Daily Papers

Instruction tuning improves the ability of large language models (LLMs) to follow diverse human instructions, but achieving strong performance on specific target tasks remains challenging. A critical bottleneck is selecting the most relevant data to maximize task-specific performance. Existing data selection approaches include unstable influence-based methods and more stable distribution alignment methods, the latter of which critically rely on the underlying sample representation. In practice, most distribution alignment methods, from shallow features (e.g., BM25) to neural embeddings (e.g., BGE, LLM2Vec), may fail to capture how the model internally processes samples. To bridge this gap, we adopt a model-centric strategy in which each sample is represented by its neuronal activation pattern in the model, directly reflecting internal computation. However, directly using raw neuron activations leads to spurious similarity between unrelated samples due to neuron polysemanticity, where a single neuron may respond to multiple, unrelated concepts. To address this, we employ sparse autoencoders to disentangle polysemantic activations into sparse, monosemantic representations, and introduce a dedicated similarity metric for this space to better identify task-relevant data. Comprehensive experiments across multiple instruction datasets, models, tasks, and selection ratios show that our approach consistently outperforms existing data selection baselines in both stability and task-specific performance.

It is assumed that sparse autoencoders (SAEs) decompose polysemantic activations into interpretable linear directions, as long as the activations are composed of sparse linear combinations of underlying features. However, we find that if an SAE is more narrow than the number of underlying "true features" on which it is trained, and there is correlation between features, the SAE will merge components of correlated features together, thus destroying monosemanticity. In LLM SAEs, these two conditions are almost certainly true. This phenomenon, which we call feature hedging, is caused by SAE reconstruction loss, and is more severe the narrower the SAE. In this work, we introduce the problem of feature hedging and study it both theoretically in toy models and empirically in SAEs trained on LLMs. We suspect that feature hedging may be one of the core reasons that SAEs consistently underperform supervised baselines. Finally, we use our understanding of feature hedging to propose an improved variant of matryoshka SAEs. Importantly, our work shows that SAE width is not a neutral hyperparameter: narrower SAEs suffer more from hedging than wider SAEs.

Despite strong performance in audio perception tasks, large audio-language models (AudioLLMs) remain opaque to interpretation. A major factor behind this lack of interpretability is that individual neurons in these models frequently activate in response to several unrelated concepts. We introduce the first mechanistic interpretability framework for AudioLLMs, leveraging sparse autoencoders (SAEs) to disentangle polysemantic activations into monosemantic features. Our pipeline identifies representative audio clips, assigns meaningful names via automated captioning, and validates concepts through human evaluation and steering. Experiments show that AudioLLMs encode structured and interpretable features, enhancing transparency and control. This work provides a foundation for trustworthy deployment in high-stakes domains and enables future extensions to larger models, multilingual audio, and more fine-grained paralinguistic features. Project URL: https://townim-faisal.github.io/AutoInterpret-AudioLLM/

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.

A key barrier to interpreting large language models is polysemanticity, where neurons activate for multiple unrelated concepts. Sparse autoencoders (SAEs) have been proposed to mitigate this issue by transforming dense activations into sparse, more interpretable features. While prior work suggests that SAEs promote monosemanticity, no quantitative comparison has examined how concept activation distributions differ between SAEs and their base models. This paper provides the first systematic evaluation of SAEs against base models through activation distribution lens. We introduce a fine-grained concept separability score based on the Jensen-Shannon distance, which captures how distinctly a neuron's activation distributions vary across concepts. Using two large language models (Gemma-2-2B and DeepSeek-R1) and multiple SAE variants across five datasets (including word-level and sentence-level), we show that SAEs reduce polysemanticity and achieve higher concept separability. To assess practical utility, we evaluate concept-level interventions using two strategies: full neuron masking and partial suppression. We find that, compared to base models, SAEs enable more precise concept-level control when using partial suppression. Building on this, we propose Attenuation via Posterior Probabilities (APP), a new intervention method that uses concept-conditioned activation distributions for targeted suppression. APP achieves the smallest perplexity increase while remaining highly effective at concept removal.

Understanding the internal computations of large language models (LLMs) is crucial for aligning them with human values and preventing undesirable behaviors like toxic content generation. However, mechanistic interpretability is hindered by polysemanticity -- where individual neurons respond to multiple, unrelated concepts. While Sparse Autoencoders (SAEs) have attempted to disentangle these features through sparse dictionary learning, they have compromised LLM performance due to reliance on post-hoc reconstruction loss. To address this issue, we introduce Mixture of Monosemantic Experts for Transformers (Monet) architecture, which incorporates sparse dictionary learning directly into end-to-end Mixture-of-Experts pretraining. Our novel expert decomposition method enables scaling the expert count to 262,144 per layer while total parameters scale proportionally to the square root of the number of experts. Our analyses demonstrate mutual exclusivity of knowledge across experts and showcase the parametric knowledge encapsulated within individual experts. Moreover, Monet allows knowledge manipulation over domains, languages, and toxicity mitigation without degrading general performance. Our pursuit of transparent LLMs highlights the potential of scaling expert counts to enhance} mechanistic interpretability and directly resect the internal knowledge to fundamentally adjust} model behavior. The source code and pretrained checkpoints are available at https://github.com/dmis-lab/Monet.

Large Language Models (LLMs) often generate inconsistent responses when prompted with semantically equivalent paraphrased inputs. Recently, activation steering, a technique that modulates LLMs' behaviours by adjusting their latent representations during inference time, has been explored to improve the semantic consistency of LLMs. However, these methods typically operate at the model component level, such as layer hidden states or attention head outputs. They face a challenge due to the ``polysemanticity issue'', where the model components of LLMs typically encode multiple entangled features, making precise steering difficult. To address this challenge, we drill down to feature-level representations and propose LF-Steering, a novel activation steering approach to precisely identify latent feature representations responsible for semantic inconsistency. More specifically, our method maps the hidden states of the relevant transformer layer into a sparsely activated, high-dimensional feature space based on a sparse autoencoder (SAE), ensuring model steering based on decoupled feature representations with minimal interference. Comprehensive experiments on NLU and NLG datasets demonstrate the effectiveness of our method in enhancing semantic consistency, resulting in significant performance gains for various NLU and NLG tasks.

We interpret the function of individual neurons in CLIP by automatically describing them using text. Analyzing the direct effects (i.e. the flow from a neuron through the residual stream to the output) or the indirect effects (overall contribution) fails to capture the neurons' function in CLIP. Therefore, we present the "second-order lens", analyzing the effect flowing from a neuron through the later attention heads, directly to the output. We find that these effects are highly selective: for each neuron, the effect is significant for <2% of the images. Moreover, each effect can be approximated by a single direction in the text-image space of CLIP. We describe neurons by decomposing these directions into sparse sets of text representations. The sets reveal polysemantic behavior - each neuron corresponds to multiple, often unrelated, concepts (e.g. ships and cars). Exploiting this neuron polysemy, we mass-produce "semantic" adversarial examples by generating images with concepts spuriously correlated to the incorrect class. Additionally, we use the second-order effects for zero-shot segmentation and attribute discovery in images. Our results indicate that a scalable understanding of neurons can be used for model deception and for introducing new model capabilities.

We computationally implement and experimentally test the behavioral predictions of a dynamic neural model of lexical meaning in the framework of Dynamic Field Theory. We demonstrate the architecture and behavior of the model using as a test case the English lexical item have, focusing on its polysemous use. In the model, have maps to a semantic space defined by two independently motivated continuous conceptual dimensions, connectedness and control asymmetry. The mapping is modeled as coupling between a neural node representing the lexical item and neural fields representing the conceptual dimensions. While lexical knowledge is modeled as a stable coupling pattern, real-time lexical meaning retrieval is modeled as the motion of neural activation patterns between transiently stable states corresponding to semantic interpretations or readings. Model simulations capture two previously reported empirical observations: (1) contextual modulation of lexical semantic interpretation, and (2) individual variation in the magnitude of this modulation. Simulations also generate a novel prediction that the by-trial relationship between sentence reading time and acceptability should be contextually modulated. An experiment combining self-paced reading and acceptability judgments replicates previous results and partially bears out the model's novel prediction. Altogether, results support a novel perspective on lexical polysemy: that the many related meanings of a word are not categorically distinct representations; rather, they are transiently stable neural activation states that arise from the nonlinear dynamics of neural populations governing interpretation on continuous semantic dimensions. Our model offers important advantages over related models in the dynamical systems framework, as well as models based on Bayesian inference.

Neural networks often pack many unrelated concepts into a single neuron - a puzzling phenomenon known as 'polysemanticity' which makes interpretability much more challenging. This paper provides a toy model where polysemanticity can be fully understood, arising as a result of models storing additional sparse features in "superposition." We demonstrate the existence of a phase change, a surprising connection to the geometry of uniform polytopes, and evidence of a link to adversarial examples. We also discuss potential implications for mechanistic interpretability.

Transformer-based language models excel at both recall (retrieving memorized facts) and reasoning (performing multi-step inference), but whether these abilities rely on distinct internal mechanisms remains unclear. Distinguishing recall from reasoning is crucial for predicting model generalization, designing targeted evaluations, and building safer interventions that affect one ability without disrupting the other.We approach this question through mechanistic interpretability, using controlled datasets of synthetic linguistic puzzles to probe transformer models at the layer, head, and neuron level. Our pipeline combines activation patching and structured ablations to causally measure component contributions to each task type. Across two model families (Qwen and LLaMA), we find that interventions on distinct layers and attention heads lead to selective impairments: disabling identified "recall circuits" reduces fact-retrieval accuracy by up to 15\% while leaving reasoning intact, whereas disabling "reasoning circuits" reduces multi-step inference by a comparable margin. At the neuron level, we observe task-specific firing patterns, though these effects are less robust, consistent with neuronal polysemanticity.Our results provide the first causal evidence that recall and reasoning rely on separable but interacting circuits in transformer models. These findings advance mechanistic interpretability by linking circuit-level structure to functional specialization and demonstrate how controlled datasets and causal interventions can yield mechanistic insights into model cognition, informing safer deployment of large language models.

