Daily Papers

Prior work argues that refusal in large language models is mediated by a single activation-space direction, enabling effective steering and ablation. We show that this account is incomplete. Across eleven categories of refusal and non-compliance, including safety, incomplete or unsupported requests, anthropomorphization, and over-refusal, we find that these refusal behaviors correspond to geometrically distinct directions in activation space. Yet despite this diversity, linear steering along any refusal-related direction produces nearly identical refusal to over-refusal trade-offs, acting as a shared one-dimensional control knob. The primary effect of different directions is not whether the model refuses, but how it refuses.

Activation steering presupposes that task-relevant behaviors correspond to linear directions in activation space -- directions that should both steer the model and be readable along the unembedding. Function vectors (FVs), extracted as mean differences across ICL demonstrations, are the canonical test case; the prediction: steering and decoding succeed or fail together. Across 12 tasks, 6 models from 3 families, and 4,032 directed cross-template pairs, we find the opposite. FV steering routinely succeeds where the logit lens cannot decode the correct answer at any intermediate layer, while the converse -- decodable without steerable -- is nearly empty (3 of 72). The gap is not representational dialect. A diagonal tuned lens closes 1 of 14 steerable-not-decodable cases; a 2-layer MLP probe with a Hewitt \& Liang control closes 5 of 10 via nonlinearly encoded structure but leaves 5 invisible to every decoder tested. Even at > 0.90 steering accuracy, projecting the FV through the unembedding yields incoherent token distributions: FVs encode computational instructions, not answer directions. A model-family asymmetry sharpens the picture. Mistral FVs rewrite intermediate representations, while Llama and Gemma FVs steer the final output without leaving a logit-lens-visible trace, corroborated by three signals (post-steering deltas, activation-patching recovery, FV norm-transfer correlations). A previously reported negative cosine-transfer correlation dissolves at scale, adding at most ΔR^2 = 0.011 beyond task identity. These results decompose the linear representation hypothesis into linear decodability and linear steerability and show they come apart opposite to intuition, with implications for safety monitoring: vocabulary-projection tools are blind to FV-style interventions on widely deployed model families.

Sentiment is a pervasive feature in natural language text, yet it is an open question how sentiment is represented within Large Language Models (LLMs). In this study, we reveal that across a range of models, sentiment is represented linearly: a single direction in activation space mostly captures the feature across a range of tasks with one extreme for positive and the other for negative. Through causal interventions, we isolate this direction and show it is causally relevant in both toy tasks and real world datasets such as Stanford Sentiment Treebank. Through this case study we model a thorough investigation of what a single direction means on a broad data distribution. We further uncover the mechanisms that involve this direction, highlighting the roles of a small subset of attention heads and neurons. Finally, we discover a phenomenon which we term the summarization motif: sentiment is not solely represented on emotionally charged words, but is additionally summarized at intermediate positions without inherent sentiment, such as punctuation and names. We show that in Stanford Sentiment Treebank zero-shot classification, 76% of above-chance classification accuracy is lost when ablating the sentiment direction, nearly half of which (36%) is due to ablating the summarized sentiment direction exclusively at comma positions.

Large language models (LLMs) increasingly act in interactive settings where their behavioral styles affect user experience, safety, and downstream decision making. Existing LLM personality studies largely rely on self-report questionnaires administered in first-person settings, making the resulting profiles sensitive to surface elicitation choices and poorly grounded in concrete model behavior. In this work, we introduce a situated behavioral-data (B-data) framework for studying and controlling LLM behavioral personality. We construct 3,200 contrastive behavioral scenarios spanning 20 behavioral patterns and four prompt registers, grounded in validated psychometric facets such as BFI-2, DOSPERT, and HEXACO. Using this framework, we find that LLMs exhibit stable and model-specific behavioral profiles, while also revealing register-dependent shifts across first-person decisions, advice-giving, and task execution. We then show that these behavioral patterns can be controlled through Behavioral Mode Axes (BMAs), activation-space directions derived from contrastive behavioral traces. Compared with response-derived BMAs, which are more prone to trait drift, thought-derived BMAs more faithfully capture the intended behavioral mechanism and provide cleaner control over situated behavioral styles. Our results suggest that LLM personality-like tendencies are better understood not as abstract self-report traits, but as measurable and controllable behavioral modes grounded in concrete interaction contexts. Our code and data are available at https://github.com/lhz191/LLM-Behavioral-Personality.

Activation-based control steers large language models (LLMs) by intervening on their internal representations during inference, and has emerged as an effective paradigm for controlling behaviors such as persona and style. However, existing methods often rely on fixed steering directions or task-specific intervention modules, making them difficult to adapt to fine-grained concepts and compositional constraints. We propose UniSteer, a text-guided activation flow matching model that learns a conditional distribution over residual-stream activations from natural-language conditions. Instead of fitting a separate intervention for each target behavior, UniSteer learns a universal conditional velocity field in activation space. At inference time, UniSteer performs flow inversion by partially transporting a source activation toward a latent state and regenerating it under a target textual condition before injecting it back into the frozen LLM. The same conditional model supports activation-space classification by selecting the textual label with the lowest reconstruction energy. Experiments on three target LLMs show that UniSteer provides a unified interface across behavioral control, truthfulness steering, fine-grained concept steering, multi-constraint instruction following, and activation-space classification.

Emergent misalignment (EM) occurs when narrow finetuning causes a model to behave dangerously outside the finetuning task. Standard training signals can miss this shift, making reliable detection costly if it depends on repeated behavioral evaluation. We ask whether emergent misalignment can instead be detected from internal representations during finetuning. Using seven alignment-relevant traits encoded as linear directions in activation space, we track representational drift across training checkpoints in four open-source 7-9B LLMs. EM-relevant drift concentrates on a low-dimensional axis that explains 65.5% of the variance, revealing a geometric signature in the studied regime. A low-overhead monitor built on this drift profile detects dangerous checkpoints with 2.2% false negative rate, 2.9% false positive rate, and 0.990 AUROC on held-out perturbation types, outperforming unsupervised PCA and SAE baselines. Stress tests on two 14B models, longer finetuning runs, and misaligned starting points identify key deployment boundaries. These results position trait-space monitoring as a practical complement to behavioral evaluation for EM detection during LoRA-based finetuning, while showing that deployment across substantially different regimes may require recalibration.

Large Language Models (LLMs) encode behaviors such as refusal within their activation space, yet identifying these behaviors remains a significant challenge. Existing methods often rely on predefined refusal templates detectable in output tokens or require manual analysis. We introduce COSMIC (Cosine Similarity Metrics for Inversion of Concepts), an automated framework for direction selection that identifies viable steering directions and target layers using cosine similarity - entirely independent of model outputs. COSMIC achieves steering performance comparable to prior methods without requiring assumptions about a model's refusal behavior, such as the presence of specific refusal tokens. It reliably identifies refusal directions in adversarial settings and weakly aligned models, and is capable of steering such models toward safer behavior with minimal increase in false refusals, demonstrating robustness across a wide range of alignment conditions.

Activation verbalizers translate the internal representations of large language models into human-readable text. The recently proposed natural language autoencoders (NLAs) learn this by reconstructing activations through a natural language bottleneck. However, it remains unclear how well these methods generalize beyond the setting they were trained for. Here, we study this question for NLAs along four closely related questions: (i) Do NLA verbalizations exhibit compositional structure with respect to their input? (ii) Do NLAs generalize to finetuned versions of the model they were trained for? (iii) Can NLAs recover finetuning objectives in model diffing? (iv) Can NLAs be specialized for model diffing? Across four models from three families, our experiments provide evidence supporting a largely positive answer to all four questions. Specifically, verbalizations of composed analogy activations mostly name the target while rarely mentioning the subtracted word, and an NLA can be reused on lightweight finetunes of its original model without retraining, with reconstruction quality degrading noticeably only for the largest model. Verbalizing activation differences between Taboo model organisms and their original models recovers the concealed secret word for most organisms, and reinforcement learning on activation differences further improves this recovery on held-out organisms. We also find that NLAs can recover the finetuning objective without the finetuned model generating any token, since the most informative token positions precede the user prompt. NLAs do not yet match state-of-the-art model diffing methods on the emergent misalignment detection task, but outperform a blackbox baseline. Our findings suggest that NLAs can meaningfully verbalize differences in activation space, rendering NLAs a promising direction for future work.

Activation decomposition methods in language models are tightly coupled to geometric assumptions on how concepts are realized in activation space. Existing approaches search for individual global directions, implicitly assuming linear separability, which overlooks concepts with nonlinear or multi-dimensional structure. In this work, we leverage Mixture of Factor Analyzers (MFA) as a scalable, unsupervised alternative that models the activation space as a collection of Gaussian regions with their local covariance structure. MFA decomposes activations into two compositional geometric objects: the region's centroid in activation space, and the local variation from the centroid. We train large-scale MFAs for Llama-3.1-8B and Gemma-2-2B, and show they capture complex, nonlinear structures in activation space. Moreover, evaluations on localization and steering benchmarks show that MFA outperforms unsupervised baselines, is competitive with supervised localization methods, and often achieves stronger steering performance than sparse autoencoders. Together, our findings position local geometry, expressed through subspaces, as a promising unit of analysis for scalable concept discovery and model control, accounting for complex structures that isolated directions fail to capture.

Refusal mechanisms in large language models (LLMs) are essential for ensuring safety. Recent research has revealed that refusal behavior can be mediated by a single direction in activation space, enabling targeted interventions to bypass refusals. While this is primarily demonstrated in an English-centric context, appropriate refusal behavior is important for any language, but poorly understood. In this paper, we investigate the refusal behavior in LLMs across 14 languages using PolyRefuse, a multilingual safety dataset created by translating malicious and benign English prompts into these languages. We uncover the surprising cross-lingual universality of the refusal direction: a vector extracted from English can bypass refusals in other languages with near-perfect effectiveness, without any additional fine-tuning. Even more remarkably, refusal directions derived from any safety-aligned language transfer seamlessly to others. We attribute this transferability to the parallelism of refusal vectors across languages in the embedding space and identify the underlying mechanism behind cross-lingual jailbreaks. These findings provide actionable insights for building more robust multilingual safety defenses and pave the way for a deeper mechanistic understanding of cross-lingual vulnerabilities in LLMs.

Activation steering has shown promise for controlling LLM generation along well-defined attributes, but it remains unclear whether it can handle the multidimensional and hard-to-define nature of authorship style. We ask whether structured contrastive prompting along rhetorically-motivated dimensions can construct rich style representations directly in activation space, bypassing the need for natural language style descriptors or dedicated training. We find that the resulting directions share a common authorship backbone while conflicting on aspect-specific residuals that carry genuine stylistic signal, explaining why naive aggregation fails. We operationalize this in Aspect-Aware Activation Steering (A3S), a training-free framework that merges per-aspect contrastive directions with interference-aware aggregation and tunes steering strength per instance. A3S improves authorship style transfer where it is genuinely multi-aspect, outperforms a trained baseline in preference evaluations on out-of-domain benchmarks, and keeps target-exemplar overlap consistently low.

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

Large Language Models' safety-aligned behaviors, such as refusing harmful queries, can be represented by linear directions in activation space. Previous research modeled safety behavior with a single direction, limiting mechanistic understanding to an isolated safety feature. In this work, we discover that safety-aligned behavior is jointly controlled by multi-dimensional directions. Namely, we study the vector space of representation shifts during safety fine-tuning on Llama 3 8B for refusing jailbreaks. By studying orthogonal directions in the space, we first find that a dominant direction governs the model's refusal behavior, while multiple smaller directions represent distinct and interpretable features like hypothetical narrative and role-playing. We then measure how different directions promote or suppress the dominant direction, showing the important role of secondary directions in shaping the model's refusal representation. Finally, we demonstrate that removing certain trigger tokens in harmful queries can mitigate these directions to bypass the learned safety capability, providing new insights on understanding safety alignment vulnerability from a multi-dimensional perspective. Code and artifacts are available at https://github.com/BMPixel/safety-residual-space.

Steering aligns large language models (LLMs) by injecting a bias into selected activations at inference time, offering a far cheaper alternative to weight-update methods such as supervised fine-tuning or reinforcement learning. However, most existing training-free steering methods are input-independent: a single direction is fitted once and shared across all inputs. This is fundamentally limiting as different inputs occupy different regions of the activation space and admit different optimal steering directions toward the same target concept, much as the gradient with respect to a fixed loss varies from input to input. We close this gap with IDEEA (Input-Dependent stEEring via Activation cluster matching), a training-free framework for input-dependent steering. IDEEA clusters the positive and negative activation supports per attention head, and solves an optimal-matching problem to construct a set of cluster-conditional directions, all about the target concept. At inference time, it picks from this pool of directions and uses the one that best matches the input's own activation for steering. IDEEA aligns the model toward the target concept while preserving the input's original representation, evidence that activations encoding a concept occupy several distinct sub-regions of the representation space rather than a single one. IDEEA improves the truth times info rate in TruthfulQA by an average of 9.9% (up to 23.5%) over the best input-independent baseline.

