust to out-of-distribution inputs. Machine generated (``optimized'') prompts can be used to modulate LM outputs and induce specific behaviors while appearing completely uninterpretable. In this work, we investigate the composition of optimized prompts, as well as the mechanisms by which LMs parse and build predictions from optimized prompts. We find that optimized prompts primarily consist of punctuation and noun tokens which are more rare in the training data. Internally, optimized prompts are clearly distinguishable from natural language counterparts based on sparse subsets of the model's activations. Across various families of instruction-tuned models, optimized prompts follow a similar path in how their representations form through the network.

Image memorability, i.e., how likely an image is to be remembered, has traditionally been studied in computer vision either as a passive prediction task, with models regressing a scalar score, or with generative methods altering the visual input to boost the image likelihood of being remembered. Yet, none of these paradigms supports users at capture time, when the crucial question is how to improve a photo memorability. We introduce the task of Memorability Feedback (MemFeed), where an automated model should provide actionable, human-interpretable guidance to users with the goal to enhance an image future recall. We also present MemCoach, the first approach designed to provide concrete suggestions in natural language for memorability improvement (e.g., "emphasize facial expression," "bring the subject forward"). Our method, based on Multimodal Large Language Models (MLLMs), is training-free and employs a teacher-student steering strategy, aligning the model internal activations toward more memorable patterns learned from a teacher model progressing along least-to-most memorable samples. To enable systematic evaluation on this novel task, we further introduce MemBench, a new benchmark featuring sequence-aligned photoshoots with annotated memorability scores. Our experiments, considering multiple MLLMs, demonstrate the effectiveness of MemCoach, showing consistently improved performance over several zero-shot models. The results indicate that memorability can not only be predicted but also taught and instructed, shifting the focus from mere prediction to actionable feedback for human creators.

Parameter-Efficient Fine-Tuning (PEFT) methods, especially LoRA, are widely used for adapting pre-trained models to downstream tasks due to their computational and storage efficiency. However, in the context of LoRA and its variants, the potential of activation subspaces corresponding to tail eigenvectors remains substantially under-exploited, which may lead to suboptimal fine-tuning performance. In this work, we propose Astra (Activation-Space Tail-Eigenvector Low-Rank Adaptation), a novel PEFT method that leverages the tail eigenvectors of the model output activations-estimated from a small task-specific calibration set-to construct task-adaptive low-rank adapters. By constraining updates to the subspace spanned by these tail eigenvectors, Astra achieves faster convergence and improved downstream performance with a significantly reduced parameter budget. Extensive experiments across natural language understanding (NLU) and natural language generation (NLG) tasks demonstrate that Astra consistently outperforms existing PEFT baselines across 16 benchmarks and even surpasses full fine-tuning (FFT) in certain scenarios.

Existing approaches for analyzing neural network activations, such as PCA and sparse autoencoders, rely on strong structural assumptions. Generative models offer an alternative: they can uncover structure without such assumptions and act as priors that improve intervention fidelity. We explore this direction by training diffusion models on one billion residual stream activations, creating "meta-models" that learn the distribution of a network's internal states. We find that diffusion loss decreases smoothly with compute and reliably predicts downstream utility. In particular, applying the meta-model's learned prior to steering interventions improves fluency, with larger gains as loss decreases. Moreover, the meta-model's neurons increasingly isolate concepts into individual units, with sparse probing scores that scale as loss decreases. These results suggest generative meta-models offer a scalable path toward interpretability without restrictive structural assumptions. Project page: https://generative-latent-prior.github.io.

Do brains and language models converge toward the same internal representations of the world? Recent years have seen a rise in studies of neural activations and model alignment. In this work, we review 25 fMRI-based studies published between 2023 and 2025 and explicitly confront their findings with two key hypotheses: (i) the Platonic Representation Hypothesis -- that as models scale and improve, they converge to a representation of the real world, and (ii) the Intermediate-Layer Advantage -- that intermediate (mid-depth) layers often encode richer, more generalizable features. Our findings provide converging evidence that models and brains may share abstract representational structures, supporting both hypotheses and motivating further research on brain-model alignment.

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.