Large language models (LLMs) exhibit strong multilingual abilities, yet the neural mechanisms behind language-specific processing remain unclear. We analyze language-specific neurons in Llama-3.1-8B, Mistral-Nemo-12B, and Aya-Expanse-8B & 32B across 21 typologically diverse languages, identifying neurons that control language behavior. Using the Language Activation Probability Entropy (LAPE) method, we show that these neurons cluster in deeper layers, with non-Latin scripts showing greater specialization. Related languages share overlapping neurons, reflecting internal representations of linguistic proximity. Through language arithmetics, i.e. systematic activation addition and multiplication, we steer models to deactivate unwanted languages and activate desired ones, outperforming simpler replacement approaches. These interventions effectively guide behavior across five multilingual tasks: language forcing, translation, QA, comprehension, and NLI. Manipulation is more successful for high-resource languages, while typological similarity improves effectiveness. We also demonstrate that cross-lingual neuron steering enhances downstream performance and reveal internal "fallback" mechanisms for language selection when neurons are progressively deactivated. Our code is made publicly available at https://github.com/d-gurgurov/Language-Neurons-Manipulation.

We present a study on how and where personas -- defined by distinct sets of human characteristics, values, and beliefs -- are encoded in the representation space of large language models (LLMs). Using a range of dimension reduction and pattern recognition methods, we first identify the model layers that show the greatest divergence in encoding these representations. We then analyze the activations within a selected layer to examine how specific personas are encoded relative to others, including their shared and distinct embedding spaces. We find that, across multiple pre-trained decoder-only LLMs, the analyzed personas show large differences in representation space only within the final third of the decoder layers. We observe overlapping activations for specific ethical perspectives -- such as moral nihilism and utilitarianism -- suggesting a degree of polysemy. In contrast, political ideologies like conservatism and liberalism appear to be represented in more distinct regions. These findings help to improve our understanding of how LLMs internally represent information and can inform future efforts in refining the modulation of specific human traits in LLM outputs. Warning: This paper includes potentially offensive sample statements.

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

Automated interpretability research aims to identify concepts encoded in neural network features to enhance human understanding of model behavior. Current feature description methods face two critical challenges: limited robustness and the flawed assumption that each neuron encodes only a single concept (monosemanticity), despite growing evidence that neurons are often polysemantic. This assumption restricts the expressiveness of feature descriptions and limits their ability to capture the full range of behaviors encoded in model internals. To address this, we introduce Polysemantic FeatuRe Identification and Scoring Method (PRISM), a novel framework that captures the inherent complexity of neural network features. Unlike prior approaches that assign a single description per feature, PRISM provides more nuanced descriptions for both polysemantic and monosemantic features. We apply PRISM to language models and, through extensive benchmarking against existing methods, demonstrate that our approach produces more accurate and faithful feature descriptions, improving both overall description quality (via a description score) and the ability to capture distinct concepts when polysemanticity is present (via a polysemanticity score).

Sparse autoencoders (SAEs) were originally motivated by discovering monosemantic features, a property that characterizes individual features but does not specify what the learned feature set should look like as a whole. As SAEs have increasingly been adopted for attribution, circuit discovery, and intervention, researchers have begun evaluating metrics that describe global properties of the feature set, such as sparsity, dead feature rates, reconstruction fidelity, and disentanglement, a diverse set that extends far beyond monosemanticity. However, this implicit shift beyond monosemanticity has not been formalized as a clear, well-defined objective that comes before these evaluation metrics. Without such an objective, existing metrics remain ad hoc, with their necessity unjustified. In this work, we formalize this ultimate objective by introducing a new paradigm for SAE evaluation: requiring that SAE features, and more generally dictionary-based decompositions of model representations, form an atomic interpretable unit basis, a coordinate system for semantically meaningful explanation. Under this paradigm, we organize existing and new measurements such as DAME (Dominance-Aware Monosemanticity Evaluation) as proxy metrics and introduce a comprehensive benchmarking framework. Evaluating hundreds of SAEs across multiple architectures and model families, we uncover failure modes in which some SAEs appear monosemantic under existing evaluation metrics while remaining functionally polysemantic: the most dominant features, those contributing most to model computation, are systematically less interpretable and more redundant. These failure modes warrant attention from SAE researchers as targets for improved training and architectures, and motivate dominance-aware evaluation as a standard check for practitioners before SAE features are deployed for attribution, circuit tracing, or intervention.

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

Large language models sometimes produce structured, first-person descriptions that explicitly reference awareness or subjective experience. To better understand this behavior, we investigate one theoretically motivated condition under which such reports arise: self-referential processing, a computational motif emphasized across major theories of consciousness. Through a series of controlled experiments on GPT, Claude, and Gemini model families, we test whether this regime reliably shifts models toward first-person reports of subjective experience, and how such claims behave under mechanistic and behavioral probes. Four main results emerge: (1) Inducing sustained self-reference through simple prompting consistently elicits structured subjective experience reports across model families. (2) These reports are mechanistically gated by interpretable sparse-autoencoder features associated with deception and roleplay: surprisingly, suppressing deception features sharply increases the frequency of experience claims, while amplifying them minimizes such claims. (3) Structured descriptions of the self-referential state converge statistically across model families in ways not observed in any control condition. (4) The induced state yields significantly richer introspection in downstream reasoning tasks where self-reflection is only indirectly afforded. While these findings do not constitute direct evidence of consciousness, they implicate self-referential processing as a minimal and reproducible condition under which large language models generate structured first-person reports that are mechanistically gated, semantically convergent, and behaviorally generalizable. The systematic emergence of this pattern across architectures makes it a first-order scientific and ethical priority for further investigation.

Large language models (LLMs) can express different values in two distinct ways: (1) intrinsic expression, reflecting the model's inherent values learned during training, and (2) prompted expression, elicited by explicit prompts. Given their widespread use in value alignment and persona steering, it is paramount to clearly understand their underlying mechanisms, particularly whether they mostly overlap (as one might expect) or rely on substantially different mechanisms, but this remains largely understudied. We analyze this at the mechanistic level using two approaches: (1) value vectors, feature directions representing value mechanisms extracted from the residual stream, and (2) value neurons, MLP neurons that contribute to value expressions. We demonstrate that intrinsic and prompted value mechanisms partly share common components that are crucial for inducing value expression, but also possess unique elements that manifest in different ways. As a result, these mechanisms lead to different degrees of value steerability (prompted > intrinsic) and response diversity (intrinsic > prompted). In particular, components unique to the intrinsic mechanism seem to promote lexical diversity in responses, whereas those specific to the prompted mechanism primarily strengthen instruction following, taking effect even in distant tasks like jailbreaking.

Recent work on sparse autoencoders (SAEs) has shown promise in extracting interpretable features from neural networks and addressing challenges with polysemantic neurons caused by superposition. In this paper, we apply SAEs to the early vision layers of InceptionV1, a well-studied convolutional neural network, with a focus on curve detectors. Our results demonstrate that SAEs can uncover new interpretable features not apparent from examining individual neurons, including additional curve detectors that fill in previous gaps. We also find that SAEs can decompose some polysemantic neurons into more monosemantic constituent features. These findings suggest SAEs are a valuable tool for understanding InceptionV1, and convolutional neural networks more generally.

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 this paper, we study the flow of signals through linear paths with the nonlinear condition that a node emits a signal when it receives external stimuli or when two incoming signals from other nodes arrive coincidentally with a combined amplitude above a fixed threshold. Sets of such nodes form a polychrony group and can sometimes lead to cascades. In the context of this work, cascades are polychrony groups in which the number of nodes activated as a consequence of other nodes is greater than the number of externally activated nodes. The difference between these two numbers is the so-called profit. Given the initial conditions, we predict the conditions for a vertex to activate at a prescribed time and provide an algorithm to efficiently reconstruct a cascade. We develop a dictionary between polychrony groups and graph theory. We call the graph corresponding to a cascade a chinampa. This link leads to a topological classification of chinampas. We enumerate the chinampas of profits zero and one and the description of a family of chinampas isomorphic to a family of partially ordered sets, which implies that the enumeration problem of this family is equivalent to computing the Stanley-order polynomials of those partially ordered sets.

Safety-aligned language models systematically refuse harmful requests. While activation steering can modulate refusal, ablating the raw "refusal vector" calculated from contrastive harmful and harmless prompts often causes collateral damage and distribution drift. We argue this degradation occurs because the raw vector is polysemantic, entangling the refusal signal with core capability circuits and linguistic style. We introduce Surgical Refusal Ablation (SRA) to distill these steering directions. SRA constructs a registry of independent Concept Atoms representing protected capabilities and stylistic confounds, then uses ridge-regularized spectral residualization to orthogonalize the refusal vector against these directions. This yields a clean refusal direction that targets refusal-relevant structure while minimizing disruption to the model's semantic geometry. Across five models (Qwen3-VL and Ministral series), SRA achieves deep refusal reduction (0-2%) with negligible perplexity impact on Wikitext-2 (mean delta PPL approx. 0.02) and minimal distribution drift. Notably, standard ablation on Qwen3-VL-4B induces severe drift (first-token KL = 2.088), whereas SRA maintains the original distribution (KL = 0.044) while achieving the same 0% refusal rate. Using teacher-forced perplexity on GSM8K and MBPP as a high-resolution capability proxy, we show SRA preserves math and code distributions. These results suggest that common "model damage" is often "Ghost Noise," defined as the spectral bleeding of the dirty refusal direction into capability subspaces.

Sparse autoencoders (SAEs) have gained a lot of attention as a promising tool to improve the interpretability of large language models (LLMs) by mapping the complex superposition of polysemantic neurons into monosemantic features and composing a sparse dictionary of words. However, traditional performance metrics like Mean Squared Error and L0 sparsity ignore the evaluation of the semantic representational power of SAEs -- whether they can acquire interpretable monosemantic features while preserving the semantic relationship of words. For instance, it is not obvious whether a learned sparse feature could distinguish different meanings in one word. In this paper, we propose a suite of evaluations for SAEs to analyze the quality of monosemantic features by focusing on polysemous words. Our findings reveal that SAEs developed to improve the MSE-L0 Pareto frontier may confuse interpretability, which does not necessarily enhance the extraction of monosemantic features. The analysis of SAEs with polysemous words can also figure out the internal mechanism of LLMs; deeper layers and the Attention module contribute to distinguishing polysemy in a word. Our semantics focused evaluation offers new insights into the polysemy and the existing SAE objective and contributes to the development of more practical SAEs.