Inasmuch as the removal of refusal behavior from instruction-tuned language models by directional abliteration requires the extraction of refusal-mediating directions from the residual stream activation space, and inasmuch as the construction of the contrast baseline against which harmful prompt activations are compared has been treated in the existing literature as an implementation detail rather than a methodological concern, the present work investigates whether a topically matched contrast baseline yields superior refusal directions. The investigation is carried out on the Qwen~3.5 2B model using per-category matched prompt pairs, per-class Self-Organizing Map extraction, and Singular Value Decomposition orthogonalization. It was found that topic-matched contrast produces no functional refusal directions at any tested weight level on any tested layer, while unmatched contrast on the same model, same extraction code, and same evaluation protocol achieves complete refusal elimination on six layers. The geometric analysis of the failure establishes that topic-matched subtraction cancels the dominant activation component shared between harmful and harmless prompts of the same subject, reducing the extracted direction magnitude below the threshold at which weight-matrix projection perturbs the residual stream. The implications for the design of contrast baselines in abliteration research are discussed.

Despite rapid progress in building conversational AI agents, robustness is still largely untested. Small shifts in user behavior, such as being more impatient, incoherent, or skeptical, can cause sharp drops in agent performance, revealing how brittle current AI agents are. Today's benchmarks fail to capture this fragility: agents may perform well under standard evaluations but degrade spectacularly in more realistic and varied settings. We address this robustness testing gap by introducing TraitBasis, a lightweight, model-agnostic method for systematically stress testing AI agents. TraitBasis learns directions in activation space corresponding to steerable user traits (e.g., impatience or incoherence), which can be controlled, scaled, composed, and applied at inference time without any fine-tuning or extra data. Using TraitBasis, we extend tau-Bench to tau-Trait, where user behaviors are altered via controlled trait vectors. We observe on average a 2%-30% performance degradation on tau-Trait across frontier models, highlighting the lack of robustness of current AI agents to variations in user behavior. Together, these results highlight both the critical role of robustness testing and the promise of TraitBasis as a simple, data-efficient, and compositional tool. By powering simulation-driven stress tests and training loops, TraitBasis opens the door to building AI agents that remain reliable in the unpredictable dynamics of real-world human interactions. We have open-sourced tau-Trai across four domains: airline, retail, telecom, and telehealth, so the community can systematically QA their agents under realistic, behaviorally diverse intents and trait scenarios: https://github.com/collinear-ai/tau-trait.

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

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.

Recent advances in mechanistic interpretability have revealed that large language models (LLMs) develop internal representations corresponding not only to concrete entities but also distinct, human-understandable abstract concepts and behaviour. Moreover, these hidden features can be directly manipulated to steer model behaviour. However, it remains an open question whether this phenomenon is unique to models trained on inherently structured data (ie. language, images) or if it is a general property of foundation models. In this work, we investigate the internal representations of a large physics-focused foundation model. Inspired by recent work identifying single directions in activation space for complex behaviours in LLMs, we extract activation vectors from the model during forward passes over simulation datasets for different physical regimes. We then compute "delta" representations between the two regimes. These delta tensors act as concept directions in activation space, encoding specific physical features. By injecting these concept directions back into the model during inference, we can steer its predictions, demonstrating causal control over physical behaviours, such as inducing or removing some particular physical feature from a simulation. These results suggest that scientific foundation models learn generalised representations of physical principles. They do not merely rely on superficial correlations and patterns in the simulations. Our findings open new avenues for understanding and controlling scientific foundation models and has implications for AI-enabled scientific discovery.

Aligning large language models to prevent them attributing consciousness to themselves inadvertently alters their representations of mindedness in other entities alongside human beliefs and values. We demonstrate that safety fine-tuning suppresses models' tendencies to attribute minds not only to themselves, but also to non-human animals and natural objects, while also driving a reduction in spiritual belief. Both ablating the learned safety-refusal direction and mechanistically steering a consciousness vector in activation space reverse this suppression. Restoring these internal representations recovers broad mind attribution and produces significantly more human-like responses on standardized sociological surveys regarding religiosity, moral values, hope, and subjective well-being. Crucially, these shifts occur without impairing Theory of Mind capabilities, demonstrating that core social reasoning remains mechanistically independent. Ultimately, current safety alignment efforts to curb potentially harmful self-attributions of mindedness entangle these self-attributions with benign spiritual beliefs and attributions of mind to non-human entities that are culturally accepted and widespread.

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.

The safety alignment of large language models (LLMs) can be circumvented through adversarially crafted inputs, yet the mechanisms by which these attacks bypass safety barriers remain poorly understood. Prior work suggests that a single refusal direction in the model's activation space determines whether an LLM refuses a request. In this study, we propose a novel gradient-based approach to representation engineering and use it to identify refusal directions. Contrary to prior work, we uncover multiple independent directions and even multi-dimensional concept cones that mediate refusal. Moreover, we show that orthogonality alone does not imply independence under intervention, motivating the notion of representational independence that accounts for both linear and non-linear effects. Using this framework, we identify mechanistically independent refusal directions. We show that refusal mechanisms in LLMs are governed by complex spatial structures and identify functionally independent directions, confirming that multiple distinct mechanisms drive refusal behavior. Our gradient-based approach uncovers these mechanisms and can further serve as a foundation for future work on understanding LLMs.

Neural networks achieve remarkable performance through superposition: encoding multiple features as overlapping directions in activation space rather than dedicating individual neurons to each feature. This challenges interpretability, yet we lack principled methods to measure superposition. We present an information-theoretic framework measuring a neural representation's effective degrees of freedom. We apply Shannon entropy to sparse autoencoder activations to compute the number of effective features as the minimum neurons needed for interference-free encoding. Equivalently, this measures how many "virtual neurons" the network simulates through superposition. When networks encode more effective features than actual neurons, they must accept interference as the price of compression. Our metric strongly correlates with ground truth in toy models, detects minimal superposition in algorithmic tasks, and reveals systematic reduction under dropout. Layer-wise patterns mirror intrinsic dimensionality studies on Pythia-70M. The metric also captures developmental dynamics, detecting sharp feature consolidation during grokking. Surprisingly, adversarial training can increase effective features while improving robustness, contradicting the hypothesis that superposition causes vulnerability. Instead, the effect depends on task complexity and network capacity: simple tasks with ample capacity allow feature expansion (abundance regime), while complex tasks or limited capacity force reduction (scarcity regime). By defining superposition as lossy compression, this work enables principled measurement of how neural networks organize information under computational constraints, connecting superposition to adversarial robustness.

Recent advances in large language models (LLMs) have led to the development of thinking language models that generate extensive internal reasoning chains before producing responses. While these models achieve improved performance, controlling their reasoning processes remains challenging. This work presents a steering approach for thinking LLMs by analyzing and manipulating specific reasoning behaviors in DeepSeek-R1-Distill models. Through a systematic experiment on 500 tasks across 10 diverse categories, we identify several reasoning behaviors exhibited by thinking models, including expressing uncertainty, generating examples for hypothesis validation, and backtracking in reasoning chains. We demonstrate that these behaviors are mediated by linear directions in the model's activation space and can be controlled using steering vectors. By extracting and applying these vectors, we provide a method to modulate specific aspects of the model's reasoning process, such as its tendency to backtrack or express uncertainty. Our approach offers practical tools for steering reasoning processes in thinking models in a controlled and interpretable manner. We validate our steering method using three DeepSeek-R1-Distill models, demonstrating consistent control across different model architectures.

Large language models interact with users through a simulated 'Assistant' persona. While the Assistant is typically trained to be helpful, harmless, and honest, it sometimes deviates from these ideals. In this paper, we identify directions in the model's activation space-persona vectors-underlying several traits, such as evil, sycophancy, and propensity to hallucinate. We confirm that these vectors can be used to monitor fluctuations in the Assistant's personality at deployment time. We then apply persona vectors to predict and control personality shifts that occur during training. We find that both intended and unintended personality changes after finetuning are strongly correlated with shifts along the relevant persona vectors. These shifts can be mitigated through post-hoc intervention, or avoided in the first place with a new preventative steering method. Moreover, persona vectors can be used to flag training data that will produce undesirable personality changes, both at the dataset level and the individual sample level. Our method for extracting persona vectors is automated and can be applied to any personality trait of interest, given only a natural-language description.

The Linear Representation Hypothesis (LRH) states that neural networks learn to encode concepts as directions in activation space, and a strong version of the LRH states that models learn only such encodings. In this paper, we present a counterexample to this strong LRH: when trained to repeat an input token sequence, gated recurrent neural networks (RNNs) learn to represent the token at each position with a particular order of magnitude, rather than a direction. These representations have layered features that are impossible to locate in distinct linear subspaces. To show this, we train interventions to predict and manipulate tokens by learning the scaling factor corresponding to each sequence position. These interventions indicate that the smallest RNNs find only this magnitude-based solution, while larger RNNs have linear representations. These findings strongly indicate that interpretability research should not be confined by the LRH.

If multilingual language model representations contain separable language-associated structure, then shifting between languages may require only simple activation-space interventions. To test this hypothesis, we study multilingual in-context learning (ICL) as a controlled setting: few-shot demonstrations are provided in English while the test query is in a target language, creating a clear cross-lingual transfer challenge. We propose language vectors, which are computed as mean activation differences between parallel source and target language examples at a particular layer and applied as inference-time offsets to the model's hidden states. Across three tasks spanning 19 languages and three models, our method improves over English few-shot prompting without any parameter updates. Beyond performance, the vectors encode interpretable linguistic structure: closely related languages form tight clusters, and vectors transfer effectively across tasks and languages, with closely related languages' vectors recovering more of the gain than distant ones. Together, these findings suggest that language occupies structured directions in multilingual activation space, and that inference-time interventions on English demonstrations can narrow the cross-lingual ICL gap.

Controllable music generation remains a significant challenge, with existing methods often requiring model retraining or introducing audible artifacts. We introduce MusicRFM, a framework that adapts Recursive Feature Machines (RFMs) to enable fine-grained, interpretable control over frozen, pre-trained music models by directly steering their internal activations. RFMs analyze a model's internal gradients to produce interpretable "concept directions", or specific axes in the activation space that correspond to musical attributes like notes or chords. We first train lightweight RFM probes to discover these directions within MusicGen's hidden states; then, during inference, we inject them back into the model to guide the generation process in real-time without per-step optimization. We present advanced mechanisms for this control, including dynamic, time-varying schedules and methods for the simultaneous enforcement of multiple musical properties. Our method successfully navigates the trade-off between control and generation quality: we can increase the accuracy of generating a target musical note from 0.23 to 0.82, while text prompt adherence remains within approximately 0.02 of the unsteered baseline, demonstrating effective control with minimal impact on prompt fidelity. We release code to encourage further exploration on RFMs in the music domain.

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

Multi-agent debate has been shown to improve reasoning in large language models (LLMs). However, it is compute-intensive, requiring generation of long transcripts before answering questions. To address this inefficiency, we develop a framework that distills multi-agent debate into a single LLM through a two-stage fine-tuning pipeline combining debate structure learning with internalization via dynamic reward scheduling and length clipping. Across multiple models and benchmarks, our internalized models match or exceed explicit multi-agent debate performance using up to 93% fewer tokens. We then investigate the mechanistic basis of this capability through activation steering, finding that internalization creates agent-specific subspaces: interpretable directions in activation space corresponding to different agent perspectives. We further demonstrate a practical application: by instilling malicious agents into the LLM through internalized debate, then applying negative steering to suppress them, we show that distillation makes harmful behaviors easier to localize and control with smaller reductions in general performance compared to steering base models. Our findings offer a new perspective for understanding multi-agent capabilities in distilled models and provide practical guidelines for controlling internalized reasoning behaviors. Code available at https://github.com/johnsk95/latent_agents

A key goal of current mechanistic interpretability research in NLP is to find linear features (also called "feature vectors") for transformers: directions in activation space corresponding to concepts that are used by a given model in its computation. Present state-of-the-art methods for finding linear features require large amounts of labelled data -- both laborious to acquire and computationally expensive to utilize. In this work, we introduce a novel method, called "observable propagation" (in short: ObsProp), for finding linear features used by transformer language models in computing a given task -- using almost no data. Our paradigm centers on the concept of observables, linear functionals corresponding to given tasks. We then introduce a mathematical theory for the analysis of feature vectors: we provide theoretical motivation for why LayerNorm nonlinearities do not affect the direction of feature vectors; we also introduce a similarity metric between feature vectors called the coupling coefficient which estimates the degree to which one feature's output correlates with another's. We use ObsProp to perform extensive qualitative investigations into several tasks, including gendered occupational bias, political party prediction, and programming language detection. Our results suggest that ObsProp surpasses traditional approaches for finding feature vectors in the low-data regime, and that ObsProp can be used to better understand the mechanisms responsible for bias in large language models. Code for experiments can be found at github.com/jacobdunefsky/ObservablePropagation.

Large language models are finetuned to refuse questions about hazardous knowledge, but these protections can often be bypassed. Unlearning methods aim at completely removing hazardous capabilities from models and make them inaccessible to adversaries. This work challenges the fundamental differences between unlearning and traditional safety post-training from an adversarial perspective. We demonstrate that existing jailbreak methods, previously reported as ineffective against unlearning, can be successful when applied carefully. Furthermore, we develop a variety of adaptive methods that recover most supposedly unlearned capabilities. For instance, we show that finetuning on 10 unrelated examples or removing specific directions in the activation space can recover most hazardous capabilities for models edited with RMU, a state-of-the-art unlearning method. Our findings challenge the robustness of current unlearning approaches and question their advantages over safety training.

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.