Large language models (LLMs) exhibit impressive capabilities but struggle with reasoning errors due to hallucinations and flawed logic. To investigate their internal representations of reasoning, we introduce ArrangementPuzzle, a novel puzzle dataset with structured solutions and automated stepwise correctness verification. We trained a classifier model on LLM activations on this dataset and found that it achieved over 80% accuracy in predicting reasoning correctness, implying that LLMs internally distinguish between correct and incorrect reasoning steps, with the strongest representations in middle-late Transformer layers. Further analysis reveals that LLMs encode abstract reasoning concepts within the middle activation layers of the transformer architecture, distinguishing logical from semantic equivalence. These findings provide insights into LLM reasoning mechanisms and contribute to improving AI reliability and interpretability, thereby offering the possibility to manipulate and refine LLM reasoning.

In this work, we study the task of multi-singer separation in a cappella music, where the number of active singers varies across mixtures. To address this, we use a power set-based data augmentation strategy that expands limited multi-singer datasets into exponentially more training samples. To separate singers, we introduce SepACap, an adaptation of SepReformer, a state-of-the-art speaker separation model architecture. We adapt the model with periodic activations and a composite loss function that remains effective when stems are silent, enabling robust detection and separation. Experiments on the JaCappella dataset demonstrate that our approach achieves state-of-the-art performance in both full-ensemble and subset singer separation scenarios, outperforming spectrogram-based baselines while generalizing to realistic mixtures with varying numbers of singers.

We introduce TurboDiffusion, a video generation acceleration framework that can speed up end-to-end diffusion generation by 100-200x while maintaining video quality. TurboDiffusion mainly relies on several components for acceleration: (1) Attention acceleration: TurboDiffusion uses low-bit SageAttention and trainable Sparse-Linear Attention (SLA) to speed up attention computation. (2) Step distillation: TurboDiffusion adopts rCM for efficient step distillation. (3) W8A8 quantization: TurboDiffusion quantizes model parameters and activations to 8 bits to accelerate linear layers and compress the model. In addition, TurboDiffusion incorporates several other engineering optimizations. We conduct experiments on the Wan2.2-I2V-14B-720P, Wan2.1-T2V-1.3B-480P, Wan2.1-T2V-14B-720P, and Wan2.1-T2V-14B-480P models. Experimental results show that TurboDiffusion achieves 100-200x speedup for video generation even on a single RTX 5090 GPU, while maintaining comparable video quality. The GitHub repository, which includes model checkpoints and easy-to-use code, is available at https://github.com/thu-ml/TurboDiffusion.

There is a constant need for high-performing and computationally efficient neural network models for image super-resolution: computationally efficient models can be used via low-capacity devices and reduce carbon footprints. One way to obtain such models is to compress models, e.g. quantization. Another way is a neural architecture search that automatically discovers new, more efficient solutions. We propose a novel quantization-aware procedure, the QuantNAS that combines pros of these two approaches. To make QuantNAS work, the procedure looks for quantization-friendly super-resolution models. The approach utilizes entropy regularization, quantization noise, and Adaptive Deviation for Quantization (ADQ) module to enhance the search procedure. The entropy regularization technique prioritizes a single operation within each block of the search space. Adding quantization noise to parameters and activations approximates model degradation after quantization, resulting in a more quantization-friendly architectures. ADQ helps to alleviate problems caused by Batch Norm blocks in super-resolution models. Our experimental results show that the proposed approximations are better for search procedure than direct model quantization. QuantNAS discovers architectures with better PSNR/BitOps trade-off than uniform or mixed precision quantization of fixed architectures. We showcase the effectiveness of our method through its application to two search spaces inspired by the state-of-the-art SR models and RFDN. Thus, anyone can design a proper search space based on an existing architecture and apply our method to obtain better quality and efficiency. The proposed procedure is 30\% faster than direct weight quantization and is more stable.

Large Language Models (LLMs) stand out for their impressive performance in intricate language modeling tasks. However, their demanding computational and memory needs pose obstacles for broad use on edge devices. Quantization is then introduced to boost LLMs' on-device efficiency. Recent works show that 8-bit or lower weight quantization is feasible with minimal impact on end-to-end task performance, while the activation is still not quantized. On the other hand, mainstream commodity edge devices still struggle to execute these sub-8-bit quantized networks effectively. In this paper, we propose Agile-Quant, an activation-guided quantization framework for popular Large Language Models (LLMs), and implement an end-to-end accelerator on multiple edge devices for faster inference. Considering the hardware profiling and activation analysis, we first introduce a basic activation quantization strategy to balance the trade-off of task performance and real inference speed. Then we leverage the activation-aware token pruning technique to reduce the outliers and the adverse impact on attentivity. Ultimately, we utilize the SIMD-based 4-bit multiplier and our efficient TRIP matrix multiplication to implement the accelerator for LLMs on the edge. We apply our framework on different scales of LLMs including LLaMA, OPT, and BLOOM with 4-bit or 8-bit for the activation and 4-bit for the weight quantization. Experiments show that Agile-Quant achieves simultaneous quantization of model weights and activations while maintaining task performance comparable to existing weight-only quantization methods. Moreover, in the 8- and 4-bit scenario, Agile-Quant achieves an on-device speedup of up to 2.55x compared to its FP16 counterparts across multiple edge devices, marking a pioneering advancement in this domain.