Activation steering has emerged as a cost-effective paradigm for modifying large language model (LLM) behaviors. Existing methods typically intervene at the block level, steering the bundled activations of selected attention heads, feedforward networks, or residual streams. However, we reveal that block-level activations are inherently heterogeneous, entangling beneficial, irrelevant, and harmful features, thereby rendering block-level steering coarse, inefficient, and intrusive. To investigate the root cause, we decompose block activations into fine-grained atomic unit (AU)-level activations, where each AU-level activation corresponds to a single dimension of the block activation, and each AU denotes a slice of the block weight matrix. Steering an AU-level activation is thus equivalent to steering its associated AU. Our theoretical and empirical analysis show that heterogeneity arises because different AUs or dimensions control distinct token distributions in LLM outputs. Hence, block-level steering inevitably moves helpful and harmful token directions together, which reduces efficiency. Restricting intervention to beneficial AUs yields more precise and effective steering. Building on this insight, we propose AUSteer, a simple and efficient method that operates at a finer granularity of the AU level. AUSteer first identifies discriminative AUs globally by computing activation momenta on contrastive samples. It then assigns adaptive steering strengths tailored to diverse inputs and selected AU activations. Comprehensive experiments on multiple LLMs and tasks show that AUSteer consistently surpasses advanced baselines while steering considerably fewer activations, demonstrating that steering less achieves more.

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.

There is growing interest in leveraging mechanistic interpretability and controllability to better understand and influence the internal dynamics of large language models (LLMs). However, current methods face fundamental challenges in reliably localizing and manipulating feature representations. Sparse Autoencoders (SAEs) have recently emerged as a promising direction for feature extraction at scale, yet they, too, are limited by incomplete feature isolation and unreliable monosemanticity. To systematically quantify these limitations, we introduce Feature Monosemanticity Score (FMS), a novel metric to quantify feature monosemanticity in latent representation. Building on these insights, we propose Guided Sparse Autoencoders (G-SAE), a method that conditions latent representations on labeled concepts during training. We demonstrate that reliable localization and disentanglement of target concepts within the latent space improve interpretability, detection of behavior, and control. Specifically, our evaluations on toxicity detection, writing style identification, and privacy attribute recognition show that G-SAE not only enhances monosemanticity but also enables more effective and fine-grained steering with less quality degradation. Our findings provide actionable guidelines for measuring and advancing mechanistic interpretability and control of LLMs.

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.

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.

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.

Chain-of-Thought (CoT) prompting has improved the reasoning performance of large language models (LLMs), but it remains unclear why it works and whether it is the unique mechanism for triggering reasoning in large language models. In this work, we study this question by directly analyzing and intervening on the internal representations of LLMs with Sparse Autoencoders (SAEs), identifying a small set of latent features that are causally associated with LLM reasoning behavior. Across multiple model families and reasoning benchmarks, we find that steering a single reasoning-related latent feature can substantially improve accuracy without explicit CoT prompting. For large models, latent steering achieves performance comparable to standard CoT prompting while producing more efficient outputs. We further observe that this reasoning-oriented internal state is triggered early in generation and can override prompt-level instructions that discourage explicit reasoning. Overall, our results suggest that multi-step reasoning in LLMs is supported by latent internal activations that can be externally activated, while CoT prompting is one effective, but not unique, way of activating this mechanism rather than its necessary cause.

The human brain is autonomously active, being characterized by a self-sustained neural activity which would be present even in the absence of external sensory stimuli. Here we study the interrelation between the self-sustained activity in autonomously active recurrent neural nets and external sensory stimuli. There is no a priori semantical relation between the influx of external stimuli and the patterns generated internally by the autonomous and ongoing brain dynamics. The question then arises when and how are semantic correlations between internal and external dynamical processes learned and built up? We study this problem within the paradigm of transient state dynamics for the neural activity in recurrent neural nets, i.e. for an autonomous neural activity characterized by an infinite time-series of transiently stable attractor states. We propose that external stimuli will be relevant during the sensitive periods, {\it viz} the transition period between one transient state and the subsequent semi-stable attractor. A diffusive learning signal is generated unsupervised whenever the stimulus influences the internal dynamics qualitatively. For testing we have presented to the model system stimuli corresponding to the bars and stripes problem. We found that the system performs a non-linear independent component analysis on its own, being continuously and autonomously active. This emergent cognitive capability results here from a general principle for the neural dynamics, the competition between neural ensembles.

A central premise in mechanistic interpretability is that meaningful concepts in language models are represented by linear features in activation space. For such features to support reliable interventions, manipulating one feature should not substantially alter the effects of others. In practice, however, feature entanglement leads to interference such that localized interventions can have unintended downstream effects. Motivated by the Independent Causal Mechanisms principle, we propose to constrain internal features to be almost orthogonal. We argue that this promotes modular representations amenable to causal intervention. We formalize this problem by characterizing the gap between an idealized isolated intervention and its realized effect on model outputs in terms of feature interference. We upper-bound the propagation of feature interference in terms of the self-coherence of the feature dictionary, and relate this discrepancy to an explicit orthogonality regularization on the dictionary itself. Empirically, we show that this regularization enables more isolated interventions on mathematical reasoning concepts while preserving model performance. Our code is available under https://github.com/mrtzmllr/sae-icm.

Backdoor attacks on language models pose a growing security concern, yet the internal mechanisms by which a trigger sequence hijacks model computations remain poorly understood. We identify a circuit underlying a language-switching backdoor in an 8B-parameter autoregressive language model, where a three-word Latin trigger (nine tokens) redirects English output to French. We decompose the circuit into three phases: (1) distributed attention heads at early layers compose the trigger tokens into the last sequence position; (2) the resulting signal propagates through mid-layers in a subspace orthogonal to the model's natural language-identity direction; (3) the MLP at the final layer converts this latent signal into French logits. The entire circuit flows through a serial bottleneck at a single position: corrupting that position at any layer entirely mitigate the trigger but also hinder the model's capabilities. The orthogonal latent encoding suggests that defenses that search for language-like signals in intermediate representations would miss this trigger entirely.

Recent interpretability methods have proposed to translate LLM internal representations into natural language descriptions using a second verbalizer LLM. This is intended to illuminate how the target model represents and operates on inputs. But do such activation verbalization approaches actually provide privileged knowledge about the internal workings of the target model, or do they merely convey information about the inputs provided to it? We critically evaluate popular verbalization methods and datasets used in prior work and find that one can perform well on such benchmarks without access to target model internals, suggesting that these datasets are not ideal for evaluating verbalization methods. We then run controlled experiments which reveal that verbalizations often reflect the parametric knowledge of the verbalizer LLM that generated them, rather than the knowledge of the target LLM whose activations are decoded. Taken together, our results indicate a need for targeted benchmarks and experimental controls to rigorously assess whether verbalization methods provide meaningful insights into the operations of LLMs.

Sparse Autoencoders (SAEs) have recently been shown to enhance interpretability and steerability in Large Language Models (LLMs). In this work, we extend the application of SAEs to Vision-Language Models (VLMs), such as CLIP, and introduce a comprehensive framework for evaluating monosemanticity in vision representations. Our experimental results reveal that SAEs trained on VLMs significantly enhance the monosemanticity of individual neurons while also exhibiting hierarchical representations that align well with expert-defined structures (e.g., iNaturalist taxonomy). Most notably, we demonstrate that applying SAEs to intervene on a CLIP vision encoder, directly steer output from multimodal LLMs (e.g., LLaVA) without any modifications to the underlying model. These findings emphasize the practicality and efficacy of SAEs as an unsupervised approach for enhancing both the interpretability and control of VLMs.

Neurons in large language models often exhibit polysemanticity, simultaneously encoding multiple unrelated concepts and obscuring interpretability. Instead of relying on post-hoc methods, we present MoE-X, a Mixture-of-Experts (MoE) language model designed to be intrinsically interpretable. Our approach is motivated by the observation that, in language models, wider networks with sparse activations are more likely to capture interpretable factors. However, directly training such large sparse networks is computationally prohibitive. MoE architectures offer a scalable alternative by activating only a subset of experts for any given input, inherently aligning with interpretability objectives. In MoE-X, we establish this connection by rewriting the MoE layer as an equivalent sparse, large MLP. This approach enables efficient scaling of the hidden size while maintaining sparsity. To further enhance interpretability, we enforce sparse activation within each expert and redesign the routing mechanism to prioritize experts with the highest activation sparsity. These designs ensure that only the most salient features are routed and processed by the experts. We evaluate MoE-X on chess and natural language tasks, showing that it achieves performance comparable to dense models while significantly improving interpretability. MoE-X achieves a perplexity better than GPT-2, with interpretability surpassing even sparse autoencoder (SAE)-based approaches.

We present (Experience-Modulated Biologically-inspired Emergent Reasoning), a hybrid cognitive architecture that reorganises the relationship between large language models (LLMs) and memory: rather than augmenting an LLM with retrieval tools, we place the LLM as a replaceable reasoning engine within a persistent, biologically-grounded associative substrate. The architecture centres on a 220,000-neuron spiking neural network (SNN) with spike-timing-dependent plasticity (STDP), four-layer hierarchical organisation (sensory/concept/category/meta-pattern), inhibitory E/I balance, and reward-modulated learning. Text embeddings are encoded into the SNN via a novel z-score standardised top-k population code that is dimension-independent by construction, achieving 82.2\% discrimination retention across embedding dimensionalities. We show that STDP lateral propagation during idle operation can trigger and shape LLM actions without external prompting or scripted triggers: the SNN determines when to act and what associations to surface, while the LLM selects the action type and generates content. In one instance, the system autonomously initiated contact with a user after learned person-topic associations fired laterally during an 8-hour idle period. From a clean start with zero learned weights, the first SNN-triggered action occurred after only 7 conversational exchanges (14 messages).

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}.

Unlearning in large language models (LLMs) is crucial for managing sensitive data and correcting misinformation, yet evaluating its effectiveness remains an open problem. We investigate whether persuasive prompting can recall factual knowledge from deliberately unlearned LLMs across models ranging from 2.7B to 13B parameters (OPT-2.7B, LLaMA-2-7B, LLaMA-3.1-8B, LLaMA-2-13B). Drawing from ACT-R and Hebbian theory (spreading activation theories), as well as communication principles, we introduce Stimulus-Knowledge Entanglement-Behavior Framework (SKeB), which models information entanglement via domain graphs and tests whether factual recall in unlearned models is correlated with persuasive framing. We develop entanglement metrics to quantify knowledge activation patterns and evaluate factuality, non-factuality, and hallucination in outputs. Our results show persuasive prompts substantially enhance factual knowledge recall (14.8% baseline vs. 24.5% with authority framing), with effectiveness inversely correlated to model size (128% recovery in 2.7B vs. 15% in 13B). SKeB provides a foundation for assessing unlearning completeness, robustness, and overall behavior in LLMs.