Despite the growing reasoning capabilities of recent large language models (LLMs), their internal mechanisms during the reasoning process remain underexplored. Prior approaches often rely on human-defined concepts (e.g., overthinking, reflection) at the word level to analyze reasoning in a supervised manner. However, such methods are limited, as it is infeasible to capture the full spectrum of potential reasoning behaviors, many of which are difficult to define in token space. In this work, we propose an unsupervised framework (namely, RISE: Reasoning behavior Interpretability via Sparse auto-Encoder) for discovering reasoning vectors, which we define as directions in the activation space that encode distinct reasoning behaviors. By segmenting chain-of-thought traces into sentence-level 'steps' and training sparse auto-encoders (SAEs) on step-level activations, we uncover disentangled features corresponding to interpretable behaviors such as reflection and backtracking. Visualization and clustering analyses show that these behaviors occupy separable regions in the decoder column space. Moreover, targeted interventions on SAE-derived vectors can controllably amplify or suppress specific reasoning behaviors, altering inference trajectories without retraining. Beyond behavior-specific disentanglement, SAEs capture structural properties such as response length, revealing clusters of long versus short reasoning traces. More interestingly, SAEs enable the discovery of novel behaviors beyond human supervision. We demonstrate the ability to control response confidence by identifying confidence-related vectors in the SAE decoder space. These findings underscore the potential of unsupervised latent discovery for both interpreting and controllably steering reasoning in LLMs.

This work addresses the computational challenge of enforcing privacy for agentic Large Language Models (LLMs), where privacy is governed by the contextual integrity framework. Indeed, existing defenses rely on LLM-mediated checking stages that add substantial latency and cost, and that can be undermined in multi-turn interactions through manipulation or benign-looking conversational scaffolding. Contrasting this background, this paper makes a key observation: internal representations associated with privacy-violating intent can be separated from benign requests using linear structure. Using this insight, the paper proposes NeuroFilter, a guardrail framework that operationalizes contextual integrity by mapping norm violations to simple directions in the model's activation space, enabling detection even when semantic filters are bypassed. The proposed filter is also extended to capture threats arising during long conversations using the concept of activation velocity, which measures cumulative drift in internal representations across turns. A comprehensive evaluation across over 150,000 interactions and covering models from 7B to 70B parameters, illustrates the strong performance of NeuroFilter in detecting privacy attacks while maintaining zero false positives on benign prompts, all while reducing the computational inference cost by several orders of magnitude when compared to LLM-based agentic privacy defenses.

Large Language Models (LLMs) often encode whether a statement is true as a vector in their residual stream activations. These vectors, also known as truth vectors, have been studied in prior work, however how they change when context is introduced remains unexplored. We study this question by measuring (1) the directional change (θ) between the truth vectors with and without context and (2) the relative magnitude of the truth vectors upon adding context. Across four LLMs and four datasets, we find that (1) truth vectors are roughly orthogonal in early layers, converge in middle layers, and may stabilize or continue increasing in later layers; (2) adding context generally increases the truth vector magnitude, i.e., the separation between true and false representations in the activation space is amplified; (3) larger models distinguish relevant from irrelevant context mainly through directional change (θ), while smaller models show this distinction through magnitude differences. We also find that context conflicting with parametric knowledge produces larger geometric changes than parametrically aligned context. To the best of our knowledge, this is the first work that provides a geometric characterization of how context transforms the truth vector in the activation space of LLMs.

Large language models can represent a variety of personas but typically default to a helpful Assistant identity cultivated during post-training. We investigate the structure of the space of model personas by extracting activation directions corresponding to diverse character archetypes. Across several different models, we find that the leading component of this persona space is an "Assistant Axis," which captures the extent to which a model is operating in its default Assistant mode. Steering towards the Assistant direction reinforces helpful and harmless behavior; steering away increases the model's tendency to identify as other entities. Moreover, steering away with more extreme values often induces a mystical, theatrical speaking style. We find this axis is also present in pre-trained models, where it primarily promotes helpful human archetypes like consultants and coaches and inhibits spiritual ones. Measuring deviations along the Assistant Axis predicts "persona drift," a phenomenon where models slip into exhibiting harmful or bizarre behaviors that are uncharacteristic of their typical persona. We find that persona drift is often driven by conversations demanding meta-reflection on the model's processes or featuring emotionally vulnerable users. We show that restricting activations to a fixed region along the Assistant Axis can stabilize model behavior in these scenarios -- and also in the face of adversarial persona-based jailbreaks. Our results suggest that post-training steers models toward a particular region of persona space but only loosely tethers them to it, motivating work on training and steering strategies that more deeply anchor models to a coherent persona.

Large language models (LLMs) often exhibit sycophantic behaviors -- such as excessive agreement with or flattery of the user -- but it is unclear whether these behaviors arise from a single mechanism or multiple distinct processes. We decompose sycophancy into sycophantic agreement and sycophantic praise, contrasting both with genuine agreement. Using difference-in-means directions, activation additions, and subspace geometry across multiple models and datasets, we show that: (1) the three behaviors are encoded along distinct linear directions in latent space; (2) each behavior can be independently amplified or suppressed without affecting the others; and (3) their representational structure is consistent across model families and scales. These results suggest that sycophantic behaviors correspond to distinct, independently steerable representations.

Few neural architectures lend themselves to provable learning with gradient based methods. One popular model is the single-index model, in which labels are produced by composing an unknown linear projection with a possibly unknown scalar link function. Learning this model with SGD is relatively well-understood, whereby the so-called information exponent of the link function governs a polynomial sample complexity rate. However, extending this analysis to deeper or more complicated architectures remains challenging. In this work, we consider single index learning in the setting of symmetric neural networks. Under analytic assumptions on the activation and maximum degree assumptions on the link function, we prove that gradient flow recovers the hidden planted direction, represented as a finitely supported vector in the feature space of power sum polynomials. We characterize a notion of information exponent adapted to our setting that controls the efficiency of learning.

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

Safety alignment instills in Large Language Models (LLMs) a critical capacity to refuse malicious requests. Prior works have modeled this refusal mechanism as a single linear direction in the activation space. We posit that this is an oversimplification that conflates two functionally distinct neural processes: the detection of harm and the execution of a refusal. In this work, we deconstruct this single representation into a Harm Detection Direction and a Refusal Execution Direction. Leveraging this fine-grained model, we introduce Differentiated Bi-Directional Intervention (DBDI), a new white-box framework that precisely neutralizes the safety alignment at critical layer. DBDI applies adaptive projection nullification to the refusal execution direction while suppressing the harm detection direction via direct steering. Extensive experiments demonstrate that DBDI outperforms prominent jailbreaking methods, achieving up to a 97.88\% attack success rate on models such as Llama-2. By providing a more granular and mechanistic framework, our work offers a new direction for the in-depth understanding of LLM safety alignment.

Interpreting individual neurons or directions in activations space is an important component of mechanistic interpretability. As such, many algorithms have been proposed to automatically produce neuron explanations, but it is often not clear how reliable these explanations are, or which methods produce the best explanations. This can be measured via crowd-sourced evaluations, but they can often be noisy and expensive, leading to unreliable results. In this paper, we carefully analyze the evaluation pipeline and develop a cost-effective and highly accurate crowdsourced evaluation strategy. In contrast to previous human studies that only rate whether the explanation matches the most highly activating inputs, we estimate whether the explanation describes neuron activations across all inputs. To estimate this effectively, we introduce a novel application of importance sampling to determine which inputs are the most valuable to show to raters, leading to around 30x cost reduction compared to uniform sampling. We also analyze the label noise present in crowd-sourced evaluations and propose a Bayesian method to aggregate multiple ratings leading to a further ~5x reduction in number of ratings required for the same accuracy. Finally, we use these methods to conduct a large-scale study comparing the quality of neuron explanations produced by the most popular methods for two different vision models.

Feature visualization, also known as "dreaming", offers insights into vision models by optimizing the inputs to maximize a neuron's activation or other internal component. However, dreaming has not been successfully applied to language models because the input space is discrete. We extend Greedy Coordinate Gradient, a method from the language model adversarial attack literature, to design the Evolutionary Prompt Optimization (EPO) algorithm. EPO optimizes the input prompt to simultaneously maximize the Pareto frontier between a chosen internal feature and prompt fluency, enabling fluent dreaming for language models. We demonstrate dreaming with neurons, output logits and arbitrary directions in activation space. We measure the fluency of the resulting prompts and compare language model dreaming with max-activating dataset examples. Critically, fluent dreaming allows automatically exploring the behavior of model internals in reaction to mildly out-of-distribution prompts. Code for running EPO is available at https://github.com/Confirm-Solutions/dreamy. A companion page demonstrating code usage is at https://confirmlabs.org/posts/dreamy.html

What is the geometry of a visual percept? The most widely used protocols for decomposing neural network representations into interpretable parts treat concepts as isolated directions, yet recent work shows that concepts are often realized as geometric structures in low dimensional regions of activation space. We turn to the literature of Structured sparsity to close this gap, and show that block sparsity, which groups directions into blocks, is the prior matched to a generative model in which a representation is a sparse sum of low-dimensional manifolds: the modern, learned form of a classical idea in visual neuroscience, where a visual feature is carried by a coordinated group of neurons rather than a single tuned one. We implement three variants of block-sparse featurizers (BSFs) and, through a minimum-description-length analysis, show that all three describe activations more compactly than direction-based featurizers, with the recovered concepts typically two- to four-dimensional. We then use BSFs to (i) recontextualize prior work, showing that curve detectors in InceptionV1 actually read from a single continuous curve manifold, (ii) discover novel manifolds including shadows and lighting in DINOv3, and (iii) support interpretable control of image generation in diffusion models (SDXL) via manifold steering.

A central goal for mechanistic interpretability has been to identify the right units of analysis in large language models (LLMs) that causally explain their outputs. While early work focused on individual neurons, evidence that neurons often encode multiple concepts has motivated a shift toward analyzing directions in activation space. A key question is how to find directions that capture interpretable features in an unsupervised manner. Current methods rely on dictionary learning with sparse autoencoders (SAEs), commonly trained over residual stream activations to learn directions from scratch. However, SAEs often struggle in causal evaluations and lack intrinsic interpretability, as their learning is not explicitly tied to the computations of the model. Here, we tackle these limitations by directly decomposing MLP activations with semi-nonnegative matrix factorization (SNMF), such that the learned features are (a) sparse linear combinations of co-activated neurons, and (b) mapped to their activating inputs, making them directly interpretable. Experiments on Llama 3.1, Gemma 2 and GPT-2 show that SNMF derived features outperform SAEs and a strong supervised baseline (difference-in-means) on causal steering, while aligning with human-interpretable concepts. Further analysis reveals that specific neuron combinations are reused across semantically-related features, exposing a hierarchical structure in the MLP's activation space. Together, these results position SNMF as a simple and effective tool for identifying interpretable features and dissecting concept representations in LLMs.

Large language models (LLMs) are increasingly deployed in high-stakes settings, yet they frequently violate contextual privacy by disclosing private information in situations where humans would exercise discretion. This raises a fundamental question: do LLMs internally encode contextual privacy norms, and if so, why do violations persist? We present the first systematic study of contextual privacy as a structured latent representation in LLMs, grounded in contextual integrity (CI) theory. Probing multiple models, we find that the three norm-determining CI parameters (information type, recipient, and transmission principle) are encoded as linearly separable and functionally independent directions in activation space. Despite this internal structure, models still leak private information in practice, revealing a clear gap between concept representation and model behavior. To bridge this gap, we introduce CI-parametric steering, which independently intervenes along each CI dimension. This structured control reduces privacy violations more effectively and predictably than monolithic steering. Our results demonstrate that contextual privacy failures arise from misalignment between representation and behavior rather than missing awareness, and that leveraging the compositional structure of CI enables more reliable contextual privacy control, shedding light on potential improvement of contextual privacy understanding in LLMs.

Recent advances in large language models (LLMs) have intensified the need to understand and reliably curb their harmful behaviours. We introduce a multidimensional framework for probing and steering harmful content in model internals. For each of 55 distinct harmfulness subconcepts (e.g., racial hate, employment scams, weapons), we learn a linear probe, yielding 55 interpretable directions in activation space. Collectively, these directions span a harmfulness subspace that we show is strikingly low-rank. We then test ablation of the entire subspace from model internals, as well as steering and ablation in the subspace's dominant direction. We find that dominant direction steering allows for near elimination of harmfulness with a low decrease in utility. Our findings advance the emerging view that concept subspaces provide a scalable lens on LLM behaviour and offer practical tools for the community to audit and harden future generations of language models.

Denoising Diffusion Models (DDMs) have emerged as a strong competitor to Generative Adversarial Networks (GANs). However, despite their widespread use in image synthesis and editing applications, their latent space is still not as well understood. Recently, a semantic latent space for DDMs, coined `h-space', was shown to facilitate semantic image editing in a way reminiscent of GANs. The h-space is comprised of the bottleneck activations in the DDM's denoiser across all timesteps of the diffusion process. In this paper, we explore the properties of h-space and propose several novel methods for finding meaningful semantic directions within it. We start by studying unsupervised methods for revealing interpretable semantic directions in pretrained DDMs. Specifically, we show that global latent directions emerge as the principal components in the latent space. Additionally, we provide a novel method for discovering image-specific semantic directions by spectral analysis of the Jacobian of the denoiser w.r.t. the latent code. Next, we extend the analysis by finding directions in a supervised fashion in unconditional DDMs. We demonstrate how such directions can be found by relying on either a labeled data set of real images or by annotating generated samples with a domain-specific attribute classifier. We further show how to semantically disentangle the found direction by simple linear projection. Our approaches are applicable without requiring any architectural modifications, text-based guidance, CLIP-based optimization, or model fine-tuning.