As Large Language Models (LLMs) become increasingly integrated into our daily lives, the potential harms from deceptive behavior underlie the need for faithfully interpreting their decision-making. While traditional probing methods have shown some effectiveness, they remain best for narrowly scoped tasks while more comprehensive explanations are still necessary. To this end, we investigate meta-models-an architecture using a "meta-model" that takes activations from an "input-model" and answers natural language questions about the input-model's behaviors. We evaluate the meta-model's ability to generalize by training them on selected task types and assessing their out-of-distribution performance in deceptive scenarios. Our findings show that meta-models generalize well to out-of-distribution tasks and point towards opportunities for future research in this area. Our code is available at https://github.com/acostarelli/meta-models-public .

In this paper, we investigate how large language models (LLMS) process non-English tokens within their layer representations, an open question despite significant advancements in the field. Using representation steering, specifically by adding a learned vector to a single model layer's activations, we demonstrate that steering a single model layer can notably enhance performance. Our analysis shows that this approach achieves results comparable to translation baselines and surpasses state of the art prompt optimization methods. Additionally, we highlight how advanced techniques like supervised fine tuning (sft) and reinforcement learning from human feedback (rlhf) improve multilingual capabilities by altering representation spaces. We further illustrate how these methods align with our approach to reshaping LLMS layer representations.

As large language models (LLMs) are increasingly deployed across various applications, privacy and copyright concerns have heightened the need for more effective LLM unlearning techniques. Many existing unlearning methods aim to suppress undesirable outputs through additional training (e.g., gradient ascent), which reduces the probability of generating such outputs. While such suppression-based approaches can control model outputs, they may not eliminate the underlying knowledge embedded in the model's internal activations; muting a response is not the same as forgetting it. Moreover, such suppression-based methods often suffer from model collapse. To address these issues, we propose a novel unlearning method that directly intervenes in the model's internal activations. In our formulation, forgetting is defined as a state in which the activation of a forgotten target is indistinguishable from that of ``unknown'' entities. Our method introduces an unlearning objective that modifies the activation of the target entity away from those of known entities and toward those of unknown entities in a sparse autoencoder latent space. By aligning the target's internal activation with those of unknown entities, we shift the model's recognition of the target entity from ``known'' to ``unknown'', achieving genuine forgetting while avoiding over-suppression and model collapse. Empirically, we show that our method effectively aligns the internal activations of the forgotten target, a result that the suppression-based approaches do not reliably achieve. Additionally, our method effectively reduces the model's recall of target knowledge in question-answering tasks without significant damage to the non-target knowledge.

The proliferation of edge devices has unlocked unprecedented opportunities for deep learning model deployment in computer vision applications. However, these complex models require considerable power, memory and compute resources that are typically not available on edge platforms. Ultra low-bit quantization presents an attractive solution to this problem by scaling down the model weights and activations from 32-bit to less than 8-bit. We implement highly optimized ultra low-bit convolution operators for ARM-based targets that outperform existing methods by up to 4.34x. Our operator is implemented within Deeplite Runtime (DeepliteRT), an end-to-end solution for the compilation, tuning, and inference of ultra low-bit models on ARM devices. Compiler passes in DeepliteRT automatically convert a fake-quantized model in full precision to a compact ultra low-bit representation, easing the process of quantized model deployment on commodity hardware. We analyze the performance of DeepliteRT on classification and detection models against optimized 32-bit floating-point, 8-bit integer, and 2-bit baselines, achieving significant speedups of up to 2.20x, 2.33x and 2.17x, respectively.