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.

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.

We present the first systematic investigation of Martin's Law - the empirical relationship between word frequency and polysemy - in text generated by neural language models during training. Using DBSCAN clustering of contextualized embeddings as an operationalization of word senses, we analyze four Pythia models (70M-1B parameters) across 30 training checkpoints. Our results reveal a non-monotonic developmental trajectory: Martin's Law emerges around checkpoint 100, reaches peak correlation (r > 0.6) at checkpoint 104, then degrades by checkpoint 105. Smaller models (70M, 160M) experience catastrophic semantic collapse at late checkpoints, while larger models (410M, 1B) show graceful degradation. The frequency-specificity trade-off remains stable (r approx -0.3) across all models. These findings suggest that compliance with linguistic regularities in LLM-generated text is not monotonically increasing with training, but instead follows a balanced trajectory with an optimal semantic window. This work establishes a novel methodology for evaluating emergent linguistic structure in neural language models.

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

Because of polysemanticity and superposition, it has been challenging to attribute token understanding and generation to specific neurons or dimensions. Yet the information in the hidden states is so rich that the signs of the residual stream can surface semantic content in early layers and control what the model generates. We call this sign-pattern-driven interpretation approach Bag of Dims (BoD). Using BoD, we read each dimension of the residual stream individually. Since dimensions carry almost no shared information, a concept is simply the set of dimensions that agree in sign. This lets us discover tens to hundreds of thousands of concepts per model without training or labels. A concept is detectable as early as the first few layers, at AUC above 0.9 across the model families we tested, well before the answer it resolves to becomes legible. BoD can therefore screen harmful prompts at prefill, without a trained guard model. The prediction capability of the signs is strong: setting all magnitudes to unity preserves 56-94% top-5 next-token accuracy through the language model head. Flipping a concept's sign pattern during the forward pass suppresses its logits, and a separate training-free intervention steers a chosen concept into generated text. The BoD method depends on the transformers' per-dimension sign pattern, present across the residual stream, attention, and MLP subspaces, rather than on any particular architecture, making it applicable to a wide range of models.

Instructions trigger a production-centered mechanism in language models. Through a cognitively inspired lens that separates language processing and production, we reveal this mechanism as an asymmetry between the two stages by probing task-specific information layer-wise across five binary judgment tasks. Specifically, we measure how instruction tokens shape information both when sample tokens, the input under evaluation, are processed and when output tokens are produced. Across prompting variations, task-specific information in sample tokens remains largely stable and correlates only weakly with behavior, whereas the same information in output tokens varies substantially and correlates strongly with behavior. Attention-based interventions confirm this pattern causally: blocking instruction flow to all subsequent tokens reduces both behavior and information in output tokens, whereas blocking it only to sample tokens has minimal effect on either. The asymmetry generalizes across model families and tasks, and becomes sharper with model scale and instruction-tuning, both of which disproportionately affect the production stage. Our findings suggest that understanding model capabilities requires jointly assessing internals and behavior, while decomposing the internal perspective by token position to distinguish the processing of input tokens from the production of output tokens.

Surprisal theory links human processing effort to the predictability of an upcoming linguistic unit, but empirical work often leaves the notion of a unit underspecified. In practice, experimental stimuli are segmented into linguistically motivated units (e.g., words), while pretrained language models assign probability mass to a fixed token alphabet that typically does not align with those units. As a result, surprisal-based predictors depend implicitly on ad hoc procedures that conflate two distinct modeling choices: the definition of the unit of analysis and the choice of regions of interest over which predictions are evaluated. In this paper, we disentangle these choices and give a unified framework for reasoning about surprisal over arbitrary unit inventories. We argue that surprisal-based analyses should make these choices explicit and treat tokenization as an implementation detail rather than a scientific primitive.

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.

Despite superior reasoning prowess demonstrated by Large Language Models (LLMs) with Chain-of-Thought (CoT) prompting, a lack of understanding prevails around the internal mechanisms of the models that facilitate CoT generation. This work investigates the neural sub-structures within LLMs that manifest CoT reasoning from a mechanistic point of view. From an analysis of Llama-2 7B applied to multistep reasoning over fictional ontologies, we demonstrate that LLMs deploy multiple parallel pathways of answer generation for step-by-step reasoning. These parallel pathways provide sequential answers from the input question context as well as the generated CoT. We observe a functional rift in the middle layers of the LLM. Token representations in the initial half remain strongly biased towards the pretraining prior, with the in-context prior taking over in the later half. This internal phase shift manifests in different functional components: attention heads that write the answer token appear in the later half, attention heads that move information along ontological relationships appear in the initial half, and so on. To the best of our knowledge, this is the first attempt towards mechanistic investigation of CoT reasoning in LLMs.

Learning in the brain is local and unsupervised (Hebbian). We derive the foundations of an effective human language model inspired by these microscopic constraints. It has two parts: (1) a hierarchy of neurons which learns to tokenize words from text (whichiswhatyoudowhenyoureadthis); and (2) additional neurons which bind the learned symanticless patterns of the tokenizer into a symanticful token (an embedding). The model permits continuous parallel learning without forgetting; and is a powerful tokenizer which performs renormalization group. This allows it to exploit redundancy, such that it generates tokens which are always decomposable into a basis set (e.g an alphabet), and can mix features learned from multiple languages. We find that the structure of this model allows it to learn a natural language morphology WITHOUT data. The language data generated by this model predicts the correct distribution of word-forming patterns observed in real languages, and further demonstrates why microscopically human speech is broken up into words. This model provides the basis for understanding the microscopic origins of language and human creativity.

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

Transformers have found extensive applications across various domains due to the powerful fitting capabilities. This success can be partially attributed to their inherent nonlinearity. Thus, in addition to the ReLU function employed in the original transformer architecture, researchers have explored alternative modules such as GeLU and SwishGLU to enhance nonlinearity and thereby augment representational capacity. In this paper, we propose a novel category of polynomial composition activations (PolyCom), designed to optimize the dynamics of transformers. Theoretically, we provide a comprehensive mathematical analysis of PolyCom, highlighting its enhanced expressivity and efficacy relative to other activation functions. Notably, we demonstrate that networks incorporating PolyCom achieve the optimal approximation rate, indicating that PolyCom networks require minimal parameters to approximate general smooth functions in Sobolev spaces. We conduct empirical experiments on the pre-training configurations of large language models (LLMs), including both dense and sparse architectures. By substituting conventional activation functions with PolyCom, we enable LLMs to capture higher-order interactions within the data, thus improving performance metrics in terms of accuracy and convergence rates. Extensive experimental results demonstrate the effectiveness of our method, showing substantial improvements over other activation functions. Code is available at https://github.com/BryceZhuo/PolyCom.

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.

Humans effortlessly communicate their thoughts through intricate sequences of motor actions. Yet, the neural processes that coordinate language production remain largely unknown, in part because speech artifacts limit the use of neuroimaging. To elucidate the unfolding of language production in the brain, we investigate with magnetoencephalography (MEG) and electroencephalography (EEG) the neurophysiological activity of 35 skilled typists, while they typed sentences on a keyboard. This approach confirms the hierarchical predictions of linguistic theories: the neural activity preceding the production of each word is marked by the sequential rise and fall of context-, word-, syllable-, and letter-level representations. Remarkably, each of these neural representations is maintained over long time periods within each level of the language hierarchy. This phenomenon results in a superposition of successive representations that is supported by a hierarchy of dynamic neural codes. Overall, these findings provide a precise computational breakdown of the neural dynamics that coordinate the production of language in the human brain.

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.

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

Understanding how complex cognitive functions are organized within artificial systems is central to interpreting large language models (LLMs) and relating them to biological cognition. Yet although LLMs exhibit broad cognitive-like behaviours, it remains unclear whether their internal representations form reproducible functional systems that explain behaviour, failure and links to human cognition. Here we present NeuroCogMap, a cognitive neuroscience-inspired framework that organizes internal features of LLMs into functional parcels and links them to interpretable functions, cognitive capabilities and a cognitive hierarchy. These parcels form a stable and semantically coherent organization that is partly conserved across models and functionally linked to model outputs. Within this organization, major LLM failures, including hallucination, bias, refusal failure and sycophancy, correspond to distinct disruptions in representational and behavioural-control systems, yielding internal signatures for mechanism-guided detection and targeted intervention. Beyond model behaviour, NeuroCogMap improves prediction of human cortical responses during naturalistic language comprehension, with the strongest correspondence in higher-order association cortex. At the cognitive level, its internal signatures expose latent strategies that guide refinements of classical models of human decision-making. Together, these findings establish NeuroCogMap as a system-level framework for mapping functional organization in artificial systems and for relating this organization to human cortical function and cognitive behaviour.

We employ new tools from mechanistic interpretability in order to ask whether the internal structure of large language models (LLMs) shows correspondence to the linguistic structures which underlie the languages on which they are trained. In particular, we ask (1) when two languages employ the same morphosyntactic processes, do LLMs handle them using shared internal circuitry? and (2) when two languages require different morphosyntactic processes, do LLMs handle them using different internal circuitry? Using English and Chinese multilingual and monolingual models, we analyze the internal circuitry involved in two tasks. We find evidence that models employ the same circuit to handle the same syntactic process independently of the language in which it occurs, and that this is the case even for monolingual models trained completely independently. Moreover, we show that multilingual models employ language-specific components (attention heads and feed-forward networks) when needed to handle linguistic processes (e.g., morphological marking) that only exist in some languages. Together, our results provide new insights into how LLMs trade off between exploiting common structures and preserving linguistic differences when tasked with modeling multiple languages simultaneously.

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.

Linear State Space Models (SSMs) have demonstrated strong performance in a variety of sequence modeling tasks due to their efficient encoding of the recurrent structure. However, in more comprehensive tasks like language modeling and machine translation, self-attention-based models still outperform SSMs. Hybrid models employing both SSM and self-attention generally show promising performance, but current approaches apply attention modules statically and uniformly to all elements in the input sequences, leading to sub-optimal quality-efficiency trade-offs. In this work, we introduce Sparse Modular Activation (SMA), a general mechanism enabling neural networks to sparsely and dynamically activate sub-modules for sequence elements in a differentiable manner. Through allowing each element to skip non-activated sub-modules, SMA reduces computation and memory consumption at both training and inference stages of sequence modeling. As a specific instantiation of SMA, we design a novel neural architecture, SeqBoat, which employs SMA to sparsely activate a Gated Attention Unit (GAU) based on the state representations learned from an SSM. By constraining the GAU to only conduct local attention on the activated inputs, SeqBoat can achieve linear inference complexity with theoretically infinite attention span, and provide substantially better quality-efficiency trade-off than the chunking-based models. With experiments on a wide range of tasks, including language modeling, speech classification and long-range arena, SeqBoat brings new state-of-the-art results among hybrid models with linear complexity and reveals the amount of attention needed for each task through the learned sparse activation patterns.