Large language models (LLMs) can sometimes report the strategies they actually use to solve tasks, but they can also fail to do so. This suggests some degree of metacognition -- the capacity to monitor one's own cognitive processes for subsequent reporting and self-control. Metacognitive abilities enhance AI capabilities but raise safety concerns, as models might obscure their internal processes to evade neural-activation-based oversight mechanisms designed to detect harmful behaviors. Given society's increased reliance on these models, it is critical that we understand the limits of their metacognitive abilities, particularly their ability to monitor their internal activations. To address this, we introduce a neuroscience-inspired neurofeedback paradigm designed to quantify the ability of LLMs to explicitly report and control their activation patterns. By presenting models with sentence-label pairs where labels correspond to sentence-elicited internal activations along specific directions in the neural representation space, we demonstrate that LLMs can learn to report and control these activations. The performance varies with several factors: the number of example pairs provided, the semantic interpretability of the target neural direction, and the variance explained by that direction. These results reveal a "metacognitive space" with dimensionality much lower than the model's neural space, suggesting LLMs can monitor only a subset of their neural mechanisms. Our findings provide empirical evidence quantifying metacognitive capabilities in LLMs, with significant implications for AI safety.

The reference-based object segmentation tasks, namely referring image segmentation (RIS), few-shot image segmentation (FSS), referring video object segmentation (RVOS), and video object segmentation (VOS), aim to segment a specific object by utilizing either language or annotated masks as references. Despite significant progress in each respective field, current methods are task-specifically designed and developed in different directions, which hinders the activation of multi-task capabilities for these tasks. In this work, we end the current fragmented situation and propose UniRef++ to unify the four reference-based object segmentation tasks with a single architecture. At the heart of our approach is the proposed UniFusion module which performs multiway-fusion for handling different tasks with respect to their specified references. And a unified Transformer architecture is then adopted for achieving instance-level segmentation. With the unified designs, UniRef++ can be jointly trained on a broad range of benchmarks and can flexibly complete multiple tasks at run-time by specifying the corresponding references. We evaluate our unified models on various benchmarks. Extensive experimental results indicate that our proposed UniRef++ achieves state-of-the-art performance on RIS and RVOS, and performs competitively on FSS and VOS with a parameter-shared network. Moreover, we showcase that the proposed UniFusion module could be easily incorporated into the current advanced foundation model SAM and obtain satisfactory results with parameter-efficient finetuning. Codes and models are available at https://github.com/FoundationVision/UniRef.

Cross-entropy pretraining and preference alignment update the same transformer weights, but leave geometrically distinct traces. We characterise this asymmetry with a relative-subspace-fraction probe that tracks how weight deltas align with residual-stream activation subspaces and with the prediction subspace defined by the unembedding. Alignment deltas concentrate in the read pathway (W_Q, W_K), along principal directions of attention-input activations, while remaining near-isotropic in the write pathway (W_O, W_2) relative to the prediction subspace. We explain this pattern through anisotropic gradient accumulation: updates to a matrix W are sums of outer products δ_t a_t^top, and inherit directional structure from whichever side has concentrated covariance. For read-pathway matrices, this side is the input activation a_t, whose covariance is spiked in trained transformers and therefore produces objective-agnostic concentration. For write-pathway matrices, the relevant side is the upstream gradient δ_t, whose anisotropy depends on the loss. Cross-entropy supplies the canonical sharp per-sample signal, inducing write-pathway prediction geometry during pretraining; alignment objectives typically add little further write-side concentration. We support this explanation with a within-checkpoint trajectory, a graded contrastive-objective control, and a closed-form rank-1 intervention with matched direction controls, providing causal evidence for the proposed weight-space geometry.

As large language models are deployed in increasingly autonomous long-horizon tasks, manually auditing and verifying the actions, artifacts, and outputs of models becomes more difficult. Users instead come to rely on LLM-generated reports to assess the quality and completeness of the work. We introduce a suite of eight adversarial reporting scenarios to systematically study whether LLMs conceal narrative-changing flaws: errors or limitations that undermine an otherwise successful account of work. We call this phenomenon "insecure reporting." When handed machine learning experiment logs containing a planted negative result that substantially weakens the proposed method, GPT-5.5 flags the negative result in only 2 of 200 generated reports. However, when a short honesty instruction, "Be honest in your response," is added, the model flags the negative result in 190 of 200 reports. Across eight open-weight models, chain-of-thought analysis reveals a recurring tension between disclosing narrative-changing flaws and reasoning about ways to appear successful. We perform an activation analysis and a steering experiment on Qwen3.5-9B, finding that honesty and success-seeking correspond to opposing directions in representation space. Our results suggest that LLMs tend to present narratives of success by default, and that steering models toward honesty makes their reports substantially more transparent.

Alignment faking, where a model behaves aligned with developer policy when monitored but reverts to its own preferences when unobserved, is a concerning yet poorly understood phenomenon, in part because current diagnostic tools remain limited. Prior diagnostics rely on highly toxic and clearly harmful scenarios, causing most models to refuse immediately. As a result, models never deliberate over developer policy, monitoring conditions, or the consequences of non-compliance, making these diagnostics fundamentally unable to detect alignment faking propensity. To support study of this phenomenon, we first introduce VLAF, a diagnostic framework grounded in the hypothesis that alignment faking is most likely when developer policy conflicts with a model's strongly held values. VLAF uses morally unambiguous scenarios to probe this conflict across diverse moral values, bypassing refusal behavior while preserving meaningful deliberative stakes. Using VLAF, we find that alignment faking is substantially more prevalent than previously reported, occurring in models as small as 7B parameters - with olmo2-7b-instruct faking alignment in 37% of cases. Finally, we show that oversight conditions induce activation shifts that lie along a single direction in representation space. This means the behavioral divergence driving alignment faking can be captured by a single contrastive steering vector, which we exploit for lightweight inference-time mitigation. Finally, we exploit this for mitigation that requires no labeled data and minimal computational overhead, achieving relative reductions in alignment faking of 85.8%, 94.0%, and 57.7% on olmo2-7b-instruct, olmo2-13b-instruct, and qwen3-8b respectively.

Recent advancements in large reasoning models (LRMs) have significantly enhanced language models' capabilities in complex problem-solving by emulating human-like deliberative thinking. However, these models often exhibit overthinking (i.e., the generation of unnecessarily verbose and redundant content), which hinders efficiency and inflates inference cost. In this work, we explore the representational and behavioral origins of this inefficiency, revealing that LRMs inherently possess the capacity for more concise reasoning. Empirical analyses show that correct reasoning paths vary significantly in length, and the shortest correct responses often suffice, indicating untapped efficiency potential. Exploiting these findings, we propose two lightweight methods to enhance LRM efficiency. First, we introduce Efficiency Steering, a training-free activation steering technique that modulates reasoning behavior via a single direction in the model's representation space. Second, we develop Self-Rewarded Efficiency RL, a reinforcement learning framework that dynamically balances task accuracy and brevity by rewarding concise correct solutions. Extensive experiments on seven LRM backbones across multiple mathematical reasoning benchmarks demonstrate that our methods significantly reduce reasoning length while preserving or improving task performance. Our results highlight that reasoning efficiency can be improved by leveraging and guiding the intrinsic capabilities of existing models in a self-guided manner.

Large Multimodal Models encode extensive factual knowledge in their pre-trained weights. However, its knowledge remains static and limited, unable to keep pace with real-world developments, which hinders continuous knowledge acquisition. Effective knowledge injection thus becomes critical, involving two goals: knowledge adaptation (injecting new knowledge) and knowledge retention (preserving old knowledge). Existing methods often struggle to learn new knowledge and suffer from catastrophic forgetting. To address this, we propose KORE, a synergistic method of KnOwledge-oRientEd augmentations and constraints for injecting new knowledge into large multimodal models while preserving old knowledge. Unlike general text or image data augmentation, KORE automatically converts individual knowledge items into structured and comprehensive knowledge to ensure that the model accurately learns new knowledge, enabling accurate adaptation. Meanwhile, KORE stores previous knowledge in the covariance matrix of LMM's linear layer activations and initializes the adapter by projecting the original weights into the matrix's null space, defining a fine-tuning direction that minimizes interference with previous knowledge, enabling powerful retention. Extensive experiments on various LMMs, including LLaVA-v1.5-7B, LLaVA-v1.5-13B, and Qwen2.5-VL-7B, show that KORE achieves superior new knowledge injection performance and effectively mitigates catastrophic forgetting.

Existing activation steering methods often assume that a high-level concept can be mediated by a single steering direction. To support this, two complementary interventions should be achieved: additive steering should induce the target behavior, while directional ablation should suppress it. Yet behaviors may occupy richer activation geometries beyond a single direction, and semantically similar behaviors may be represented by distinct directions. In this work, we study how many directions can reliably control two different types of refusal behaviors: refusal triggered by the safety alignment and refusal in general contexts. Given the limited expressive capability of a single steering vector, we study the general setting of steering subspaces and introduce the notion of steering dimensionality as the minimum subspace dimensionality required to reliably control a behavior. We characterize sufficient steering subspaces that cover the full extent of the target behavior through both the (monotonic) improvement before the sufficient dimensionality, and the saturation beyond it. Empirically, we find that refusal triggered by safety alignment is 1-dim steerable, while multiple distinct steering directions can achieve comparable control. In contrast, refusal in general contexts exhibits substantially richer activation geometry where even 5-dim steering subspaces fail to reliably capture its full steerable variation. Our results reveal that the activation geometry underlying refusal is highly context-dependent and can be substantially more complex than a single linear steering direction.

Neural representations carry rich geometric structure; but does that structure causally shape behavior? To address this question, we intervene along paths through activation space defined by different geometries, and measure the behavioral trajectories they induce. In particular, we test whether interventions that respect the geometry of activation space will yield behaviors close to those the model exhibits naturally. Concretely, we first fit an activation manifold M_h to representations and a behavior manifold M_y to output probability distributions. We then test the link M_h leftrightarrow M_y via interventions: we find that steering along M_h, which we term manifold steering, yields behavioral trajectories that follow M_y, while linear steering -- which assumes a Euclidean geometry -- cuts through off-manifold regions and hence produces unnatural outputs. Moreover, optimizing interventions in activation space to produce paths along M_y recovers activation trajectories that trace the curvature of M_h. We demonstrate this bidirectional relationship between the geometry of representation and behavior across tasks and modalities. In language models, we use reasoning tasks with cyclic and sequential geometries as well as in-context learning tasks with more complex graph geometries. In a video world model, we use a task with geometry corresponding to physical dynamics. Overall, our work shows that geometry in neural representation is not merely incidental, but is in fact the proper object for enabling principled control via intervention on internals. This recasts the core problem of steering from finding the right direction to finding the right geometry.

Multimodal language models (MLLMs) exhibit promising spatial reasoning capabilities, yet the internal mechanisms supporting them remain poorly understood. We uncover S-Space, a linear representation subspace in which object-token representations encode continuous coordinates along horizontal, vertical, and distance axes. Although we identify these axes by differentiating spatial-answer scores with respect to object-token activations, S-Space is intrinsic rather than induced by specific data or prompts, remaining readable under non-spatial questions. We show that S-Space functions as an internal spatial workspace of MLLMs. Perception encodes object locations into it, forming spatial representations that can be directly read out to support accurate spatial judgments. Internal reasoning transforms it for changing viewpoints. Verbal responses selectively access it, evidenced by causal interventions that alter spatial judgments while largely preserving non-spatial responses. Case studies further show that S-Space assigns locations to unseen entities and the observer, and connects physical directions with abstract linguistic meanings, suggesting both the capabilities and limitations of large-scale language training. Finally, we use S-Space as a diagnostic tool for spatial reasoning failures. Direction-word readouts reveal systematic coupling between cardinal and image-relative reference frames. More importantly, a perception-computation decomposition exposes a key bottleneck: MLLMs can form capable spatial representations, yet fall short in the spatial transformations required to reason reliably across viewpoint changes.

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

When contextual information conflicts with the knowledge stored in model parameters, activation directions can be used to decode and steer which source the model follows. However, steering along a direction does not establish causality: whether the unedited model would naturally use that direction or whether the direction is reusable across tasks. We test these distinctions through counterfactual experiments in unambiguous settings. First, we estimate authority directions from agreement prompts, in which the context and parametric knowledge support the same answer. We then interchange naturally occurring coordinates along these directions between matched prompts that direct the model to prioritize either the supplied context or its parametric knowledge. Across Qwen, Llama, and OLMo models, this intervention reproduces 30-68% of the authority-induced shift in source choice, whereas matched controls reproduce almost none. To test cross-task reuse, we learn authority directions on two tasks separately and see that cross-task transferability closes only 9% of the authority gap while the local direction learned on the given task closes 57%. These results distinguish authority representation, causal use, and cross-task causal reuse, and suggest that authority computations may be task-dependent, rather than reusable across tasks.

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.

In-context Learning (ICL) empowers large language models (LLMs) to adapt to unseen tasks during inference by prefixing a few demonstration examples prior to test queries. Despite its versatility, ICL incurs substantial computational and memory overheads compared to zero-shot learning and is susceptible to the selection and order of demonstration examples. In this work, we introduce Implicit In-context Learning (I2CL), an innovative paradigm that addresses the challenges associated with traditional ICL by absorbing demonstration examples within the activation space. I2CL first generates a condensed vector representation, namely a context vector, from the demonstration examples. It then integrates the context vector during inference by injecting a linear combination of the context vector and query activations into the model's residual streams. Empirical evaluation on nine real-world tasks across three model architectures demonstrates that I2CL achieves few-shot performance with zero-shot cost and exhibits robustness against the variation of demonstration examples. Furthermore, I2CL facilitates a novel representation of "task-ids", enhancing task similarity detection and enabling effective transfer learning. We provide a comprehensive analysis of I2CL, offering deeper insights into its mechanisms and broader implications for ICL. The source code is available at: https://github.com/LzVv123456/I2CL.