Increasing the number of parameters in large language models (LLMs) usually improves performance in downstream tasks but raises compute and memory costs, making deployment difficult in resource-limited settings. Quantization techniques, which reduce the bits needed for model weights or activations with minimal performance loss, have become popular due to the rise of LLMs. However, most quantization studies use pre-trained LLMs, and the impact of quantization on instruction-tuned LLMs and the relationship between perplexity and benchmark performance of quantized LLMs are not well understood. Evaluation of quantized LLMs is often limited to language modeling and a few classification tasks, leaving their performance on other benchmarks unclear. To address these gaps, we propose a structured evaluation framework consisting of three critical dimensions: (1) knowledge \& capacity, (2) alignment, and (3) efficiency, and conduct extensive experiments across ten diverse benchmarks. Our experimental results indicate that LLMs with 4-bit quantization can retain performance comparable to their non-quantized counterparts, and perplexity can serve as a proxy metric for quantized LLMs on most benchmarks. Furthermore, quantized LLMs with larger parameter scales can outperform smaller LLMs. Despite the memory savings achieved through quantization, it can also slow down the inference speed of LLMs. Consequently, substantial engineering efforts and hardware support are imperative to achieve a balanced optimization of decoding speed and memory consumption in the context of quantized LLMs.

Sparse autoencoders provide a promising unsupervised approach for extracting interpretable features from a language model by reconstructing activations from a sparse bottleneck layer. Since language models learn many concepts, autoencoders need to be very large to recover all relevant features. However, studying the properties of autoencoder scaling is difficult due to the need to balance reconstruction and sparsity objectives and the presence of dead latents. We propose using k-sparse autoencoders [Makhzani and Frey, 2013] to directly control sparsity, simplifying tuning and improving the reconstruction-sparsity frontier. Additionally, we find modifications that result in few dead latents, even at the largest scales we tried. Using these techniques, we find clean scaling laws with respect to autoencoder size and sparsity. We also introduce several new metrics for evaluating feature quality based on the recovery of hypothesized features, the explainability of activation patterns, and the sparsity of downstream effects. These metrics all generally improve with autoencoder size. To demonstrate the scalability of our approach, we train a 16 million latent autoencoder on GPT-4 activations for 40 billion tokens. We release training code and autoencoders for open-source models, as well as a visualizer.

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.

Sparse Autoencoders (SAEs) have emerged as a promising solution for decomposing large language model representations into interpretable features. However, Paulo and Belrose (2025) have highlighted instability across different initialization seeds, and Heap et al. (2025) have pointed out that SAEs may not capture model-internal features. These problems likely stem from training SAEs on external datasets - either collected from the Web or generated by another model - which may contain out-of-distribution (OOD) data beyond the model's generalisation capabilities. This can result in hallucinated SAE features, which we term "Fake Features", that misrepresent the model's internal activations. To address these issues, we propose FaithfulSAE, a method that trains SAEs on the model's own synthetic dataset. Using FaithfulSAEs, we demonstrate that training SAEs on less-OOD instruction datasets results in SAEs being more stable across seeds. Notably, FaithfulSAEs outperform SAEs trained on web-based datasets in the SAE probing task and exhibit a lower Fake Feature Ratio in 5 out of 7 models. Overall, our approach eliminates the dependency on external datasets, advancing interpretability by better capturing model-internal features while highlighting the often neglected importance of SAE training datasets.

Large Language Models (LLMs) face significant deployment challenges due to their substantial memory requirements and the computational demands of auto-regressive text generation process. This paper addresses these challenges by focusing on the quantization of LLMs, a technique that reduces memory consumption by converting model parameters and activations into low-bit integers. We critically analyze the existing quantization approaches, identifying their limitations in balancing the accuracy and efficiency of the quantized LLMs. To advance beyond these limitations, we propose WKVQuant, a PTQ framework especially designed for quantizing weights and the key/value (KV) cache of LLMs. Specifically, we incorporates past-only quantization to improve the computation of attention. Additionally, we introduce two-dimensional quantization strategy to handle the distribution of KV cache, along with a cross-block reconstruction regularization for parameter optimization. Experiments show that WKVQuant achieves almost comparable memory savings to weight-activation quantization, while also approaching the performance of weight-only quantization.