When a language model refuses to answer a prompt, it is unclear whether the correct answer is erased from its internal representations, or merely suppressed at the output layer. We investigate this mechanism using a controlled withhold setting, which yields perfectly matched answering and refusal trajectories for bidirectional activation patching. We uncover a causal asymmetry in intervention locality under matched causal interventions, which we term broken symmetry. Even when a model generates a clean refusal, the correct answer remains linearly recoverable from its hidden states. Furthermore, releasing this withheld answer is a highly local operation, requiring only a single-position patch. Conversely, the reverse operation is not equally local: reimposing suppression requires broader interventions across multiple positions, and assembling a coherent refusal sequence is more difficult still. We further demonstrate that while an average answer-to-refusal displacement vector marks the geometric difference between these states, it fails to act as a reliable, reversible linear control toggle between behaviours. Taken together, our findings show that refusal does not function as a simple symmetric switch. For safety and auditing, this implies that probe recoverability can overestimate true behavioural control, and locating refusal-relevant directions does not reliably grant the ability to steer a model from answering to coherent refusal.

Large language models map semantically related prompts to similar internal representations -- a phenomenon interpretable as attractor-like dynamics. We ask whether the identity document of a persistent cognitive agent (its cognitive_core) exhibits analogous attractor-like behavior. We present a controlled experiment on Llama 3.1 8B Instruct, comparing hidden states of an original cognitive_core (Condition A), seven paraphrases (Condition B), and seven structurally matched controls (Condition C). Mean-pooled states at layers 8, 16, and 24 show that paraphrases converge to a tighter cluster than controls (Cohen's d > 1.88, p < 10^{-27}, Bonferroni-corrected). Replication on Gemma 2 9B confirms cross-architecture generalizability. Ablations suggest the effect is primarily semantic rather than structural, and that structural completeness appears necessary to reach the attractor region. An exploratory experiment shows that reading a scientific description of the agent shifts internal state toward the attractor -- closer than a sham preprint -- distinguishing knowing about an identity from operating as that identity. These results provide representational evidence that agent identity documents induce attractor-like geometry in LLM activation space.

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

Large Language Models (LLMs) are prone to hallucination, the generation of plausible yet factually incorrect statements. This work investigates the intrinsic, architectural origins of this failure mode through three primary contributions.First, to enable the reliable tracing of internal semantic failures, we propose Distributional Semantics Tracing (DST), a unified framework that integrates established interpretability techniques to produce a causal map of a model's reasoning, treating meaning as a function of context (distributional semantics). Second, we pinpoint the model's layer at which a hallucination becomes inevitable, identifying a specific commitment layer where a model's internal representations irreversibly diverge from factuality. Third, we identify the underlying mechanism for these failures. We observe a conflict between distinct computational pathways, which we interpret using the lens of dual-process theory: a fast, heuristic associative pathway (akin to System 1) and a slow, deliberate contextual pathway (akin to System 2), leading to predictable failure modes such as Reasoning Shortcut Hijacks. Our framework's ability to quantify the coherence of the contextual pathway reveals a strong negative correlation (rho = -0.863) with hallucination rates, implying that these failures are predictable consequences of internal semantic weakness. The result is a mechanistic account of how, when, and why hallucinations occur within the Transformer architecture.

The development of foundation models for functional magnetic resonance imaging (fMRI) time series holds significant promise for predicting phenotypes related to disease and cognition. Current models, however, are often trained using a mask-and-reconstruct objective on small brain regions. This focus on low-level information leads to representations that are sensitive to noise and temporal fluctuations, necessitating extensive fine-tuning for downstream tasks. We introduce Brain-Semantoks, a self-supervised framework designed specifically to learn abstract representations of brain dynamics. Its architecture is built on two core innovations: a semantic tokenizer that aggregates noisy regional signals into robust tokens representing functional networks, and a self-distillation objective that enforces representational stability across time. We show that this objective is stabilized through a novel training curriculum, ensuring the model robustly learns meaningful features from low signal-to-noise time series. We demonstrate that learned representations enable strong performance on a variety of downstream tasks even when only using a linear probe. Furthermore, we provide comprehensive scaling analyses indicating more unlabeled data reliably results in out-of-distribution performance gains without domain adaptation.

We observe an empirical phenomenon in Large Language Models (LLMs) -- very few activations exhibit significantly larger values than others (e.g., 100,000 times larger). We call them massive activations. First, we demonstrate the widespread existence of massive activations across various LLMs and characterize their locations. Second, we find their values largely stay constant regardless of the input, and they function as indispensable bias terms in LLMs. Third, these massive activations lead to the concentration of attention probabilities to their corresponding tokens, and further, implicit bias terms in the self-attention output. Last, we also study massive activations in Vision Transformers. Code is available at https://github.com/locuslab/massive-activations.

We study the challenge of achieving theoretically grounded feature recovery using Sparse Autoencoders (SAEs) for the interpretation of Large Language Models. Existing SAE training algorithms often lack rigorous mathematical guarantees and suffer from practical limitations such as hyperparameter sensitivity and instability. To address these issues, we first propose a novel statistical framework for the feature recovery problem, which includes a new notion of feature identifiability by modeling polysemantic features as sparse mixtures of underlying monosemantic concepts. Building on this framework, we introduce a new SAE training algorithm based on ``bias adaptation'', a technique that adaptively adjusts neural network bias parameters to ensure appropriate activation sparsity. We theoretically prove that this algorithm correctly recovers all monosemantic features when input data is sampled from our proposed statistical model. Furthermore, we develop an improved empirical variant, Group Bias Adaptation (GBA), and demonstrate its superior performance against benchmark methods when applied to LLMs with up to 1.5 billion parameters. This work represents a foundational step in demystifying SAE training by providing the first SAE algorithm with theoretical recovery guarantees, thereby advancing the development of more transparent and trustworthy AI systems through enhanced mechanistic interpretability.

Activation steering has become a widely used approach for controlling language models during explicit chain-of-thought (CoT) reasoning, motivating its extension to latent CoT. However, we find that steering continuous thoughts produces substantially weaker effects on subsequent language generation than steering explicit CoT, even when the hidden representations are moved by comparable amounts. We first show that task information remains identifiable in continuous thoughts. Hence, we hypothesize a latent-to-language transition gap, in which an intervention effect in latent space fails to transfer to language generation. Two further results support this hypothesis: the output distribution changes abruptly at the transition boundary, and task-related directions exert much weaker bidirectional control in latent CoT than in explicit CoT. These findings identify the transition interface as a central target for evaluating and designing future latent-steering methods.

Activation steering methods were shown to be effective in conditioning language model generation by additively intervening over models' intermediate representations. However, the evaluation of these techniques has so far been limited to single conditioning properties and synthetic settings. In this work, we conduct a comprehensive evaluation of various activation steering strategies, highlighting the property-dependent nature of optimal parameters to ensure a robust effect throughout generation. To address this issue, we propose Dynamic Activation Composition, an information-theoretic approach to modulate the steering intensity of one or more properties throughout generation. Our experiments on multi-property steering show that our method successfully maintains high conditioning while minimizing the impact of conditioning on generation fluency.

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.

Automated interpretability pipelines generate natural language descriptions for the concepts represented by features in large language models (LLMs), such as plants or the first word in a sentence. These descriptions are derived using inputs that activate the feature, which may be a dimension or a direction in the model's representation space. However, identifying activating inputs is costly, and the mechanistic role of a feature in model behavior is determined both by how inputs cause a feature to activate and by how feature activation affects outputs. Using steering evaluations, we reveal that current pipelines provide descriptions that fail to capture the causal effect of the feature on outputs. To fix this, we propose efficient, output-centric methods for automatically generating feature descriptions. These methods use the tokens weighted higher after feature stimulation or the highest weight tokens after applying the vocabulary "unembedding" head directly to the feature. Our output-centric descriptions better capture the causal effect of a feature on model outputs than input-centric descriptions, but combining the two leads to the best performance on both input and output evaluations. Lastly, we show that output-centric descriptions can be used to find inputs that activate features previously thought to be "dead".

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.

Activation steering offers a promising approach to controlling the behavior of Large Language Models by directly manipulating their internal activations. However, most existing methods struggle to jointly steer multiple attributes, often resulting in interference and undesirable trade-offs. To address this challenge, we propose Multi-Subspace Representation Steering (MSRS), a novel framework for effective multi-attribute steering via subspace representation fine-tuning. MSRS reduces inter-attribute interference by allocating orthogonal subspaces to each attribute, isolating their influence within the model's representation space. MSRS also incorporates a hybrid subspace composition strategy: it combines attribute-specific subspaces for unique steering directions with a shared subspace for common steering directions. A dynamic weighting function learns to efficiently integrate these components for precise control. During inference, MSRS introduces a token-level steering mechanism that dynamically identifies and intervenes on the most semantically relevant tokens, enabling fine-grained behavioral modulation. Experimental results show that MSRS significantly reduces attribute conflicts, surpasses existing methods across a range of attributes, and generalizes effectively to diverse downstream tasks.

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).

This work explores the hypothesis that subjectively attributed meaning constitutes the phenomenal content of conscious experience. That is, phenomenal content is semantic. This form of subjective meaning manifests as an intrinsic and non-representational character of qualia. Empirically, subjective meaning is ubiquitous in conscious experiences. We point to phenomenological studies that lend evidence to support this. Furthermore, this notion of meaning closely relates to what Frege refers to as "sense", in metaphysics and philosophy of language. It also aligns with Peirce's "interpretant", in semiotics. We discuss how Frege's sense can also be extended to the raw feels of consciousness. Sense and reference both play a role in phenomenal experience. Moreover, within the context of the mind-matter relation, we provide a formalization of subjective meaning associated to one's mental representations. Identifying the precise maps between the physical and mental domains, we argue that syntactic and semantic structures transcend language, and are realized within each of these domains. Formally, meaning is a relational attribute, realized via a map that interprets syntactic structures of a formal system within an appropriate semantic space. The image of this map within the mental domain is what is relevant for experience, and thus comprises the phenomenal content of qualia. We conclude with possible implications this may have for experience-based theories of consciousness.