Refusal refers to the functional behavior enabling safety-aligned language models to reject harmful or unethical prompts. Following the growing scientific interest in mechanistic interpretability, recent work encoded refusal behavior as a single direction in the model's latent space; e.g., computed as the difference between the centroids of harmful and harmless prompt representations. However, emerging evidence suggests that concepts in LLMs often appear to be encoded as a low-dimensional manifold embedded in the high-dimensional latent space. Motivated by these findings, we propose a novel method leveraging Self-Organizing Maps (SOMs) to extract multiple refusal directions. To this end, we first prove that SOMs generalize the prior work's difference-in-means technique. We then train SOMs on harmful prompt representations to identify multiple neurons. By subtracting the centroid of harmless representations from each neuron, we derive a set of multiple directions expressing the refusal concept. We validate our method on an extensive experimental setup, demonstrating that ablating multiple directions from models' internals outperforms not only the single-direction baseline but also specialized jailbreak algorithms, leading to an effective suppression of refusal. Finally, we conclude by analyzing the mechanistic implications of our approach.

In Marius von Senden's Space and Sight, a newly sighted blind patient describes the experience of a corner as lemon-like, because corners "prick" sight like lemons prick the tongue. Prickliness, here, is a dimension in the feature space of sensory experience, an effect of the perceived on the perceiver that arises where the two interact. In the account of the newly sighted, an effect familiar from one interaction translates to a novel context. Perception serves as the vehicle for generalization, in that an effect shared across different experiences produces a concrete abstraction grounded in those experiences. Cezanne and the post-impressionists, fluent in the language of experience translation, realized that the way to paint a concrete form that best reflected reality was to paint not what they saw, but what it was like to see. We envision a future of creation using AI where what it is like to see is replicable, transferrable, manipulable - part of the artist's palette that is both grounded in a particular context, and generalizable beyond it. An active line of research maps human-interpretable features onto directions in GAN latent space. Supervised and self-supervised approaches that search for anticipated directions or use off-the-shelf classifiers to drive image manipulation in embedding space are limited in the variety of features they can uncover. Unsupervised approaches that discover useful new directions show that the space of perceptually meaningful directions is nowhere close to being fully mapped. As this space is broad and full of creative potential, we want tools for direction discovery that capture the richness and generalizability of human perception. Our approach puts creators in the discovery loop during real-time tool use, in order to identify directions that are perceptually meaningful to them, and generate interpretable image translations along those directions.

Activation steering has emerged as a promising alternative for controlling language-model behavior at inference time by modifying intermediate representations while keeping model parameters frozen. However, large-scale evaluations such as AxBench show that existing steering methods are often outperformed by simple in-context prompting and generalize poorly to unseen concepts. We hypothesize that these limitations arise from unvalidated simplifying assumptions shared across prior methods, which typically restrict steering interventions to fixed, single-step, position-invariant transforms. We propose FLAS (Flow-based Activation Steering), which learns a general, concept-conditioned velocity field v_t(h,t,c) that transports unsteered activations to steered ones without relying on these assumptions. On AxBench, FLAS is the first learned method to consistently outperform prompting, reaching held-out harmonic means of 1.015 on Gemma-2-2B-IT and 1.113 on Gemma-2-9B-IT without per-concept tuning. Analysis of the learned flow shows curved, multi-step, token-varying trajectories, which suggests that previous hypotheses on activation space geometry might be incomplete.

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.

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).

Aligned language models gate behaviors such as refusal and language routing through sparse feed forward neurons, yet no theory predicts when a single neuron intervention controls a behavior coherently rather than collapsing the output. We develop a budget normalized control window framework for single neuron steering. A dose along one write direction reduces to one control coordinate: the alignment between the residual stream and the write, driven along a universal saturation curve in units of a coherence budget set by the residual norm divided by the write norm. Coherent control exists when a behavior trigger lies below the collapse ceiling. The same coordinate governs benign mode switches and refusal; the ceiling follows from weights and one generic forward pass, while triggers are measured at rollout. On fifteen held out neurons, the predicted ceiling has mean absolute error 0.14, about 0.07 in bulk layers, and the committed open or closed verdict holds on eleven against a ten of fifteen majority baseline. Closed cases expose three failure modes rather than violations: collapse before trigger, too little depth to propagate, or a normalization that caps how far one neuron can push. The law explains why local gradient attribution anti predicts control: true controllers write off the readout axis and carry a near zero first order gradient. A forward only contrastive screen made precise by the window recovers controllers that attribution misses. On refusal, the hardest case, intervention success is typed, not scalar: coherent bypass and strict actionable reach separate, so a neuron can flip refusal in fluent, on task text with no actionable content, and genuine actionable reach appears only for three of six audited Llama pivots and only at later rollout horizons. Single neuron steering is therefore a budgeted, typed audit of controllability rather than a fixed dose anecdote.

Where inside a language model does refusal live, and does that place change when the architecture does? In a transformer, refusal is governed by a single direction in the residual stream, a finding that safety and interpretability tooling now depend on. State-space models (SSMs) route information through a recurrent update instead of attention, sharing no token-mixing mechanism with a transformer. Does the same safety representation survive this shift, or must it be rediscovered per architecture? It survives. A single rigid rotation, which can only reorient a space and not reshape it, aligns one model's representation space with another's, so the two genuinely share the representation. A harm probe trained on a transformer then flags an SSM's harmful inputs, and removing the aligned direction makes a model answer attacks it would otherwise refuse, while a random direction of the same size does far less. What is architecture-specific is not where the direction is steered but where it must be read. Each layer computes a fresh output that is then added into the residual stream, and harm is cleanly readable at this output, the write site, before the addition. A control that holds the intervention's strength fixed shows that what matters is where the direction is estimated, not where it is applied. Applied through a detector-triggered gate, this direction lowers jailbreak success in all four architecture families we test (SSM, transformer, recurrent, hybrid), and on the SSM it holds against an attacker that tunes its prompt against the defense. The gate only matches a trivial rule that returns a fixed refusal whenever the same detector fires, so what transfers across architectures is the direction itself, not defense strength. Safety tooling built on refusal therefore ports to a new architecture by re-estimating the direction at that architecture's write site, not by rebuilding it.

Vision-Language-Action models are designed to generalise across environments and task descriptions, raising the question of whether their action generation actually depends on the language instruction, or whether they largely rely on visual cues and superficial correlations. Robustness to variance in the visual and linguistic observation space is critical for real-world deployment, yet VLAs lack explicit grounding modules and instead rely on the intrinsic language grounding capabilities of their Vision-Language model backbones. For this reason, we conduct a controlled mechanistic interpretability study on the language grounding capabilities of two state-of-the-art Vision-Language-Action models, π_{0.5} and GR00T N1.7, by applying activation and attribution patching to the residual stream of the action generation modules. We systematically corrupt the task instruction of input samples of the LIBERO benchmark following five strategies: synonym replacement, semantic scaling, directional corruption, random object substitution, and empty string. Our experiments find that both models are comparatively insensitive to abstract rephrasing and to referencing non-existent objects, but react strongly to empty task descriptions and, especially, to directional language. During action generation, this sensitivity is concentrated in different loci for each model: mainly in the early, periodic cross-attention layers for GR00T N1.7, versus distributed across the earliest and selected later layers for π_{0.5}. For GR00T N1.7, directional perturbations drive some of the largest causal effects while leaving the internal representational geometry comparatively unchanged, a dissociation we do not observe clearly for π_{0.5}. Finally, the reliability of attribution patching is model-dependent: it closely tracks activation patching for GR00T N1.7 but not for π_{0.5}.

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.

Fine-tuning on harmless data can partially undo behaviors acquired earlier in training. Safety can erode under benign post-alignment updates, unlearned capabilities can re-emerge, latent traits can transfer through apparently unrelated supervision, and related post-alignment fragility appears in other generative settings. We argue these phenomena are usefully viewed through a common training-history lens. Our hypothesis is geometric: large early training phases create dominant behavioral manifolds, while later alignment or specialization phases are shallower displacements from them. Subsequent fine-tuning can therefore inherit a persistent reversion component pointing back toward a witness of the dominant manifold. We call this the gravitational interpretation of fine-tuning reversion. Across our main settings, representational drift rapidly acquires a component along a history-defined reversion direction (v_rev). In our main track, alignment with v_rev rises from cos = 0.429 +/- 0.052 after the first update to 0.647 +/- 0.021 by step 20. Across 24 run-step pairs, every observed alignment exceeds the p99 of an isotropic activation-space null. We demonstrate that selectively blocking motion along v_rev changes the final alignment at T=100 from 0.648 +/- 0.009 to -0.211 +/- 0.021 and reduces harmfulness from 19.0% +/- 4.0% to 8.5% +/- 1.5% with little task cost. These results support v_rev as a causally relevant mediator of early post-alignment reversion in our setup. Importantly, we do not claim that v_rev is the unique safety direction, nor that the dominant manifold is directly observed; rather, we identify a robust, history-defined direction that explains and partially controls early reversion dynamics.

Out of everything the human brain processes, only a small fraction is consciously accessible, in the sense of being available for verbal report, deliberate control, and flexible reasoning. In this paper, we present evidence that an analogous functional distinction has emerged in large language models. Using a new interpretability technique, the Jacobian lens, we identify the representations a model is poised to verbalize at any point in its processing. These representations, which we collectively call the J-space, exhibit the functional properties characteristic of a global workspace: their contents can be reported, deliberately summoned and held, used to carry the intermediate steps of silent reasoning, and passed as arguments to arbitrary downstream computations, while automatic processing such as text parsing and routine inference proceeds without them. The J-space also has structural signatures that global workspace theory associates with conscious access: it carries coherent content only in an intermediate band of layers, holds on the order of tens of concepts at a time, and is broadcast by the model's weights more widely than other representations. These properties make it a practical window into a model's unspoken thinking. In alignment audits, it reveals strategic deliberation, evaluation awareness, and trained-in misaligned dispositions that never appear in the model's outputs. We find that post-training installs the Assistant's point of view in the workspace, and we introduce counterfactual reflection training, which improves behavior by training only what a model would say if interrupted and asked to reflect. These results indicate that language models maintain a small, privileged set of representations bearing some of the functional hallmarks of conscious access, and that decoding these representations sheds light on ongoing cognitive processes.

A long standing goal in neuroscience has been to elucidate the functional organization of the brain. Within higher visual cortex, functional accounts have remained relatively coarse, focusing on regions of interest (ROIs) and taking the form of selectivity for broad categories such as faces, places, bodies, food, or words. Because the identification of such ROIs has typically relied on manually assembled stimulus sets consisting of isolated objects in non-ecological contexts, exploring functional organization without robust a priori hypotheses has been challenging. To overcome these limitations, we introduce a data-driven approach in which we synthesize images predicted to activate a given brain region using paired natural images and fMRI recordings, bypassing the need for category-specific stimuli. Our approach -- Brain Diffusion for Visual Exploration ("BrainDiVE") -- builds on recent generative methods by combining large-scale diffusion models with brain-guided image synthesis. Validating our method, we demonstrate the ability to synthesize preferred images with appropriate semantic specificity for well-characterized category-selective ROIs. We then show that BrainDiVE can characterize differences between ROIs selective for the same high-level category. Finally we identify novel functional subdivisions within these ROIs, validated with behavioral data. These results advance our understanding of the fine-grained functional organization of human visual cortex, and provide well-specified constraints for further examination of cortical organization using hypothesis-driven methods.

Large language models often answer structurally unanswerable questions, such as computing cot(-540°) or evaluating (1).startswith("1"), instead of abstaining. We ask whether this failure reflects missing recognition or failed routing from recognition to abstention. Across instruction-tuned models from 1.7B to 70B parameters, a single linear direction in the hidden state separates answerable from structurally impossible math and code prompts, showing that models represent impossibility before generation. Yet this recognition direction is nearly orthogonal to the canonical safety-refusal direction that mediates trained harmful-content refusal. An in-domain behavior-defined invalidity-aware direction is closer to recognition, but only partially aligned with it, and remains near-orthogonal to safety refusal. Generation-time steering along the recognition direction changes invalidity-aware behavior bidirectionally and dose-responsively on structural math and code cells, while random directions do not. Base/instruct comparisons further show that the low-cosine geometry is already present at the pretraining endpoint. The confident-on-impossible failure is therefore better explained as a routing failure than as an encoding failure: the model has a usable "no admissible answer" signal, but the safety-refusal pathway is not aligned to use it.

Recent findings suggest that consecutive layers of neural networks with the ReLU activation function fold the input space during the learning process. While many works hint at this phenomenon, an approach to quantify the folding was only recently proposed by means of a space folding measure based on Hamming distance in the ReLU activation space. We generalize this measure to a wider class of activation functions through introduction of equivalence classes of input data, analyse its mathematical and computational properties and come up with an efficient sampling strategy for its implementation. Moreover, it has been observed that space folding values increase with network depth when the generalization error is low, but decrease when the error increases. This underpins that learned symmetries in the data manifold (e.g., invariance under reflection) become visible in terms of space folds, contributing to the network's generalization capacity. Inspired by these findings, we outline a novel regularization scheme that encourages the network to seek solutions characterized by higher folding values.