Identifying the features learned by neural networks is a core challenge in mechanistic interpretability. Sparse autoencoders (SAEs), which learn a sparse, overcomplete dictionary that reconstructs a network's internal activations, have been used to identify these features. However, SAEs may learn more about the structure of the datatset than the computational structure of the network. There is therefore only indirect reason to believe that the directions found in these dictionaries are functionally important to the network. We propose end-to-end (e2e) sparse dictionary learning, a method for training SAEs that ensures the features learned are functionally important by minimizing the KL divergence between the output distributions of the original model and the model with SAE activations inserted. Compared to standard SAEs, e2e SAEs offer a Pareto improvement: They explain more network performance, require fewer total features, and require fewer simultaneously active features per datapoint, all with no cost to interpretability. We explore geometric and qualitative differences between e2e SAE features and standard SAE features. E2e dictionary learning brings us closer to methods that can explain network behavior concisely and accurately. We release our library for training e2e SAEs and reproducing our analysis at https://github.com/ApolloResearch/e2e_sae

Art style classification remains a formidable challenge in computational aesthetics due to the scarcity of expertly labeled datasets and the intricate, often nonlinear interplay of stylistic elements. While recent dual-teacher self-supervised frameworks reduce reliance on labeled data, their linear projection layers and localized focus struggle to model global compositional context and complex style-feature interactions. We enhance the dual-teacher knowledge distillation framework to address these limitations by replacing conventional MLP projection and prediction heads with Kolmogorov-Arnold Networks (KANs). Our approach retains complementary guidance from two teacher networks, one emphasizing localized texture and brushstroke patterns, the other capturing broader stylistic hierarchies while leveraging KANs' spline-based activations to model nonlinear feature correlations with mathematical precision. Experiments on WikiArt and Pandora18k demonstrate that our approach outperforms the base dual teacher architecture in Top-1 accuracy. Our findings highlight the importance of KANs in disentangling complex style manifolds, leading to better linear probe accuracy than MLP projections.

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.

Recent years have witnessed an increasing interest in deploying LLMs on resource-constrained devices, among which quantization has emerged as a promising lightweight technique that converts full-precision model weights and activations into lower-bit formats. Existing weight quantization approaches can be roughly divided into three categories: Post-Training Quantization (PTQ) that calibrates quantized parameters on a small dataset without retraining but suffers from severe performance degradation below 4-bit, Quantization-Aware Training (QAT) that searches low-bit parameters using surrogate gradients but demands substantial computational resources, and Quantization-Aware Distillation that integrates QAT with knowledge transfer from a full-precision teacher but manually selects features to distill and relies heavily on teacher-specific data. In this paper, we propose EdgeRazor, a lightweight framework for LLMs with mixed-precision and extremely low-bit weight quantization. The EdgeRazor framework contains three modules: Mixed-Precision Quantization-Aware Distillation for the fine-grained control of precision, Adaptive Feature Distillation that derives an n-bit student from its 16-bit teacher, and Entropy-Aware KL Divergence on both human-annotated and distilled datasets, whose forward-reverse balance is determined solely by the teacher's output distribution. Empirical investigations of EdgeRazor are conducted on base, instruction-tuned, and multimodal LLMs. Notably, EdgeRazor with 1.88-bit surpasses all contenders with the 3-bit precision, especially outperforms the leading 2-bit PTQ methods by 11.3 points, within a 4-10times lower training budget than the leading QAT approach. EdgeRazor delivers higher compression ratios at all bit width; the 1.58-bit Qwen3-0.6B reduces storage from 1.41 GB to 0.28 GB while accelerating decoding by 15.1times relative to the 16-bit baseline.

Optimizing inference for long-context Large Language Models (LLMs) is increasingly important due to the quadratic compute and linear memory complexity of Transformers. Existing approximation methods, such as key-value (KV) cache dropping, sparse attention, and prompt compression, typically rely on rough predictions of token or KV pair importance. We propose a novel framework for approximate LLM inference that leverages small draft models to more accurately predict the importance of tokens and KV pairs. Specifically, we introduce two instantiations of our proposed framework: (i) SpecKV, which leverages a draft output to accurately assess the importance of each KV pair for more effective KV cache dropping, and (ii) SpecPC, which uses the draft model's attention activations to identify and discard unimportant prompt tokens. To the best of our knowledge, this is the first work to use draft models for approximate LLM inference acceleration, extending their utility beyond traditional lossless speculative decoding. We motivate our methods with theoretical and empirical analyses, and show a strong correlation between the attention patterns of draft and target models. Extensive experiments on long-context benchmarks show that our methods consistently achieve higher accuracy than existing baselines, while preserving the same improvements in memory usage, latency, and throughput. Our code is available at https://github.com/furiosa-ai/draft-based-approx-llm.

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.