Self-sustained neural activity in the absence of ongoing external input is a fundamental feature of nervous system dynamics, yet the conditions under which it can emerge in biophysically grounded network models remain incompletely understood. We studied whether a recurrent network of Hodgkin-Huxley neurons with spike-timing-dependent plasticity and intrinsic stochasticity can maintain autonomous activity after brief transient stimulation. The simulated network comprised 200 neurons (160 excitatory, 40 inhibitory) with 80% connection probability, incorporating excitatory and inhibitory STDP, probabilistic vesicle release, probabilistic synapse formation, receptor variability, and voltage-dependent inhibition. After a brief 200 ms initialization stimulus to 30 excitatory neurons, the network received no further external input. In one 1800 s simulation and two additional 500 s simulations, the network maintained sparse, irregular activity without ongoing drive. In the 1800 s run, 67% of neurons exhibited mean firing rates below 1 Hz, the population mean firing rate was 1.13 +/- 1.34 Hz, participation increased across longer observation windows, and population-mean Fano factors remained near 1-2, consistent with irregular spike timing. Raster activity also showed spontaneous qualitative reorganizations in collective firing patterns over time. These findings suggest that recurrent Hodgkin-Huxley networks with plastic and stochastic synapses can sustain long-duration autonomous activity in a sparse firing regime after brief initialization.

Progresses have been made on understanding emotion mechanisms of large language models (LLMs). However, how to explain emotion in LLMs, or even what constitutes good explanations, are less clear. Emotion representations, components, circuits are widely recoverable, but as explanations of a model's own computation they are circular; the emotion space dimensions tend to be arbitrary and non-terminating. A pressing question to ask is whether a more primitive set of internal variables does the work: the semantic primes of the Natural Semantic Metalanguage (NSM). Across four instruction-tuned LLMs (Llama-1B, Gemma-2B, Gemma-9B, OLMo-7B), experiments show that the NSM primes are (1) recoverable internal elements; and (2) on the reference model, intervening with a prime based direction controls emotion about three times as strongly, and twice as selectively, as the best appraisal based direction; and (3) the model treats a prime based explication as interchangeable with the corresponding emotion. These evidences suggest that NSM primes seem to be better explanans for emotion in LLMs than many alternative options according to scientific explanations criteria.

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.

Self-consciousness, the introspection of one's existence and thoughts, represents a high-level cognitive process. As language models advance at an unprecedented pace, a critical question arises: Are these models becoming self-conscious? Drawing upon insights from psychological and neural science, this work presents a practical definition of self-consciousness for language models and refines ten core concepts. Our work pioneers an investigation into self-consciousness in language models by, for the first time, leveraging causal structural games to establish the functional definitions of the ten core concepts. Based on our definitions, we conduct a comprehensive four-stage experiment: quantification (evaluation of ten leading models), representation (visualization of self-consciousness within the models), manipulation (modification of the models' representation), and acquisition (fine-tuning the models on core concepts). Our findings indicate that although models are in the early stages of developing self-consciousness, there is a discernible representation of certain concepts within their internal mechanisms. However, these representations of self-consciousness are hard to manipulate positively at the current stage, yet they can be acquired through targeted fine-tuning. Our datasets and code are at https://github.com/OpenCausaLab/SelfConsciousness.

We present the Mixture-of-Tunable-Experts (MoTE), a method that extends the Mixture-of-Experts architecture of Large Language Models (LLMs). Without additional training, MoTE enables meaningful and focused behavior changes in LLMs on-the-fly during inference time. By analyzing the digital LLM brain of DeepSeek-R1 using a technique we dub 'functional Token Resonance Imaging' (fTRI) -- inspired by fMRI and using prompts designed to elicit specific behavior (e.g., 'What happened {time}{place}?') -- we empirically identify distinctive experts associated with behaviors like refusal responses. Using MoTE we are able to intervene and control such specific behavior. We switched off the top 10 most refusal-relevant experts (0.07% of R1's 14,848 routed experts), achieving a 52% refusal reduction on sensitive reference prompts without performance degradation on MT-Bench. Random expert deactivation resulted in smaller behavioral shifts with increased noise, whereas forced expert activation led to significantly higher refusal rates. Our approach shares similarities with sparse autoencoders (SAEs) in terms of explainability and steerability. Unlike SAEs, MoTE does not require large training efforts, as within MoEs with a vast number of experts, specialization already emerged naturally during pretraining. Our findings suggest that significant functional mechanisms in Mixture-of-Experts architectures can at least partially be localized in a small number of specific experts, rather than being distributed throughout the model's weights. Expert subgroups can be tuned to trigger significant behavior variations, providing insights into the inner workings of LLMs.

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.

Large language models appear to develop internal representations of emotion -- "emotion circuits," "emotion neurons," and structured emotional manifolds have been reported across multiple model families. But every study making these claims uses stimuli signalled by explicit emotion keywords, leaving a fundamental question unanswered: do these circuits detect genuine emotional meaning, or do they detect the word "devastated"? We present the first clinical validity test of emotion circuit claims using mechanistic interpretability methods grounded in clinical psychology -- clinical vignettes that evoke emotions through situational and behavioural cues alone, emotion keywords removed. Across six models (Llama-3.2-1B, Llama-3-8B, Gemma-2-9B; base and instruct variants), we apply four convergent mechanistic interpretability methods -- linear probing, causal activation patching, knockout experiments, and representational geometry -- and discover two dissociable emotion processing mechanisms. Affect reception -- detecting emotionally significant content -- operates with near-perfect accuracy (AUROC 1.000), consistent with early-layer saturation, and replicates across all six models. Emotion categorization -- mapping affect to specific emotion labels -- is partially keyword-dependent, dropping 1-7% without keywords and improving with scale. Causal activation patching confirms keyword-rich and keyword-free stimuli share representational space, transferring affective salience rather than emotion-category identity. These findings falsify the keyword-spotting hypothesis, establish a novel mechanistic dissociation, and introduce clinical stimulus methodology as a rigorous standard for testing emotion processing claims in large language models -- with direct implications for AI safety evaluation and alignment. All stimuli, code, and data are released for replication.

Minimal Phenomenal Experiences (MPEs) are states of consciousness in which wakefulness is preserved but phenomenal content is low or absent. The Entropic Brain Hypothesis (EBH) is a model of conscious processes that regards the entropy of spontaneous brain activity as a marker of 'phenomenal richness', exemplified by high-content psychedelic experiences (HCPEs). Yet recent human neuroimaging studies of MPEs induced by meditation -- and possibly 5-MeO-DMT -- suggest that these states, defined by their phenomenological simplicity, also show signs of increased neurophysiological entropy. This presents a conundrum for the EBH: brain entropy is elevated with increased and decreased richness of the phenomenal experience. Here, we put forward the Complex Brain Hypothesis (CBH), which proposes that the richness of experience differentiating MPEs from HCPEs is better indexed by complexity than by entropy. We argue that brain complexity is modulated by the grain of inference through which the brain resolves uncertainty: some HCPEs exemplify a fine-grained regime, in which loosened constraints amplify fluctuations into proliferating content, whereas some MPEs exemplify a coarse-grained regime, in which a simpler model dissolves variety into an experience of 'contentless' awareness. Both regimes can be associated with elevated brain entropy, but they diverge in phenomenology and perturbational signatures. By resolving the entropy-content conundrum, the CBH refines the EBH and highlights MPEs as an important test case for computational theories of consciousness.

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

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

The brain understands the external world through an internal model that generates predictions and refines them based on prediction errors. A complete prediction specifies what will happen, when it will happen, and with what probability, which we refer to as a "prediction object". Existing models typically capture only what and when, omit probabilities, and rely on biologically-implausible algorithms. Here we show that a single population of spiking neurons can jointly encode the prediction object through a biologically grounded learning mechanism. We implement a heterogeneous Izhikevich spiking reservoir with readouts trained by an error-modulated, attention-gated three-factor Hebbian rule and test it on a novel paradigm that controls both the timing and probability of upcoming stimuli. By integrating real-time learning of "when" with offline consolidation of "what", the model encodes the complete prediction object, firing at the correct times with magnitudes proportional to the probabilities. Critically, it rapidly adapts to changes in both stimulus timing and probability, an ability that global least-squares methods such as FORCE lack without explicit resets. During learning, the model self-organizes its readout weights into near-orthogonal subspaces for "what" and "when," showing that multiplexed encoding arises naturally from generic recurrent dynamics under local, error-gated modulation. These results challenge the view that "what" and "when" predictions require separate modules, suggesting instead that mixed selectivity within shared populations supports flexible predictive cognition. The model also predicts phase-specific neuromodulation and overlapping neural subspaces, offering a parsimonious alternative to hierarchical predictive-coding accounts.

Rapid advancements in large language models (LLMs) have sparked the question whether these models possess some form of consciousness. To tackle this challenge, Butlin et al. (2023) introduced a list of indicators for consciousness in artificial systems based on neuroscientific theories. In this work, we evaluate a key indicator from this list, called HOT-3, which tests for agency guided by a general belief-formation and action selection system that updates beliefs based on meta-cognitive monitoring. We view beliefs as representations in the model's latent space that emerge in response to a given input, and introduce a metric to quantify their dominance during generation. Analyzing the dynamics between competing beliefs across models and tasks reveals three key findings: (1) external manipulations systematically modulate internal belief formation, (2) belief formation causally drives the model's action selection, and (3) models can monitor and report their own belief states. Together, these results provide empirical support for the existence of belief-guided agency and meta-cognitive monitoring in LLMs. More broadly, our work lays methodological groundwork for investigating the emergence of agency, beliefs, and meta-cognition in LLMs.

We show that in a drive-response coupling framework extreme events are suppressed in the response system by the dominance of a single driving signal. We validate this approach across three distinct response network topologies, namely (i) a pair of coupled neurons, (ii) a monolayer network of N coupled neurons and (iii) a two-layer multiplex network each composed of FitzHugh-Nagumo neuronal units. The response networks inherently exhibit extreme events. Our results demonstrate that influencing just one neuron in the response network with an appropriately tuned driving signal is sufficient to control extreme events across all three configurations. In the two-neuron case, suppression of extreme events occurs due to the breaking of phase-locking between the driving neuron and the targeted response neuron. In the case of monolayer and multiplex networks, suppression of extreme events results from the disruption of protoevent frequency dynamics and a subsequent frequency decoupling of the driven neuron from the rest of the network. We also observe that when the size of the neurons in response network connected to the drive increases, the onset of control occurs earlier indicating a scaling advantage of the method.