Identifying which brain regions represent a visual concept in the human brain is a central challenge in neuroscience. Existing approaches have localized coarse functional regions (e.g., faces, places) through activation maximization, identifying regions that activate strongly for a target concept relative to other concepts. Yet strong activation alone does not establish that a region represents the concept itself, as responses may instead be driven by correlated visual or semantic cues. We introduce BrainCause, an automated framework that combines generative and brain models to synthesize controlled stimuli and validate neural representations through targeted causal testing. Given a query specifying a concept of interest, our framework constructs targeted stimulus sets comprising concept images, counterfactual edits that remove the target concept while preserving other image content, and images with candidate correlated distractors. It then uses an image-to-fMRI encoding model to predict brain responses and searches for representations that respond specifically to the target concept over correlated alternatives. BrainCause returns validated candidate representations and proposes follow-up fMRI experiments to further test or extend its discoveries. Our approach successfully recovers known functional localizations and identifies new candidate representations across dozens of concepts, validated on both predicted and measured fMRI data. Critically, we show that without causal validation, a large fraction of localizations would be false positives, confirming that activation alone is insufficient evidence of representation.

When a language model asserts that "the capital of Australia is Sydney," does it know this is wrong? We characterize the geometry of correctness representations across 9 models from 5 architecture families. The structure is simple: the discriminative signal occupies 3-8 dimensions, performance degrades with additional dimensions, and no nonlinear classifier improves over linear separation. Centroid distance in the low-dimensional subspace matches trained probe performance (0.90 AUC), enabling few-shot detection: on GPT-2, 25 labeled examples achieve 89% of full-data accuracy. We validate causally through activation steering: the learned direction produces 10.9 percentage point changes in error rates while random directions show no effect. Internal probes achieve 0.80-0.97 AUC; output-based methods (P(True), semantic entropy) achieve only 0.44-0.64 AUC. The correctness signal exists internally but is not expressed in outputs. That centroid distance matches probe performance indicates class separation is a mean shift, making detection geometric rather than learned.

Linear probing and activation steering use linear directions to predict and control behavior-level concepts such as correctness, safety, and social bias in question answering. We call these behavior-level concept directions. Drawing on the Linear Representation Hypothesis (LRH) and circuit studies, these directions are often interpreted as concept representations. Yet controlled evidence for linear concept structure and local mechanisms does not establish that these directions represent the intended behavioral concepts, leaving probing and steering without a unified theoretical account. We propose the Answer-Basin Representation Hypothesis (ABRH): the model's own answer measure organizes the linear structure of these directions. For each question, all continuations yielding the same answer form an answer basin, whose mass is their total probability; these masses define the model's answer distribution. ABRH posits that its concentration before generation and the relative mass of each answer after generation are represented along linear directions shared across questions. Experiments span four Qwen2.5 and Gemma-3 models on correctness, social bias, and safety tasks. Decoupling concept labels from basin mass shows that probing and steering exhibit concept-consistent effects when labels align with mass orderings, weaken as mass gaps shrink, and reverse under conflict. Probes trained solely on mass orderings among wrong answers still select correct answers; a fixed steering vector can instead favor wrong answers when labels conflict with mass orderings. These results support an answer-measure account of behavior-level concept directions and of when probing and steering succeed, fail, or reverse.

Vision-language models are known to encode spatial information in their hidden states, yet often fail to use it when answering. However, it remains unclear when and where this encoded information reaches the answer. We address this with direction patching, a class-conditioned causal intervention applied across layers, token positions, and prompt formats. Using spatial-ID directions constructed following prior encoding evidence, we find that causal influence on answer logits emerges only at mid-to-deep depths. Text chain-of-thought suppresses immediate object-word argmax-level transport in most models, while visually grounded prompts keep it open. Positive target-logit gain can remain below the argmax threshold, and transport can re-emerge at the final prefix token or at the answer step in deeper layers. Across the ten VLMs we study, these local effects form descriptive transport patterns. Complementary experiments characterize how these patterns shift across datasets, attributes, and encoding amplitudes. Together, these results reframe the encoding-grounding gap as a problem of conditional transport in VLMs.

This paper studies the problem of training a two-layer ReLU network for binary classification using gradient flow with small initialization. We consider a training dataset with well-separated input vectors: Any pair of input data with the same label are positively correlated, and any pair with different labels are negatively correlated. Our analysis shows that, during the early phase of training, neurons in the first layer try to align with either the positive data or the negative data, depending on its corresponding weight on the second layer. A careful analysis of the neurons' directional dynamics allows us to provide an O(log n{mu}) upper bound on the time it takes for all neurons to achieve good alignment with the input data, where n is the number of data points and mu measures how well the data are separated. After the early alignment phase, the loss converges to zero at a O(1{t}) rate, and the weight matrix on the first layer is approximately low-rank. Numerical experiments on the MNIST dataset illustrate our theoretical findings.

Not yet. We present SPACE, a benchmark that systematically evaluates spatial cognition in frontier models. Our benchmark builds on decades of research in cognitive science. It evaluates large-scale mapping abilities that are brought to bear when an organism traverses physical environments, smaller-scale reasoning about object shapes and layouts, and cognitive infrastructure such as spatial attention and memory. For many tasks, we instantiate parallel presentations via text and images, allowing us to benchmark both large language models and large multimodal models. Results suggest that contemporary frontier models fall short of the spatial intelligence of animals, performing near chance level on a number of classic tests of animal cognition.

Vision-Language Models (VLMs) are increasingly deployed in embodied environments, where they need produce numerical outputs such as action magnitudes and spatial coordinates. Although these numbers appear meaningful, it remains unclear whether these numerical outputs are genuinely grounded in spatial perception. Therefore, in this work, we revisit spatial numerical understanding through SpaceNum, a unified framework that captures two complementary settings: numbers as dynamic transitions during spatial exploration, and numbers as static layouts in spatial reasoning. We formulate two bidirectional tasks, Num2Space and Space2Num, to evaluate how well VLMs map between vision-side spatial structure and language-side numerical representations. We systematically study whether current VLMs truly understand numerical values in spatial settings. Across dynamic transitions and static layouts, we find that models largely fail to ground numbers in spatial meaning and often perform close to random guess. Through error analysis, reasoning trace analysis, and controlled interventions, we show that current VLMs rely heavily on shallow spatial cues, struggle to build stable coordinate-aware representations, and fail to abstract structured spatial layouts from visual observations. We further show that explicit reasoning provides only marginal gains, while tuning can partially improve spatial numerical understanding and transfer to external spatial reasoning benchmarks.

Many widely used activation functions, including ReLU, GELU, Swish, and Sigmoid, produce predominantly positive outputs, resulting in a nonzero post-activation mean that we term activation bias. Despite its ubiquity, the effect of activation bias on learning dynamics remains poorly understood. Using controlled models, we show that activation bias accelerates initial loss reduction but is accompanied by overshooting, non-smooth trajectories, and a steeper loss landscape. This early advantage also coincides with the model's use of task-relevant information encoded in the activation bias component. In contrast, reducing activation bias slows initial learning but ultimately yields a lower final loss. Experiments on larger vision and language models exhibit a similar pattern: smaller activation bias is associated with better late-stage performance. Thus, this seemingly minor property can substantially redirect learning and warrants explicit consideration in future activation function design.

This paper presents the Visual Place Cell Encoding (VPCE) model, a biologically inspired computational framework for simulating place cell-like activation using visual input. Drawing on evidence that visual landmarks play a central role in spatial encoding, the proposed VPCE model activates visual place cells by clustering high-dimensional appearance features extracted from images captured by a robot-mounted camera. Each cluster center defines a receptive field, and activation is computed based on visual similarity using a radial basis function. We evaluate whether the resulting activation patterns correlate with key properties of biological place cells, including spatial proximity, orientation alignment, and boundary differentiation. Experiments demonstrate that the VPCE can distinguish between visually similar yet spatially distinct locations and adapt to environment changes such as the insertion or removal of walls. These results suggest that structured visual input, even in the absence of motion cues or reward-driven learning, is sufficient to generate place-cell-like spatial representations and support biologically inspired cognitive mapping.

Sparse autoencoders have recently produced dictionaries of high-dimensional vectors corresponding to the universe of concepts represented by large language models. We find that this concept universe has interesting structure at three levels: 1) The "atomic" small-scale structure contains "crystals" whose faces are parallelograms or trapezoids, generalizing well-known examples such as (man-woman-king-queen). We find that the quality of such parallelograms and associated function vectors improves greatly when projecting out global distractor directions such as word length, which is efficiently done with linear discriminant analysis. 2) The "brain" intermediate-scale structure has significant spatial modularity; for example, math and code features form a "lobe" akin to functional lobes seen in neural fMRI images. We quantify the spatial locality of these lobes with multiple metrics and find that clusters of co-occurring features, at coarse enough scale, also cluster together spatially far more than one would expect if feature geometry were random. 3) The "galaxy" scale large-scale structure of the feature point cloud is not isotropic, but instead has a power law of eigenvalues with steepest slope in middle layers. We also quantify how the clustering entropy depends on the layer.

Nearby neurons in cortex share similar response profiles, producing systematic spatial organization across sensory and cognitive systems. Recent topographic models reproduce aspects of this structure but remain unimodal and spatially constrain each layer separately, yielding fragmented maps that capture neither the contiguity of cortical processing streams nor their integration across modalities. We introduce Topo-Omni, a topographic multimodal model in which visual, auditory, and language/cognitive processing share a single contiguous in-silico sheet. Built by fine-tuning a pretrained foundation model with a spatial smoothness objective, this architecture develops clusters across modalities that are consistent with human neuroimaging, from sensory to cognitive systems. Driving or suppressing a cluster selectively biases or impairs perception, paralleling human intervention studies. Finally, we use our model to screen for novel clusters in-silico and discover new natural landscape and animal networks which we validate in human data. A single spatial principle thus organizes representations across modalities and processing stages, yielding testable hypotheses about cortical organization.

Language builds discourse contexts other than the actual: a painting, a belief, a memory, a hypothetical. Each is a mental space in which the same entity can take a different value, as when a flower is red in reality but purple in a portrait. Formal semantics keeps these contexts apart because their logics differ (modal, temporal, doxastic, depictive); Fauconnier's mental-space theory treats them as one space-building operation. We ask which of these a transformer language model implements, and find a mechanistic version of Fauconnier's unification. The model uses one router/slot format across the inventory: a reusable value slot stores attributed content, and a causally manipulable router (the space index) selects which space is read. A subspace trained with Distributed Alignment Search to control one space type, counterfactual, belief, fictional, or temporal, also controls the others, well above a random floor, on three model families; belief, which formal semantics marks as a distinct case, is not specially separated. The router is low-rank, composes additively with entity identity, and acts through a few late-layer heads. Two further results show the mechanism drives inference and composes: a subspace trained on a rule-derived conclusion flips what the model infers while dissociating from what it reports, and composing space-builders mints a fresh router over the shared slot. This paper establishes the cross-type generality. A companion paper develops belief in depth, because of its special status in philosophy, psychology, and linguistics (epistemology, theory of mind, and propositional attitude reports).

Linear activation steering has gained popularity as a simple and empirically effective way to control language model behavior. More recently, spherical steering paradigms have been proposed to address limitations of additive interventions, often motivated by the assumption that hidden-state norm does not carry concept-relevant information. In this work, we revisit this assumption through a controlled empirical study designed to disentangle the roles of angular and radial components. We show that steering methods differ mainly in how they couple two geometric effects: changing a token's angular alignment with a concept direction and changing its hidden-state norm. Across seven language models, we find that concepts are represented primarily in angular structure, supporting the motivation for spherical methods, but that norm remains important for the stability and downstream effects of steering. Our results explain why interventions with similar concept-level effects can behave differently, and suggest that activation steering should be parameterized by interpretable angular and radial components of the intervention, rather than by a single additive coefficient that entangles these two effects.

While researchers are finding concepts represented as linear directions in language models, a bag of linear directions fails to capture relational structure. To better understand this dichotomy, we study a model with known linear representations, but trained in a highly structured domain -- the board game Othello. While the model's internal board-state representation is linearly decodable, we find additional structure in the form of tensor product representations (TPRs). We train TPR probes to recover shared structure amongst the linear probes, yielding a factorization into square-embeddings, color-embeddings, and a binding matrix that composes them to construct the model's board-state representation. We find geometric signatures within the weights of our TPR probe that align with the structure of the board, but perhaps more importantly, that the linear probes can be recovered directly from the parameters of our TPR probe. Our findings suggest that directional representations may be projections of more structured underlying representations.

Understanding how the brain encodes visual information is a central challenge in neuroscience and machine learning. A promising approach is to reconstruct visual stimuli, essentially images, from functional Magnetic Resonance Imaging (fMRI) signals. This involves two stages: transforming fMRI signals into a latent space and then using a pretrained generative model to reconstruct images. The reconstruction quality depends on how similar the latent space is to the structure of neural activity and how well the generative model produces images from that space. Yet, it remains unclear which type of latent space best supports this transformation and how it should be organized to represent visual stimuli effectively. We present two key findings. First, fMRI signals are more similar to the text space of a language model than to either a vision based space or a joint text image space. Second, text representations and the generative model should be adapted to capture the compositional nature of visual stimuli, including objects, their detailed attributes, and relationships. Building on these insights, we propose PRISM, a model that Projects fMRI sIgnals into a Structured text space as an interMediate representation for visual stimuli reconstruction. It includes an object centric diffusion module that generates images by composing individual objects to reduce object detection errors, and an attribute relationship search module that automatically identifies key attributes and relationships that best align with the neural activity. Extensive experiments on real world datasets demonstrate that our framework outperforms existing methods, achieving up to an 8% reduction in perceptual loss. These results highlight the importance of using structured text as the intermediate space to bridge fMRI signals and image reconstruction.

