Daily Papers

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.

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.

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.

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.

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.

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.

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

LLMs are increasingly deployed as orchestrators that coordinate specialized subagents to solve complex tasks through natural language. However, in many important domains like game playing and robotics, the strongest available agents are not language models. Integrating non-language agents with LLMs would require verbalization: compressing their rich continuous representations into sparse textual summaries at each interaction step. To study whether verbalization constitutes a bottleneck, we introduce LLAMIA-Bench, a suite of six diverse collaborative chess tasks spanning three facets: behavioral imitation, state assessment, and natural-language explanation. Each task instantiates a well-established chess problem that neither the LLM nor the chess engine can solve alone. To solve LLM collaboration with non-language agents, we introduce latent state internalization, which projects the subagent's continuous representations directly into the LLM's token stream as learned state tokens, with dynamic re-encoding as actions advance the environment state. Comparing internalization to verbalized integration, our experiments reveal a consistent verbalization debt: the performance gap widens throughout training and persists as the LLM scales from 4B to 14B parameters. A single 14B model, LLAMIA, trained with latent state internalization, matches or exceeds task specialists and frontier models including GPT-5.1 with tool access across all benchmark tasks, and generalizes out-of-distribution where task-specific finetunes collapse

The field of large language models (LLMs) has grown rapidly in recent years, driven by the desire for better efficiency, interpretability, and safe use. Building on the novel approach of "activation engineering," this study explores personality modification in LLMs, drawing inspiration from research like Refusal in LLMs Is Mediated by a Single Direction (arXiv:2406.11717) and Steering Llama 2 via Contrastive Activation Addition (arXiv:2312.06681). We leverage activation engineering to develop a method for identifying and adjusting activation directions related to personality traits, which may allow for dynamic LLM personality fine-tuning. This work aims to further our understanding of LLM interpretability while examining the ethical implications of such developments.

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.

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.

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.

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

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".

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

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.

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

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 .

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.

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

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

Recent work on domain-specific reasoning with large language models (LLMs) often relies on training-intensive approaches that require parameter updates. While activation steering has emerged as a parameter efficient alternative, existing methods apply static, manual interventions that fail to adapt to the dynamic nature of complex reasoning. To address this limitation, we propose RISER (Router-based Intervention for Steerable Enhancement of Reasoning), a plug-and-play intervention framework that adaptively steers LLM reasoning in activation space. RISER constructs a library of reusable reasoning vectors and employs a lightweight Router to dynamically compose them for each input. The Router is optimized via reinforcement learning under task-level rewards, activating latent cognitive primitives in an emergent and compositional manner. Across seven diverse benchmarks, RISER yields 3.4-6.5% average zero-shot accuracy improvements over the base model while surpassing CoT-style reasoning with 2-3x higher token efficiency and robust accuracy gains. Further analysis shows that RISER autonomously combines multiple vectors into interpretable, precise control strategies, pointing toward more controllable and efficient LLM reasoning.

Mental manipulation severely undermines mental wellness by covertly and negatively distorting decision-making. While there is an increasing interest in mental health care within the natural language processing community, progress in tackling manipulation remains limited due to the complexity of detecting subtle, covert tactics in conversations. In this paper, we propose Intent-Aware Prompting (IAP), a novel approach for detecting mental manipulations using large language models (LLMs), providing a deeper understanding of manipulative tactics by capturing the underlying intents of participants. Experimental results on the MentalManip dataset demonstrate superior effectiveness of IAP against other advanced prompting strategies. Notably, our approach substantially reduces false negatives, helping detect more instances of mental manipulation with minimal misjudgment of positive cases. The code of this paper is available at https://github.com/Anton-Jiayuan-MA/Manip-IAP.

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).

Spoken communication occurs in a "noisy channel" characterized by high levels of environmental noise, variability within and between speakers, and lexical and syntactic ambiguity. Given these properties of the received linguistic input, robust spoken word recognition -- and language processing more generally -- relies heavily on listeners' prior knowledge to evaluate whether candidate interpretations of that input are more or less likely. Here we compare several broad-coverage probabilistic generative language models in their ability to capture human linguistic expectations. Serial reproduction, an experimental paradigm where spoken utterances are reproduced by successive participants similar to the children's game of "Telephone," is used to elicit a sample that reflects the linguistic expectations of English-speaking adults. When we evaluate a suite of probabilistic generative language models against the yielded chains of utterances, we find that those models that make use of abstract representations of preceding linguistic context (i.e., phrase structure) best predict the changes made by people in the course of serial reproduction. A logistic regression model predicting which words in an utterance are most likely to be lost or changed in the course of spoken transmission corroborates this result. We interpret these findings in light of research highlighting the interaction of memory-based constraints and representations in language processing.