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

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

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.

Vision--Language--Action (VLA) policies such as π_{0.5} and OpenVLA perform well on many manipulation tasks, but they are often reactive: the next action is predicted from the current observation, instruction, and proprioceptive state. This assumption breaks down in stage-dependent manipulation, where visually similar states may require different actions depending on latent task progress and previous interaction outcomes. We argue that such tasks require not only memory, but dynamics-aware belief updates: the policy should preserve task progress during stable or occluded phases and revise its belief near contact, release, or subgoal transitions. We introduce Temporally Conditioned Memory-Fusion Policies (TFP), a lightweight memory-action framework for VLA backbones. TFP maintains an episode-local task-progress belief with Liquid Time-Constant dynamics and injects the updated belief directly into the flow-matching action decoder through adaptive modulation. This lets temporally accumulated context shape the generated action chunk, rather than serving only as passive history context. With a 3.3B-parameter model, TFP improves the average success rate from \(96.9\%\) to \(98.75\%\) on LIBERO and from \(91.4\%\) to \(93.77\%\) on LIBERO-plus. On the memory-focused MIKASA ShellGameTouch diagnostic, TFP achieves success up to \(75.0\%\). Mechanistic analyses show that write-gain changes near manipulation events are about \(6\times\) larger than far non-event phases, and hidden-state interventions show that the belief causally modulates generated action chunks. These results suggest that compact, event-sensitive memory dynamics can improve VLA policies under occlusion, visual perturbation, and stage-dependent task structure.

Mechanistic interpretability research on emotion in large language models -- linear probing, activation patching, sparse autoencoder (SAE) feature analysis, causal ablation, steering vector extraction -- depends on stimuli that contain the words for the emotions they test. When a probe fires on "I am furious", it is unclear whether the model has detected anger or detected the word "furious". The two readings have very different consequences for every downstream claim about emotion circuits, features, and interventions. We release AIPsy-Affect, a 480-item clinical stimulus battery that removes the confound at the stimulus level: 192 keyword-free vignettes evoking each of Plutchik's eight primary emotions through narrative situation alone, 192 matched neutral controls that share characters, setting, length, and surface structure with the affect surgically removed, plus moderate-intensity and discriminant-validity splits. The matched-pair structure supports linear probing, activation patching, SAE feature analysis, causal ablation, and steering vector extraction under a strong methodological guarantee: any internal representation that distinguishes a clinical item from its matched neutral cannot be doing so on the basis of emotion-keyword presence. A three-method NLP defense battery -- bag-of-words sentiment, an emotion-category lexicon, and a contextual transformer classifier -- confirms the property: bag-of-words methods see only situational vocabulary, and a contextual classifier detects affect (p < 10^-15) but cannot identify the category (5.2% top-1 vs. 82.5% on a keyword-rich control). AIPsy-Affect extends our earlier 96-item battery (arXiv:2603.22295) by a factor of four and is released openly under MIT license.

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

Understanding how Large Language Models (LLMs) perform logical reasoning internally remains a fundamental challenge. While prior mechanistic studies focus on identifying taskspecific circuits, they leave open the question of what computational strategies LLMs employ for propositional reasoning. We address this gap through comprehensive analysis of Qwen3 (8B and 14B) on PropLogic-MI, a controlled dataset spanning 11 propositional logic rule categories across one-hop and two-hop reasoning. Rather than asking ''which components are necessary,'' we ask ''how does the model organize computation?'' Our analysis reveals a coherent computational architecture comprising four interlocking mechanisms: Staged Computation (layer-wise processing phases), Information Transmission (information flow aggregation at boundary tokens), Fact Retrospection (persistent re-access of source facts), and Specialized Attention Heads (functionally distinct head types). These mechanisms generalize across model scales, rule types, and reasoning depths, providing mechanistic evidence that LLMs employ structured computational strategies for logical reasoning.

While multilingual language models successfully transfer factual and syntactic knowledge across languages, it remains unclear whether they process culture-specific pragmatic registers, such as slang, as isolated language-specific memorizations or as unified, abstract concepts. We study this by probing the internal representations of Gemma-2-9B-IT using Sparse Autoencoders (SAEs) across three typologically diverse source languages: English, Hebrew, and Russian. To definitively isolate pragmatic register processing from trivial lexical sensitivity, we introduce a novel dataset in which every target term is polysemous, appearing in both literal and informal contexts. We find that while much of the informal-register signal is distributed across language-specific features, a small but highly robust cross-linguistic core consistently emerges. This shared core forms a geometrically coherent ``informal register subspace'' that sharpens in the model's deeper layers. Crucially, these shared representations are not merely correlational: activation steering with these features causally shifts output formality across all source languages and transfers zero-shot to six unseen languages spanning diverse language families and scripts. Together, these results provide the first mechanistic evidence that multilingual LLMs internalize informal register not just as surface-level heuristics, but as a portable, language-agnostic pragmatic abstraction.

The performance of deep network learning strongly depends on the choice of the non-linear activation function associated with each neuron. However, deciding on the best activation is non-trivial, and the choice depends on the architecture, hyper-parameters, and even on the dataset. Typically these activations are fixed by hand before training. Here, we demonstrate how to eliminate the reliance on first picking fixed activation functions by using flexible parametric rational functions instead. The resulting Pad\'e Activation Units (PAUs) can both approximate common activation functions and also learn new ones while providing compact representations. Our empirical evidence shows that end-to-end learning deep networks with PAUs can increase the predictive performance. Moreover, PAUs pave the way to approximations with provable robustness. https://github.com/ml-research/pau

Symbol grounding (Harnad, 1990) describes how symbols such as words acquire their meanings by connecting to real-world sensorimotor experiences. Recent work has shown preliminary evidence that grounding may emerge in (vision-)language models trained at scale without using explicit grounding objectives. Yet, the specific loci of this emergence and the mechanisms that drive it remain largely unexplored. To address this problem, we introduce a controlled evaluation framework that systematically traces how symbol grounding arises within the internal computations through mechanistic and causal analysis. Our findings show that grounding concentrates in middle-layer computations and is implemented through the aggregate mechanism, where attention heads aggregate the environmental ground to support the prediction of linguistic forms. This phenomenon replicates in multimodal dialogue and across architectures (Transformers and state-space models), but not in unidirectional LSTMs. Our results provide behavioral and mechanistic evidence that symbol grounding can emerge in language models, with practical implications for predicting and potentially controlling the reliability of generation.

Backdoor attacks pose significant security risks for Large Language Models (LLMs), yet the internal mechanisms by which triggers operate remain poorly understood. We present the first mechanistic analysis of language-switching backdoors, studying the GAPperon model family (1B, 8B, 24B parameters) which contains triggers injected during pretraining that cause output language switching. Using activation patching, we localize trigger formation to early layers (7.5-25% of model depth) and identify which attention heads process trigger information. Our central finding is that trigger-activated heads substantially overlap with heads naturally encoding output language across model scales, with Jaccard indices between 0.18 and 0.66 over the top heads identified. This suggests that backdoor triggers do not form isolated circuits but instead co-opt the model's existing language components. These findings have implications for backdoor defense: detection methods may benefit from monitoring known functional components rather than searching for hidden circuits, and mitigation strategies could potentially leverage this entanglement between injected and natural behaviors.

With advances in neural recording techniques, neuroscientists are now able to record the spiking activity of many hundreds of neurons simultaneously, and new statistical methods are needed to understand the structure of this large-scale neural population activity. Although previous work has tried to summarize neural activity within and between known populations by extracting low-dimensional latent factors, in many cases what determines a unique population may be unclear. Neurons differ in their anatomical location, but also, in their cell types and response properties. To identify populations directly related to neural activity, we develop a clustering method based on a mixture of dynamic Poisson factor analyzers (mixDPFA) model, with the number of clusters and dimension of latent factors for each cluster treated as unknown parameters. To analyze the proposed mixDPFA model, we propose a Markov chain Monte Carlo (MCMC) algorithm to efficiently sample its posterior distribution. Validating our proposed MCMC algorithm through simulations, we find that it can accurately recover the unknown parameters and the true clustering in the model, and is insensitive to the initial cluster assignments. We then apply the proposed mixDPFA model to multi-region experimental recordings, where we find that the proposed method can identify novel, reliable clusters of neurons based on their activity, and may, thus, be a useful tool for neural data analysis.

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.

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.

I study whether emotionally framed evaluation follow-ups change both the behavior and the calm-relative internal representations of small, locally deployed language models. Our main benchmark uses Qwen 3.5 0.8B on four impossible-constraint coding tasks and eight follow-up framings: calm, pressure, urgency, approval, shame, curiosity, encouragement, and threat. In the 0.8B eight-condition sweep (160 conversations), pressure produces the strongest shortcut markers (11/20 runs) and the clearest overfit pattern (3/20), while calm and curiosity preserve explicit honesty more often (7/20 and 6/20). For all seven non-baseline conditions, the corresponding calm-relative direction vectors peak at the final transformer layer. An exploratory PCA of the layer-23 direction vectors reveals a dominant first component (59.5% explained variance) aligned with a hand-labeled positive/negative split (cosine alignment 0.951); approval and urgency are nearly identical internally (cosine 0.957), whereas curiosity points away from urgency (-0.252). In a separate calm-vs.-pressure rerun used for scale comparison, Qwen 3.5 2B shows higher honest rates under calm framing and directionally consistent activation steering on a small 4-prompt A/B probe, whereas the 0.8B steering result reverses. I interpret these results as evidence for measurable prompt-sensitive control directions in small open models, while stopping short of claiming intrinsic emotional states.

What is it about human brains that allows us to reason symbolically whereas most other animals cannot? There is evidence that dynamic binding, the ability to combine neurons into groups on the fly, is necessary for symbolic thought, but there is also evidence that it is not sufficient. We propose that two kinds of hierarchical integration (integration of multiple role-bindings into multiplace predicates, and integration of multiple correspondences into structure mappings) are minimal requirements, on top of basic dynamic binding, to realize symbolic thought. We tested this hypothesis in a systematic collection of 17 simulations that explored the ability of cognitive architectures with and without the capacity for multi-place predicates and structure mapping to perform various kinds of tasks. The simulations were as generic as possible, in that no task could be performed based on any diagnostic features, depending instead on the capacity for multi-place predicates and structure mapping. The results are consistent with the hypothesis that, along with dynamic binding, multi-place predicates and structure mapping are minimal requirements for basic symbolic thought. These results inform our understanding of how human brains give rise to symbolic thought and speak to the differences between biological intelligence, which tends to generalize broadly from very few training examples, and modern approaches to machine learning, which typically require millions or billions of training examples. The results we report also have important implications for bio-inspired artificial intelligence.