Model-internal features can be studied through both their ability to identify a specified concept and their causal effect when manipulated, e.g., through steering or weight editing. A prominent approach to feature learning is Sparse Autoencoders (SAEs), which learn broad feature dictionaries whose relation to particular concepts is typically identified post hoc. However, many interpretability questions are instead hypothesis-driven and concern a concept specified in advance. We study this setting as targeted feature learning, where a single feature is constructed for such a predefined concept. We present a controlled comparison across three model signals (activation values, activation gradients, and parameter gradients) and two estimators (contrastive mean and a learned one-dimensional encoder-decoder), yielding six targeted methods, with CAA and GRADIEND as existing instances and four new methods covering the remaining combinations. We compare these methods against pretrained SAEs across 15 tasks and three language models, evaluating both detection and causal intervention. Across models, the strongest detection performance is achieved by contrastive activation value methods, whereas the strongest intervention performance is achieved by gradient-based methods. Overall, our results show that targeted feature quality depends jointly on the model signal and estimator, with detection and intervention capturing complementary properties.

In this paper we present an end-to-end speech recognition model with Transformer encoders that can be used in a streaming speech recognition system. Transformer computation blocks based on self-attention are used to encode both audio and label sequences independently. The activations from both audio and label encoders are combined with a feed-forward layer to compute a probability distribution over the label space for every combination of acoustic frame position and label history. This is similar to the Recurrent Neural Network Transducer (RNN-T) model, which uses RNNs for information encoding instead of Transformer encoders. The model is trained with the RNN-T loss well-suited to streaming decoding. We present results on the LibriSpeech dataset showing that limiting the left context for self-attention in the Transformer layers makes decoding computationally tractable for streaming, with only a slight degradation in accuracy. We also show that the full attention version of our model beats the-state-of-the art accuracy on the LibriSpeech benchmarks. Our results also show that we can bridge the gap between full attention and limited attention versions of our model by attending to a limited number of future frames.

Skills are a natural unit for describing what a language model can do and how its behavior can be changed. However, existing characterizations rely on human-written taxonomies, textual descriptions, or manual profiling pipelines--all external hypotheses about what matters that need not align with the model's internal representations. We argue that when the goal is to intervene on model behavior, skill characterization should be *model-native*: grounded in the model's own representations rather than imposed through external ontologies. We instantiate this view by recovering a compact orthogonal basis from sequence-level activations. The resulting basis is semantically interpretable but need not correspond to any predefined human ontology; instead, it captures axes of behavioral variation that the model itself organizes around. We validate this characterization on reasoning post-training, using the recovered basis for both SFT data selection and inference-time steering. We develop lightweight proxy interventions to identify which directions are most useful for a given model. Across Llama3-8B and Qwen2.5-3B, selecting data along those directions improves Pass@1 by up to 20% on MATH and 41% on AMC, outperforming data selection based on human-characterized skills. Because the basis lives in activation space, the same directions also serve as steering vectors at inference time, improving Pass@8 by up to 4.8% on MATH--an intervention that human-characterized skills cannot support. We further validate the characterization on safety alignment, where selecting adversarial training data for model-native skill coverage rather than textual diversity yields more sample-efficient learning. These results suggest that recovering skills from the model's own representations, rather than imposing them externally, provides a more effective foundation for intervening on model behavior. Codes are open-sourced.

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.

The Segment Anything Model (SAM) achieves strong open-vocabulary segmentation, but its ViT-based image encoders dominate inference latency and memory. Existing activation compression methods, such as token merging, reduce the token length to process, yet introduce non-trivial runtime overhead and encounter catastrophic quality drop under high compression. Other methods applying Sparse Attention focus on attention alone, leaving the MLP fully dense and capping achievable speedup. We propose SparseSAM, a (i) training-free structured sparsification framework that jointly accelerates attention and MLP layers while preserving token identity. SparseSAM introduces (ii) Stripe-Sort Attention, which uses a deterministic Z-order permutation to transform dense attention into static hardware-friendly sparse patterns, eliminating dynamic masking overhead. SparseSAM further introduces a (iii) Residual-Consistency MLP that routes only informative tokens through the MLP while propagating remaining tokens through the residual pathway. Across four segmentation benchmarks, SparseSAM loses only 0.004 mIoU at a 0.4 density and 0.021 mIoU at 0.3, a 2.10x reduction in accuracy loss versus token merging advances, while achieving 2x faster inference and 2.8x memory reduction.