On-policy self-distillation (OPSD) trains a large language model using supervision from a teacher instantiated from the same model as the student. The teacher is conditioned on privileged information (PI), such as verified solutions. The student, which does not observe PI, learns from the teacher's next-token distributions along its own sampled responses. However, PI can bias supervision toward shortcut reasoning with limited verification or reconsideration, accompanied by a sharp drop in the entropy of the output distribution. To address this challenge, we introduce self-distillation with verbal regularization (VRSD), which uses system prompts to regulate how the teacher uses PI, thereby encouraging rethinking and exploration to counteract premature commitment. Theoretically, we model prompting as attenuating the influence of PI on the teacher distribution and derive conditions for entropy preservation. Across three models and four tasks, VRSD improves Pass@1 over OPSD in 11 of 12 model-task settings, by 3.9 percentage points on average, while generally maintaining higher token entropy.

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.

Subliminal learning is a phenomenon in which a distillation dataset transmits traits from the teacher model that are not legibly encoded in the dataset itself. This introduces a new challenge for model development and creates new risks from data poisoning. In this work, we use text optimization to detect subliminal learning effects and describe them as legible prompts. Subliminal learning from a prompted teacher motivates our approach. We observe that this is a special case of context distillation and leverage this observation to show that, in theory, the prompted subliminal learning dataset identifies the teacher's prompt. We reduce recovering this prompt to a text optimization problem and present a method to approximately solve it. Our method, SALVE (Search-Aided Latent Verbalization), optimizes a soft prompt, queries the same model to verbalize it as text, and uses beam search to make the verbalization reliable. In the standard subliminal learning setting, SALVE reliably recovers legible prompts that name the teacher's trait, while common text optimization methods fail to do so. In addition, we find that there are settings in which SALVE recovers the teacher's trait from a dataset even when subliminal learning fails, but that modifying student training to improve context distillation can create subliminal learning effects. We lastly show that SALVE detects subliminal learning effects in three additional settings: (1) mixtures of subliminal learning data and unrelated data, (2) data generated when the teacher is biased via activation steering, and (3) subsets of real preference data selected via Logit-Linear Selection. Overall, our results deepen our understanding of subliminal learning and present SALVE as a method to proactively detect subliminal learning effects.

Large language models can be steered at inference time through prompting or activation interventions, but activation steering methods often underperform compared to prompt-based approaches. We propose a framework that formulates prompt steering as a form of activation steering and investigates whether distilling successful prompt steering behavior into simpler, interpretable models can close this gap. Our analysis reveals that popular activation steering methods are not faithful to the mechanics of prompt steering, which applies strong interventions on some tokens while barely affecting others. Based on these insights, we introduce Prompt Steering Replacement (PSR) models that estimate token-specific steering coefficients from the activations themselves and are trained to imitate prompt-based interventions. Experiments on three steering benchmarks across multiple language models show that PSR models outperform existing activation steering methods, especially when controlling for high-coherence completions, and also compare favorably to prompting on AxBench and persona steering.

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.

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.

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

Hallucination has emerged as the most vulnerable aspect of contemporary Large Language Models (LLMs). In this paper, we introduce the Sorry, Come Again (SCA) prompting, aimed to avoid LLM hallucinations by enhancing comprehension through: (i) optimal paraphrasing and (ii) injecting [PAUSE] tokens to delay LLM generation. First, we provide an in-depth analysis of linguistic nuances: formality, readability, and concreteness of prompts for 21 LLMs, and elucidate how these nuances contribute to hallucinated generation. Prompts with lower readability, formality, or concreteness pose comprehension challenges for LLMs, similar to those faced by humans. In such scenarios, an LLM tends to speculate and generate content based on its imagination (associative memory) to fill these information gaps. Although these speculations may occasionally align with factual information, their accuracy is not assured, often resulting in hallucination. Recent studies reveal that an LLM often neglects the middle sections of extended prompts, a phenomenon termed as lost in the middle. While a specific paraphrase may suit one LLM, the same paraphrased version may elicit a different response from another LLM. Therefore, we propose an optimal paraphrasing technique to identify the most comprehensible paraphrase of a given prompt, evaluated using Integrated Gradient (and its variations) to guarantee that the LLM accurately processes all words. While reading lengthy sentences, humans often pause at various points to better comprehend the meaning read thus far. We have fine-tuned an LLM with injected [PAUSE] tokens, allowing the LLM to pause while reading lengthier prompts. This has brought several key contributions: (i) determining the optimal position to inject [PAUSE], (ii) determining the number of [PAUSE] tokens to be inserted, and (iii) introducing reverse proxy tuning to fine-tune the LLM for [PAUSE] insertion.