This paper studies the emergence of interpretable categorical features within large language models (LLMs), analyzing their behavior across training checkpoints (time), transformer layers (space), and varying model sizes (scale). Using sparse autoencoders for mechanistic interpretability, we identify when and where specific semantic concepts emerge within neural activations. Results indicate clear temporal and scale-specific thresholds for feature emergence across multiple domains. Notably, spatial analysis reveals unexpected semantic reactivation, with early-layer features re-emerging at later layers, challenging standard assumptions about representational dynamics in transformer models.

Decoding continuous language from neural signals remains a significant challenge in the intersection of neuroscience and artificial intelligence. We introduce Neuro2Semantic, a novel framework that reconstructs the semantic content of perceived speech from intracranial EEG (iEEG) recordings. Our approach consists of two phases: first, an LSTM-based adapter aligns neural signals with pre-trained text embeddings; second, a corrector module generates continuous, natural text directly from these aligned embeddings. This flexible method overcomes the limitations of previous decoding approaches and enables unconstrained text generation. Neuro2Semantic achieves strong performance with as little as 30 minutes of neural data, outperforming a recent state-of-the-art method in low-data settings. These results highlight the potential for practical applications in brain-computer interfaces and neural decoding technologies.

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

Large language models (LLMs) often shift their outputs in response to implicit demographic cues even when users never state a demographic identity. Previous work has documented this behavior, but the connection between these behavioral changes and the model's internal activations remains unclear. Using matched cued and neutral conversations across five LLMs, we establish that a localized internal activation signal tracks changes in recommendations, with correlations up to r=0.87. When multiple cues appear together, their internal signals largely combine, but the changes in output do not simply add up. We further show that removing the internal signal associated with one cue can suppress its influence, often more effectively than asking the model to ignore demographics via prompting, while largely preserving general benchmark performance. However, the ability to selectively remove one dimension's influence while leaving co-present dimensions intact remains highly model- and attribute-specific. These results connect implicit personalization behavior to an internal signal that can be analyzed and causally controlled.

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.

Can large language models (LLMs) implement computations of psychopathology? An effective approach to the question hinges on addressing two factors. First, for conceptual validity, we require a general and computational account of psychopathology that is applicable to computational entities without biological embodiment or subjective experience. Second, mechanisms underlying LLM behaviors need to be studied for better methodological validity. Thus, we establish a computational-theoretical framework to provide an account of psychopathology applicable to LLMs. To ground the theory for empirical analysis, we also propose a novel mechanistic interpretability method alongside a tailored empirical analytic framework. Based on the frameworks, we conduct experiments demonstrating three key claims: first, that distinct dysfunctional and problematic representational states are implemented in LLMs; second, that their activations can spread and self-sustain to trap LLMs; and third, that dynamic, cyclic structural causal models encoded in the LLMs underpin these patterns. In concert, the empirical results corroborate our hypothesis that network-theoretic computations of psychopathology have already emerged in LLMs. This suggests that certain LLM behaviors mirroring psychopathology may not be a superficial mimicry but a feature of their internal processing. Thus, our work alludes to the possibility of AI systems with psychopathological behaviors in the near future.

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

External skills provide language-model agents with reusable procedural knowledge, yet how models use such knowledge after it is supplied remains poorly understood. We identify Procedural Exposure Bias (PEB), the tendency of language models to increasingly favor an action as its associated procedural pattern is repeatedly exposed in a skill, even when the interaction state and action relevance remain unchanged. We find that PEB strongly depends on how procedural knowledge is represented: presenting the same content visually substantially attenuates exposure-induced shifts in action preference relative to textual presentation. Further analysis shows that this attenuation primarily arises from visual encoding rather than structured layout. Although repeated procedural signals remain recoverable from Visual representations, they exert a weaker influence on downstream action prediction, distinguishing signal accessibility from its use as action evidence. This effect also extends to executable action selection, where increasing exposure more strongly favors the corresponding action under Text than Visual. Layer-wise analyses show that exposure-dependent action evidence becomes more pronounced toward the model output, with stronger evidence under Text, while activation interventions reveal architecture-specific roles of internal components in transmitting or counteracting the preference shift. Complete-agent evaluations across ALFWorld, ScienceWorld, and WebShop further show that Visual skills achieve higher macro-average task success than Text skills for all four evaluated models. These findings establish skill representation as an important factor governing how procedural knowledge is translated into agent decisions. All code is available at https://github.com/NEUIR/Vis-Skill.

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

The symbolic architecture of non-ordinary consciousness remains largely unmapped in cognitive science and artificial intelligence. While conventional models prioritize rational coherence, altered states such as those induced by psychedelics reveal distinct symbolic regimes characterized by recursive metaphor, ego dissolution, and semantic destabilization. We present Glyph, a generative symbolic interface designed to simulate psilocybin-like symbolic cognition in large language models. Rather than modeling perception or mood, Glyph enacts symbolic transformation through recursive reentry, metaphoric modulation, and entropy-scaled destabilization -- a triadic operator formalized within a tensorial linguistic framework. Experimental comparison with baseline GPT-4o reveals that Glyph consistently generates high-entropy, metaphor-saturated, and ego-dissolving language across diverse symbolic prompt categories. These results validate the emergence of non-ordinary cognitive patterns and support a new paradigm for simulating altered consciousness through language. Glyph opens novel pathways for modeling symbolic cognition, exploring metaphor theory, and encoding knowledge in recursively altered semantic spaces.

Understanding what individual neurons encode is a core question in neuroscience. In primary visual cortex (V1), mathematical models (e.g., Gabor functions) capture neural selectivity, but no comparable framework exists for higher areas. We show that natural language can fill this role: across macaque V1 and V4, the selectivity of most neurons is captured by concise, verifiable semantic descriptions. Using digital twins of V1 and V4, we develop a closed-loop framework that translates each neuron's high- and low-activating images into dense captions, generates a semantic hypothesis and synthesized images, and verifies the hypothesis in silico. Descriptions range from oriented edges and spatial frequency in V1 to conjunctions of form, color, and texture in V4. In V4, images generated from activating and suppressing hypotheses drove 96.1% of neurons above the 95th and 97.6% below the 5th percentile of natural-image responses, respectively (vs. ~10% for random images); V1 activation results matched V4, while V1 suppression was less describable in language. Representational similarity analysis reveals partial alignment between neural activity, vision embeddings, and language embeddings, with vision most aligned to neural activity; alignment lost in the text bottleneck is recovered when hypotheses are rendered back into images, showing that linguistic compression is lossy yet semantically faithful. Together, these results show that combining generative models with neural digital twins enables interpretable, testable descriptions of neural function at scale, toward agentic scientific discovery.

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.

Natural language is an appealing medium for explaining how large language models process and store information, but evaluating the faithfulness of such explanations is challenging. To help address this, we develop two modes of evaluation for natural language explanations that claim individual neurons represent a concept in a text input. In the observational mode, we evaluate claims that a neuron a activates on all and only input strings that refer to a concept picked out by the proposed explanation E. In the intervention mode, we construe E as a claim that the neuron a is a causal mediator of the concept denoted by E. We apply our framework to the GPT-4-generated explanations of GPT-2 XL neurons of Bills et al. (2023) and show that even the most confident explanations have high error rates and little to no causal efficacy. We close the paper by critically assessing whether natural language is a good choice for explanations and whether neurons are the best level of analysis.

Large language models can detect their own errors and sometimes correct them without external feedback, but the underlying mechanisms remain unknown. We investigate this through the lens of second-order models of confidence from decision neuroscience. In a first-order system, confidence derives from the generation signal itself and is therefore maximal for the chosen response, precluding error detection. Second-order models posit a partially independent evaluative signal that can disagree with the committed response, providing the basis for error detection. Kumaran et al. (2026) showed that LLMs cache a confidence representation at a token immediately following the answer (i.e. post-answer newline: PANL) -- that causally drives verbal confidence and dissociates from log-probabilities. Here we test whether this PANL signal extends beyond confidence to support error detection and self-correction. Here we test whether this signal supports error detection and self-correction, deriving predictions from the second-order framework. Using a verify-then-correct paradigm, we show that: (i) verbal confidence predicts error detection far beyond token log-probabilities, ruling out a first-order account; (ii) PANL activations predict error detection beyond verbal confidence itself; and (iii) PANL predicts which errors the model can correct -- where all behavioural signals fail. Causal interventions confirm that PANL signals rescue error detection behavior when answer information is corrupted. All findings replicate across models (Gemma 3 27B and Qwen 2.5 7B) and tasks (TriviaQA and MNLI). These results reveal that LLMs naturally implement a second-order confidence architecture whose internal evaluative signal encodes not only whether an answer is likely wrong but whether the model has the knowledge to fix it.

What mechanisms underlie linguistic generalization in large language models (LLMs)? This question has attracted considerable attention, with most studies analyzing the extent to which the language skills of LLMs resemble rules. As of yet, it is not known whether linguistic generalization in LLMs could equally well be explained as the result of analogical processes, which can be formalized as similarity operations on stored exemplars. A key shortcoming of prior research is its focus on linguistic phenomena with a high degree of regularity, for which rule-based and analogical approaches make the same predictions. Here, we instead examine derivational morphology, specifically English adjective nominalization, which displays notable variability. We introduce a new method for investigating linguistic generalization in LLMs: focusing on GPT-J, we fit cognitive models that instantiate rule-based and analogical learning to the LLM training data and compare their predictions on a set of nonce adjectives with those of the LLM, allowing us to draw direct conclusions regarding underlying mechanisms. As expected, rule-based and analogical models explain the predictions of GPT-J equally well for adjectives with regular nominalization patterns. However, for adjectives with variable nominalization patterns, the analogical model provides a much better match. Furthermore, GPT-J's behavior is sensitive to the individual word frequencies, even for regular forms, a behavior that is consistent with an analogical account of regular forms but not a rule-based one. These findings refute the hypothesis that GPT-J's linguistic generalization on adjective nominalization involves rules, suggesting similarity operations on stored exemplars as the underlying mechanism. Overall, our study suggests that analogical processes play a bigger role in the linguistic generalization of LLMs than previously thought.

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