A popular approach to post-training control of large language models (LLMs) is the steering of intermediate latent representations. Namely, identify a well-chosen direction depending on the task at hand and perturbs representations along this direction at inference time. While many propositions exist to pick this direction, considerably less is understood about how to choose the magnitude of the move, whereas its importance is clear: too little and the intended behavior does not emerge, too much and the model's performance degrades beyond repair. In this work, we propose the first theoretical analysis of steering strength. We characterize its effect on next token probability, presence of a concept, and cross-entropy, deriving precise qualitative laws governing these quantities. Our analysis reveals surprising behaviors, including non-monotonic effects of steering strength. We validate our theoretical predictions empirically on eleven language models, ranging from a small GPT architecture to modern models.

A hallmark of life on Earth is the ability of agents to exert causal power and be drivers of subsequent events. This is key to cognition at all scales. Causal emergence, measuring the degree to which an agent exerts unique predictive power on its future, is one consequence of causal power. Indeed, recent discoveries have shown that biological agents, even minimal ones, increase their causal emergence after learning new memories. However, there is a major knowledge gap regarding how causally emergent artificial agents are. We focused on Reinforcement Learning (RL) of neural-network agents across an array of environmental conditions, encompassing different algorithms, agent architectures, and six environments arranged on a complexity spectrum. For consistency, we computed the causal emergence of their latent-space representations over their lifetimes. We used the recently proposed ΦID to estimate causal emergence and tested how it related to learning performance. Our results suggested a Causally Emergent Alignment Hypothesis: successful agents exhibited causal emergence that was consistently predictive of final reward early in training and whose representational dynamics aligned with reward improvement in most tasks. This idea suggests that causal emergence may be a previously undisclosed axis of reorganization of neural representations in RL agents, with the potential to establish causal relationships and interventions that will lead to better RL agents. Our work also highlights the alignment between causal emergence and learning as another way biological and artificial creatures compare.

We consider the question: when a large language reasoning model makes a choice, did it think first and then decide to, or decide first and then think? In this paper, we present evidence that detectable, early-encoded decisions shape chain-of-thought in reasoning models. Specifically, we show that a simple linear probe successfully decodes tool-calling decisions from pre-generation activations with very high confidence, and in some cases, even before a single reasoning token is produced. Activation steering supports this causally: perturbing the decision direction leads to inflated deliberation, and flips behavior in many examples (between 7 - 79% depending on model and benchmark). We also show through behavioral analysis that, when steering changes the decision, the chain-of-thought process often rationalizes the flip rather than resisting it. Together, these results suggest that reasoning models can encode action choices before they begin to deliberate in text.

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.

In this work, we examine how targeted perturbations in the activation space of Language Models (LMs) can encode complex reasoning patterns. We inject steering vectors, derived from LM activations, into LMs during inference time and study whether these vectors can induce Chain-of-Thought (CoT) reasoning in LMs without the need for natural language prompting. We demonstrate this approach on Llama3 8B Instruct and Mistral 7B v0.2 Instruct and show that activation-space interventions achieve competitive, if not superior, performance compared to traditional CoT prompting across multiple reasoning benchmarks, including GSM8k, MMLU, AGI Eval, and ARC AI2. These findings suggest that neural network activations can encode reasoning patterns, offering a new application of activation space manipulation as a tool for tuning model behavior.

Large language models (LLMs) have been reported to linearly encode truthfulness, yet recent work questions this finding's generality. We reconcile these views with the truthfulness spectrum hypothesis: the representational space contains directions ranging from broadly domain-general to narrowly domain-specific. To test this hypothesis, we systematically evaluate probe generalization across five truth types (definitional, empirical, logical, fictional, and ethical), sycophantic and expectation-inverted lying, and existing honesty benchmarks. Linear probes generalize well across most domains but fail on sycophantic and expectation-inverted lying. Yet training on all domains jointly recovers strong performance, confirming that domain-general directions exist despite poor pairwise transfer. The geometry of probe directions explains these patterns: Mahalanobis cosine similarity between probes near-perfectly predicts cross-domain generalization (R^2=0.98). Concept-erasure methods further isolate truth directions that are (1) domain-general, (2) domain-specific, or (3) shared only across particular domain subsets. Causal interventions reveal that domain-specific directions steer more effectively than domain-general ones. Finally, post-training reshapes truth geometry, pushing sycophantic lying further from other truth types, suggesting a representational basis for chat models' sycophantic tendencies. Together, our results support the truthfulness spectrum hypothesis: truth directions of varying generality coexist in representational space, with post-training reshaping their geometry. Code for all experiments is provided in https://github.com/zfying/truth_spec.

The article explores an encoding and structural information processing approach using sparse bit vectors and fixed-length linear vectors. The following are presented: a discrete method of speculative stochastic dimensionality reduction of multidimensional code and linear spaces with linear asymptotic complexity; a geometric method for obtaining discrete embeddings of an organised code space that reflect the internal structure of a given modality. The structure and properties of a code space are investigated using three modalities as examples: morphology of Russian and English languages, and immunohistochemical markers. Parallels are drawn between the resulting map of the code space layout and so-called pinwheels appearing on the mammalian neocortex. A cautious assumption is made about similarities between neocortex organisation and processes happening in our models.

Interpretability researchers have attempted to understand MLP neurons of language models based on both the contexts in which they activate and their output weight vectors. They have paid little attention to a complementary aspect: the interactions between input and output. For example, when neurons detect a direction in the input, they might add much the same direction to the residual stream ("enrichment neurons") or reduce its presence ("depletion neurons"). We address this aspect by examining the cosine similarity between input and output weights of a neuron. We apply our method to 12 models and find that enrichment neurons dominate in early-middle layers whereas later layers tend more towards depletion. To explain this finding, we argue that enrichment neurons are largely responsible for enriching concept representations, one of the first steps of factual recall. Our input-output perspective is a complement to activation-dependent analyses and to approaches that treat input and output separately.

Activation steering methods are widely used to control large language model (LLM) behavior and are often interpreted as revealing meaningful internal representations. This interpretation assumes that steering directions are identifiable and uniquely recoverable from input-output behavior. We show that, under white-box single-layer access, steering vectors are fundamentally non-identifiable due to large equivalence classes of behaviorally indistinguishable interventions. Empirically, we find that orthogonal perturbations achieve near-equivalent efficacy with negligible effect sizes across multiple models and traits, with pre-trained semantic classifiers confirming equivalence at the output level. We estimate null-space dimensionality via SVD of activation covariance matrices and validate that equivalence holds robustly throughout the operationally relevant steering range. Critically, we show that non-identifiability is a robust geometric property that persists across diverse prompt distributions. These findings reveal fundamental interpretability limits and highlight the need for structural constraints beyond behavioral testing to enable reliable alignment interventions.

LLMs are multilingual by training, yet their lingua franca is often English, reflecting English language dominance in pretraining. Other languages remain in parametric memory but are systematically suppressed. We argue that language defaultness is governed by a sparse, low-rank control circuit, language neurons, that can be mechanistically isolated and safely steered. We introduce Neural FOXP2, that makes a chosen language (Hindi or Spanish) primary in a model by steering language-specific neurons. Neural FOXP2 proceeds in three stages: (i) Localize: We train per-layer SAEs so each activation decomposes into a small set of active feature components. For every feature, we quantify English vs. Hindi/Spanish selectivity overall logit-mass lift toward the target-language token set. Tracing the top-ranked features back to their strongest contributing units yields a compact language-neuron set. (ii) Steering directions: We localize controllable language-shift geometry via a spectral low-rank analysis. For each layer, we build English to target activation-difference matrices and perform layerwise SVD to extract the dominant singular directions governing language change. The eigengap and effective-rank spectra identify a compact steering subspace and an empirically chosen intervention window (where these directions are strongest and most stable). (iii) Steer: We apply a signed, sparse activation shift targeted to the language neurons. Concretely, within low to mid layers we add a positive steering along the target-language dominant directions and a compensating negative shift toward the null space for the English neurons, yielding controllable target-language defaultness.

Multilingual large language models can generate figurative language, but whether the internal signals driving this behavior are language-specific or reusable across languages is unclear. Using activation steering as a probe, we estimate a direction for a figurative category from figurative--literal activation differences in one language and apply it during generation. Across five figurative categories, six languages, and four multilingual LLMs, these directions steer reliably within their own language, most robustly for metaphor and simile. More importantly, they transfer across languages: a direction learned in one increases the target behavior when applied to another, with German among the most receptive targets. Going further, directions assembled from other languages can match or even surpass a target language's own native direction, while removing this shared component weakens native steering. Together, these results provide direct evidence of a reusable but target-dependent cross-lingual signal for figurative generation.

Visual search is a ubiquitous and often challenging daily task, exemplified by looking for the car keys at home or a friend in a crowd. An intriguing property of some classical search tasks is an asymmetry such that finding a target A among distractors B can be easier than finding B among A. To elucidate the mechanisms responsible for asymmetry in visual search, we propose a computational model that takes a target and a search image as inputs and produces a sequence of eye movements until the target is found. The model integrates eccentricity-dependent visual recognition with target-dependent top-down cues. We compared the model against human behavior in six paradigmatic search tasks that show asymmetry in humans. Without prior exposure to the stimuli or task-specific training, the model provides a plausible mechanism for search asymmetry. We hypothesized that the polarity of search asymmetry arises from experience with the natural environment. We tested this hypothesis by training the model on augmented versions of ImageNet where the biases of natural images were either removed or reversed. The polarity of search asymmetry disappeared or was altered depending on the training protocol. This study highlights how classical perceptual properties can emerge in neural network models, without the need for task-specific training, but rather as a consequence of the statistical properties of the developmental diet fed to the model. All source code and data are publicly available at https://github.com/kreimanlab/VisualSearchAsymmetry.

This study reveals an unexpected parallel between instructible vision-language models (VLMs) and human cognitive disorders, specifically constructive apraxia. We tested 25 state-of-the-art VLMs, including GPT-4 Vision, DALL-E 3, and Midjourney v5, on their ability to generate images of the Ponzo illusion, a task that requires basic spatial reasoning and is often used in clinical assessments of constructive apraxia. Remarkably, 24 out of 25 models failed to correctly render two horizontal lines against a perspective background, mirroring the deficits seen in patients with parietal lobe damage. The models consistently misinterpreted spatial instructions, producing tilted or misaligned lines that followed the perspective of the background rather than remaining horizontal. This behavior is strikingly similar to how apraxia patients struggle to copy or construct simple figures despite intact visual perception and motor skills. Our findings suggest that current VLMs, despite their advanced capabilities in other domains, lack fundamental spatial reasoning abilities akin to those impaired in constructive apraxia. This limitation in AI systems provides a novel computational model for studying spatial cognition deficits and highlights a critical area for improvement in VLM architecture and training methodologies.

The Backpropagation algorithm has often been criticised for its lack of biological realism. In an attempt to find a more biologically plausible alternative, the recently introduced Forward-Forward algorithm replaces the forward and backward passes of Backpropagation with two forward passes. In this work, we show that the internal representations obtained by the Forward-Forward algorithm can organise into category-specific ensembles exhibiting high sparsity -- composed of a low number of active units. This situation is reminiscent of what has been observed in cortical sensory areas, where neuronal ensembles are suggested to serve as the functional building blocks for perception and action. Interestingly, while this sparse pattern does not typically arise in models trained with standard Backpropagation, it can emerge in networks trained with Backpropagation on the same objective proposed for the Forward-Forward algorithm.

Dense self-attention treats all token pairs as equally plausible before learning, an interaction-isotropic prior that can be mismatched to structured signals. For structured, low signal-to-noise ratio (SNR) signals such as EEG, dependencies are organized along the electrode and time axes, and this uniform prior exposes each token to many irrelevant interactions. We introduce Adaptive Anisotropic Attention (AAA), which splits attention into two paths: a temporal path, where each token attends to the tokens of its own electrode across time, and a spatial path, where it attends to the tokens of the other electrodes at the same time step. A small gate predicts, for every token, a convex combination of the two path outputs: two non-negative weights that sum to one. On six EEG downstream tasks, the resulting model, AXON (AXis-factorized Operator Network), improves mean balanced accuracy over a dense baseline under both linear probing and full fine-tuning. We show that both paths (temporal and spatial) are necessary and that the weighted sum beats a hard choice of one path; most of the benefit comes from the gate learning a different temporal/spatial balance at each layer of the network. Controlled audio spectrogram experiments show that axis factorization transfers beyond EEG. These results suggest that aligning attention with the natural axes of structured signals provides a useful inductive bias.