To correctly use in-context information, language models (LMs) must bind entities to their attributes. For example, given a context describing a "green square" and a "blue circle", LMs must bind the shapes to their respective colors. We analyze LM representations and identify the binding ID mechanism: a general mechanism for solving the binding problem, which we observe in every sufficiently large model from the Pythia and LLaMA families. Using causal interventions, we show that LMs' internal activations represent binding information by attaching binding ID vectors to corresponding entities and attributes. We further show that binding ID vectors form a continuous subspace, in which distances between binding ID vectors reflect their discernability. Overall, our results uncover interpretable strategies in LMs for representing symbolic knowledge in-context, providing a step towards understanding general in-context reasoning in large-scale LMs.

Language model detoxification aims to minimize the risk of generating offensive or harmful content in pretrained language models (PLMs) for safer deployment. Existing methods can be roughly categorized as finetuning-based and decoding-based. However, the former is often resource-intensive, while the latter relies on additional components and potentially compromises the generation fluency. In this paper, we propose a more lightweight approach that enables the PLM itself to achieve "self-detoxification". Our method is built upon the observation that prepending a negative steering prompt can effectively induce PLMs to generate toxic content. At the same time, we are inspired by the recent research in the interpretability field, which formulates the evolving contextualized representations within the PLM as an information stream facilitated by the attention layers. Drawing on this idea, we devise a method to identify the toxification direction from the normal generation process to the one prompted with the negative prefix, and then steer the generation to the reversed direction by manipulating the information movement within the attention layers. Experimental results show that our approach, without any fine-tuning or extra components, can achieve comparable performance with state-of-the-art methods.

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

Large language models (LLMs) tend to verbalize confidence scores that are largely detached from their actual accuracy, yet the geometric relationship governing this behavior remain poorly understood. In this work, we present a mechanistic interpretability analysis of verbalized confidence, using linear probes and contrastive activation addition (CAA) steering to show that calibration and verbalized confidence signals are encoded linearly but are orthogonal to one another -- a finding consistent across three open-weight models and four datasets. Interestingly, when models are prompted to simultaneously reason through a problem and verbalize a confidence score, the reasoning process disrupts the verbalized confidence direction, exacerbating miscalibration. We term this the "Reasoning Contamination Effect." Leveraging this insight, we introduce a two-stage adaptive steering pipeline that reads the model's internal accuracy estimate and steers verbalized output to match it, substantially improving calibration alignment across all evaluated models.

While numerous works have assessed the generative performance of language models (LMs) on tasks requiring Theory of Mind reasoning, research into the models' internal representation of mental states remains limited. Recent work has used probing to demonstrate that LMs can represent beliefs of themselves and others. However, these claims are accompanied by limited evaluation, making it difficult to assess how mental state representations are affected by model design and training choices. We report an extensive benchmark with various LM types with different model sizes, fine-tuning approaches, and prompt designs to study the robustness of mental state representations and memorisation issues within the probes. Our results show that the quality of models' internal representations of the beliefs of others increases with model size and, more crucially, with fine-tuning. We are the first to study how prompt variations impact probing performance on theory of mind tasks. We demonstrate that models' representations are sensitive to prompt variations, even when such variations should be beneficial. Finally, we complement previous activation editing experiments on Theory of Mind tasks and show that it is possible to improve models' reasoning performance by steering their activations without the need to train any probe.

In this paper, we study how training data creates associations between the tokens at the start of a base model's response and the reasoning behavior that follows. First, we demonstrate that fixing particular starting token cues makes a base model's performance competitive with that of its reinforcement learning (RL)-trained counterparts on math and coding. For instance, the cue ".\n\nOkay" raises Olmo-3-7B's MATH-500 pass@1 accuracy from 42% to 78%, while "Alright," raises Qwen3-14B's from 72% to 87%. Second, RL makes these cues more likely, while fixing them recovers much of its performance gain over the base model. Third, we trace the reasoning effects of token cues to the training data. We perform causal data interventions to turn an arbitrary word, such as "chicken", into an effective reasoning cue, or remove an existing cue's effect. A similar edit makes the prompt instruction "Think duck duck goose" as effective as "Think step by step" at eliciting reasoning. We also find that the hidden state representations induced by different cues correlate with different document types from the training set. Finally, we extend our study of token cues with a case study in language model safety, finding that different cues elicit distinct refusal and compliance behaviors that correspond to different types of training data.

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

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.

The recent advancements in Large Language Models (LLMs) have garnered widespread acclaim for their remarkable emerging capabilities. However, the issue of hallucination has parallelly emerged as a by-product, posing significant concerns. While some recent endeavors have been made to identify and mitigate different types of hallucination, there has been a limited emphasis on the nuanced categorization of hallucination and associated mitigation methods. To address this gap, we offer a fine-grained discourse on profiling hallucination based on its degree, orientation, and category, along with offering strategies for alleviation. As such, we define two overarching orientations of hallucination: (i) factual mirage (FM) and (ii) silver lining (SL). To provide a more comprehensive understanding, both orientations are further sub-categorized into intrinsic and extrinsic, with three degrees of severity - (i) mild, (ii) moderate, and (iii) alarming. We also meticulously categorize hallucination into six types: (i) acronym ambiguity, (ii) numeric nuisance, (iii) generated golem, (iv) virtual voice, (v) geographic erratum, and (vi) time wrap. Furthermore, we curate HallucInation eLiciTation (HILT), a publicly available dataset comprising of 75,000 samples generated using 15 contemporary LLMs along with human annotations for the aforementioned categories. Finally, to establish a method for quantifying and to offer a comparative spectrum that allows us to evaluate and rank LLMs based on their vulnerability to producing hallucinations, we propose Hallucination Vulnerability Index (HVI). We firmly believe that HVI holds significant value as a tool for the wider NLP community, with the potential to serve as a rubric in AI-related policy-making. In conclusion, we propose two solution strategies for mitigating hallucinations.

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

Modelling persuasion strategies as predictors of task outcome has several real-world applications and has received considerable attention from the computational linguistics community. However, previous research has failed to account for the resisting strategies employed by an individual to foil such persuasion attempts. Grounded in prior literature in cognitive and social psychology, we propose a generalised framework for identifying resisting strategies in persuasive conversations. We instantiate our framework on two distinct datasets comprising persuasion and negotiation conversations. We also leverage a hierarchical sequence-labelling neural architecture to infer the aforementioned resisting strategies automatically. Our experiments reveal the asymmetry of power roles in non-collaborative goal-directed conversations and the benefits accrued from incorporating resisting strategies on the final conversation outcome. We also investigate the role of different resisting strategies on the conversation outcome and glean insights that corroborate with past findings. We also make the code and the dataset of this work publicly available at https://github.com/americast/resper.

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.

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.

Large language models (LLMs) have achieved remarkable performance across many generation tasks. Nevertheless, effectively aligning them with desired behaviors remains a significant challenge. Activation steering is an effective and cost-efficient approach that directly modifies the activations of LLMs during the inference stage, aligning their responses with the desired behaviors and avoiding the high cost of fine-tuning. Existing methods typically indiscriminately intervene to all generations or rely solely on the question to determine intervention, which limits the accurate assessment of the intervention strength. To this end, we propose the Flexible Activation Steering with Backtracking (FASB) framework, which dynamically determines both the necessity and strength of intervention by tracking the internal states of the LLMs during generation, considering both the question and the generated content. Since intervening after detecting a deviation from the desired behavior is often too late, we further propose the backtracking mechanism to correct the deviated tokens and steer the LLMs toward the desired behavior. Extensive experiments on the TruthfulQA dataset and six multiple-choice datasets demonstrate that our method outperforms baselines. Our code will be released at https://github.com/gjw185/FASB.

The prevalent use of large language models (LLMs) in various domains has drawn attention to the issue of "hallucination," which refers to instances where LLMs generate factually inaccurate or ungrounded information. Existing techniques for hallucination detection in language assistants rely on intricate fuzzy, specific free-language-based chain of thought (CoT) techniques or parameter-based methods that suffer from interpretability issues. Additionally, the methods that identify hallucinations post-generation could not prevent their occurrence and suffer from inconsistent performance due to the influence of the instruction format and model style. In this paper, we introduce a novel pre-detection self-evaluation technique, referred to as SELF-FAMILIARITY, which focuses on evaluating the model's familiarity with the concepts present in the input instruction and withholding the generation of response in case of unfamiliar concepts. This approach emulates the human ability to refrain from responding to unfamiliar topics, thus reducing hallucinations. We validate SELF-FAMILIARITY across four different large language models, demonstrating consistently superior performance compared to existing techniques. Our findings propose a significant shift towards preemptive strategies for hallucination mitigation in LLM assistants, promising improvements in reliability, applicability, and interpretability.