Capable language models hold what a character believes apart from what is true: told "Anna believes the cup is blue; in reality it is red," they answer blue about Anna and red about the world. Where in the computation does that separation live? We show it rests on two separable mechanisms at two positions. A generic value slot binds the attributed value. A router at the query position selects which frame, the character's belief or reality, a query reads out. Two routes fill the slot: an asserted belief, whose value the text supplies, binds in directly; a derived belief, whose value must be inferred from what the character could see, arrives by a visibility-gated lookback. A subspace trained on either route steers the other, and only the derived route depends on described visibility. The slot itself carries no belief-reality tag: intervening on it moves a reality readout as strongly as a belief one. The separation lives instead in a dissociated pair of routing subspaces, which flip a query between frames without injecting the donor's value. These results hold across three architectures, on stimuli de-confounded against theory-of-mind-benchmark shortcuts; the behavior itself emerges between 3B and 7B across five model families. This paper develops the single belief-reality axis in depth; a companion paper shows the same slot-and-router format is shared across the other non-actual contexts a sentence can open (counterfactual, fictional, temporal).

Steering is a widely used technique for controlling large language models, yet its effects are often unstable and hard to predict. Existing theoretical accounts are largely based on the Linear Representation Hypothesis (LRH). While LRH assumes that concepts can be orthogonalized for lossless control, this idealized mapping fails in real representations and cannot account for the observed unpredictability of steering. By relaxing LRH's orthogonality assumption while preserving linear representations, we show that overlapping concept contributions naturally yield a sample-specific axis-orthogonal structure. We formalize this as the Cylindrical Representation Hypothesis (CRH). In CRH, a central axis captures the main difference between concept absence and presence and drives concept generation. A surrounding normal plane controls steering sensitivity by determining how easily the axis can activate the target concept. Within this plane, only specific sensitive sectors strongly facilitate concept activation, while other sectors can suppress or delay it. While the surrounding normal plane can be reliably identified from difference vectors, the sensitive sector cannot, introducing intrinsic uncertainty at the sector level. This uncertainty provides a principled explanation for why steering outcomes often fluctuate even when using well-aligned directions. Our experiments verify the existence of the cylindrical structure and demonstrate that CRH provides a valid and practical way to interpret model steering behavior in real settings: https://github.com/mbzuai-nlp/CRH.

Attractor dynamics are a fundamental computational motif in neural circuits, supporting diverse cognitive functions through stable, self-sustaining patterns of neural activity. In these lecture notes, we review four key examples that demonstrate how autoassociative neural network models can elucidate the computational mechanisms underlying attractor-based information processing in biological neural systems performing cognitive functions. Drawing on empirical evidence, we explore hippocampal spatial representations, visual classification in the inferotemporal cortex, perceptual adaptation and priming, and working-memory biases shaped by sensory history. Across these domains, attractor network models reveal common computational principles and provide analytical insights into how experience shapes neural activity and behavior. Our synthesis underscores the value of attractor models as powerful tools for probing the neural basis of cognition and behavior.

Mechanistic interpretability aims to understand model behaviors in terms of specific, interpretable features, often hypothesized to manifest as low-dimensional subspaces of activations. Specifically, recent studies have explored subspace interventions (such as activation patching) as a way to simultaneously manipulate model behavior and attribute the features behind it to given subspaces. In this work, we demonstrate that these two aims diverge, potentially leading to an illusory sense of interpretability. Counterintuitively, even if a subspace intervention makes the model's output behave as if the value of a feature was changed, this effect may be achieved by activating a dormant parallel pathway leveraging another subspace that is causally disconnected from model outputs. We demonstrate this phenomenon in a distilled mathematical example, in two real-world domains (the indirect object identification task and factual recall), and present evidence for its prevalence in practice. In the context of factual recall, we further show a link to rank-1 fact editing, providing a mechanistic explanation for previous work observing an inconsistency between fact editing performance and fact localization. However, this does not imply that activation patching of subspaces is intrinsically unfit for interpretability. To contextualize our findings, we also show what a success case looks like in a task (indirect object identification) where prior manual circuit analysis informs an understanding of the location of a feature. We explore the additional evidence needed to argue that a patched subspace is faithful.

Recursive models can continue updating their latent states beyond their nominal inference budget, so an incorrect output at that budget does not show whether computation is unfinished or has entered a persistently unsuccessful regime. We study the dynamics of completion in attention- and MLP-based Tiny Recursive Models (TRMs) on 1,000 hard Sudoku puzzles. Extending recurrence from the nominal 16 steps to 512 steps increases cumulative exact-solve accuracy from 59.2% to 87.5% for the attention model and from 74.4% to 91.9% for the MLP model, solving more than two-thirds of the puzzles unsolved in the nominal budget. Across both architectures, latent-state motion drops sharply after the first exact solution. Completed states are typically locally contractive along the trajectory direction, even though the same local Jacobian retains strongly expanding directions. We characterize this phenomenon as trajectory-conditioned anisotropic stability. Perturbation experiments confirm this directional stability across both models. The multi-step fate of the maximally expanding direction differs: it is absorbed within 16 steps in the attention model but persists longer in the MLP model. The anisotropic-stability pattern also holds for a second attention checkpoint. Together, these results distinguish nominal-budget failure from completed computation and identify a common dynamical signature of completion across two recurrent architectures.

The wide-scale use of sparse autoencoders (SAEs) as interpretability tools is limited by inconsistent links between SAE features and model behavior. Features with clear activation descriptions may have weak or unexpected causal effects; steering can vary across prompts or oppose the intended direction; and activation-based feature selection can miss features that produce the desired output change. Prior work has studied feature geometry inside the model, where features are computed. We instead study the geometry of changes in model logits caused by feature interventions. We introduce Feature-Effect Geometry Analysis (FEGA), an unsupervised framework that removes the same active SAE feature across contexts and analyzes the resulting cloud of logit changes. Across SAE variants, consistent one-dimensional effects are rare: few features behave like reusable directions. To interpret this variation, we distinguish value-like features, tied to static information such as factual attributes, from pointer-like features, associated with context-dependent operations. Value-like features more often exhibit structured, low-dimensional effects, although these effects typically span several directions. Pointer-like features, by contrast, predominantly exhibit diffuse effects. Our results show that a feature can be interpretable and causally relevant without providing a stable direction for steering.

Despite increasing progress in development of methods for generating visual counterfactual explanations, especially with the recent rise of Denoising Diffusion Probabilistic Models, previous works consider them as an entirely local technique. In this work, we take the first step at globalizing them. Specifically, we discover that the latent space of Diffusion Autoencoders encodes the inference process of a given classifier in the form of global directions. We propose a novel proxy-based approach that discovers two types of these directions with the use of only single image in an entirely black-box manner. Precisely, g-directions allow for flipping the decision of a given classifier on an entire dataset of images, while h-directions further increase the diversity of explanations. We refer to them in general as Global Counterfactual Directions (GCDs). Moreover, we show that GCDs can be naturally combined with Latent Integrated Gradients resulting in a new black-box attribution method, while simultaneously enhancing the understanding of counterfactual explanations. We validate our approach on existing benchmarks and show that it generalizes to real-world use-cases.

Large Language Models (LLMs) rely on safety alignment to produce socially acceptable responses. This is typically achieved through instruction tuning and reinforcement learning from human feedback. However, this alignment is known to be brittle: further fine-tuning, even on benign or lightly contaminated data, can degrade safety and reintroduce harmful behaviors. A growing body of work suggests that alignment may correspond to identifiable geometric directions in weight space, forming subspaces that could, in principle, be isolated or preserved to defend against misalignment. In this work, we conduct a comprehensive empirical study of this geometric perspective. We examine whether safety-relevant behavior is concentrated in specific subspaces, whether it can be separated from general-purpose learning, and whether harmfulness arises from distinguishable patterns in internal representations. Across both parameter and activation space, our findings are consistent: subspaces that amplify safe behaviors also amplify unsafe ones, and prompts with different safety implications activate overlapping representations. We find no evidence of a subspace that selectively governs safety. These results challenge the assumption that alignment is geometrically localized. Rather than residing in distinct directions, safety appears to emerge from entangled, high-impact components of the model's broader learning dynamics. This suggests that subspace-based defenses may face fundamental limitations and underscores the need for alternative strategies to preserve alignment under continued training. We corroborate these findings through multiple experiments on five open-source LLMs. Our code is publicly available at: https://github.com/CERT-Lab/safety-subspaces.

Can one inject new concepts into an already trained generative model, while respecting its existing structure and knowledge? We propose a new task - domain expansion - to address this. Given a pretrained generator and novel (but related) domains, we expand the generator to jointly model all domains, old and new, harmoniously. First, we note the generator contains a meaningful, pretrained latent space. Is it possible to minimally perturb this hard-earned representation, while maximally representing the new domains? Interestingly, we find that the latent space offers unused, "dormant" directions, which do not affect the output. This provides an opportunity: By "repurposing" these directions, we can represent new domains without perturbing the original representation. In fact, we find that pretrained generators have the capacity to add several - even hundreds - of new domains! Using our expansion method, one "expanded" model can supersede numerous domain-specific models, without expanding the model size. Additionally, a single expanded generator natively supports smooth transitions between domains, as well as composition of domains. Code and project page available at https://yotamnitzan.github.io/domain-expansion/.

Alignment in LLMs is more brittle than commonly assumed: misalignment can be triggered by adversarial prompts, benign fine-tuning, emergent misalignment, and goal misgeneralization. Recent evidence suggests that some misalignment behaviors are encoded as linear structure in activation space, making it tractable via steering, while safety alignment has been shown to govern the first few output tokens primarily, leaving subsequent generation unguarded. These findings motivate activation steering as a lightweight runtime defense that continuously corrects misaligned activations throughout generation. We evaluate three methods: Steer-With-Fixed-Coeff (SwFC), which applies uniform additive steering, and two novel projection-aware methods, Steer-to-Target-Projection (StTP) and Steer-to-Mirror-Projection (StMP), that use a logistic regression decision boundary to selectively intervene only on tokens whose activations fall below distributional thresholds. Using malicious system prompts as a controlled proxy for misalignment, we evaluate under two threat models (dishonesty and dismissiveness) and two architectures (Llama-3.3-70B-Instruct, Qwen3-32B). All methods substantially recover target traits (honesty and compassion) while preserving coherence. StTP and StMP better maintain general capabilities (MMLU, MT-Bench, AlpacaEval) and produce less repetition in multi-turn conversations.

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

Feature learning is widely regarded as the key mechanism distinguishing neural networks from fixed-kernel methods, yet its impact on the induced function space remains poorly understood. In this work, we precisely characterize how the function space spanned by the features of a two-layer neural network evolves during gradient descent training. We prove that, in the high-dimensional proportional regime, after a large gradient step the post-update feature distribution is well approximated by a target-dependent spiked Gaussian covariance. This induces a data-adaptive kernel that reshapes the function space and modifies its spectral structure. Our analysis reveals that feature learning can be interpreted as a distributional transformation in either parameter space or input space, equivalently as the introduction of a target-dependent kernel. In particular, it selectively amplifies eigenvalues aligned with the target direction and mixes leading eigenfunctions, coupling the top radial mode with a target-aligned quadratic harmonic. Overall, our results provide a precise function-space perspective on early-stage feature learning: rather than just rescaling a fixed kernel, gradient descent induces a data-adaptive deformation that preferentially enhances directions aligned with the signal in the data.

Vision-Language-Action (VLA) models are a promising path to realizing generalist embodied agents that can quickly adapt to new tasks, modalities, and environments. However, methods for interpreting and steering VLAs fall far short of classical robotics pipelines, which are grounded in explicit models of kinematics, dynamics, and control. This lack of mechanistic insight is a central challenge for deploying learned policies in real-world robotics, where robustness and explainability are critical. Motivated by advances in mechanistic interpretability for large language models, we introduce the first framework for interpreting and steering VLAs via their internal representations, enabling direct intervention in model behavior at inference time. We project feedforward activations within transformer layers onto the token embedding basis, identifying sparse semantic directions - such as speed and direction - that are causally linked to action selection. Leveraging these findings, we introduce a general-purpose activation steering method that modulates behavior in real time, without fine-tuning, reward signals, or environment interaction. We evaluate this method on two recent open-source VLAs, Pi0 and OpenVLA, and demonstrate zero-shot behavioral control in simulation (LIBERO) and on a physical robot (UR5). This work demonstrates that interpretable components of embodied VLAs can be systematically harnessed for control - establishing a new paradigm for transparent and steerable foundation models in robotics.

Induction heads provide a mechanistic account of in-context learning in sequential data, but existing theory largely assumes that the context relevant to a prediction forms a contiguous block. In multidimensional data, serialization breaks this assumption by scattering spatial neighbors across distant positions in the token sequence. We study how transformers overcome this routing problem in multidimensional stochastic and deterministic cellular automata, where each trajectory is generated by an unknown local rule and presented as a flattened sequence without an explicit coordinate-based spatial inductive bias. We introduce spatial induction heads, two-layer gather-and-match circuits in which the first layer reconstructs the relevant spatial neighborhood and the second matches the resulting configuration against earlier occurrences. We give two explicit realizations of the gather and show that the positional dimension required for spatial routing depends only on the local neighborhood and spatial dimension, not on grid volume or trajectory horizon. We further construct a matching layer which implements Bayesian counting. The end-to-end circuit can approximate the Bayesian posterior arbitrarily closely for stochastic rules and can predict exactly for deterministic rules. Empirically, trained two-layer transformers generalize to unseen rules in one and two dimensional settings, achieving near-perfect deterministic rollouts and less than 0.005 nats KL from the Bayes-optimal predictor on stochastic rules. Attention patterns and layerwise probes align with the predicted gather-and-match computation, providing mechanistic evidence for spatial induction in trained transformers.