Full-duplex spoken language models (FD-SLMs) enable seamless speech interaction by allowing models to listen and speak simultaneously, yet the internal mechanism by which they coordinate listening and speaking remains underexplored. We analyze the predictive behavior encoded in FD-SLM hidden representations and find that they exhibit stream-specific predictive patterns: during listening, they preferentially predict the incoming user stream, whereas during speaking, they preferentially predict the model output stream. Building on this observation, we show that FD-SLMs dynamically modulate their internal predictive focus between two states: a generative state aligned with model output generation and a perceptive state aligned with incoming user input. However, this modulation can lag behind abrupt changes in conversational context. During user interruptions, the model remains transiently biased toward the generative state before transitioning into the perceptive state, causing it to miss the beginning of the incoming input. We term this delayed internal transition state inertia. To quantify its downstream impact, we introduce the Zero-Buffer Benchmark (ZBB), a diagnostic benchmark for evaluating immediate interruption comprehension when user speech begins abruptly. We evaluate this setting using response correctness and initial-word occurrence rate (IWOR). Finally, we mitigate state inertia through activation steering with a perception vector, a training-free intervention with little additional computational overhead. Across multiple state-of-the-art FD-SLMs, activation steering substantially improves interruption handling; for example, on PersonaPlex, it improves correctness from 28% to 45% and IWOR from 40% to 72% without any fine-tuning.

Generating natural language text from graph-structured data is essential for conversational information seeking. Semantic triples derived from knowledge graphs can serve as a valuable source for grounding responses from conversational agents by providing a factual basis for the information they communicate. This is especially relevant in the context of large language models, which offer great potential for conversational interaction but are prone to hallucinating, omitting, or producing conflicting information. In this study, we conduct an empirical analysis of conversational large language models in generating natural language text from semantic triples. We compare four large language models of varying sizes with different prompting techniques. Through a series of benchmark experiments on the WebNLG dataset, we analyze the models' performance and identify the most common issues in the generated predictions. Our findings show that the capabilities of large language models in triple verbalization can be significantly improved through few-shot prompting, post-processing, and efficient fine-tuning techniques, particularly for smaller models that exhibit lower zero-shot performance.

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

While instruction-tuned models have shown remarkable success in various natural language processing tasks, accurately evaluating their ability to follow instructions remains challenging. Existing benchmarks primarily focus on common instructions that align well with what the model learned during training. However, proficiency in responding to these instructions does not necessarily imply strong ability in instruction following. In this paper, we propose a novel instruction-following evaluation protocol called verbalizer manipulation. It instructs the model to verbalize the task label with words aligning with model priors to different extents, adopting verbalizers from highly aligned (e.g., outputting ``postive'' for positive sentiment), to minimally aligned (e.g., outputting ``negative'' for positive sentiment). Verbalizer manipulation can be seamlessly integrated with any classification benchmark to examine the model's reliance on priors and its ability to override them to accurately follow the instructions. We conduct a comprehensive evaluation of four major model families across nine datasets, employing twelve sets of verbalizers for each of them. We observe that the instruction-following abilities of models, across different families and scales, are significantly distinguished by their performance on less natural verbalizers. Even the strongest GPT-4 model struggles to perform better than random guessing on the most challenging verbalizer, emphasizing the need for continued advancements to improve their instruction-following abilities.

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).

As large language models (LLMs) become more integrated into societal systems, the risk of them perpetuating and amplifying harmful biases becomes a critical safety concern. Traditional methods for mitigating bias often rely on data filtering or post-hoc output moderation, which treat the model as an opaque black box. In this work, we introduce a complete, end-to-end system that uses techniques from mechanistic interpretability to both identify and actively mitigate bias directly within a model's internal workings. Our method involves two primary stages. First, we train linear "probes" on the internal activations of a model to detect the latent representations of various biases (e.g., gender, race, age). Our experiments on gpt2-large demonstrate that these probes can identify biased content with near-perfect accuracy, revealing that bias representations become most salient in the model's later layers. Second, we leverage these findings to compute "steering vectors" by contrasting the model's activation patterns for biased and neutral statements. By adding these vectors during inference, we can actively steer the model's generative process away from producing harmful, stereotypical, or biased content in real-time. We demonstrate the efficacy of this activation steering technique, showing that it successfully alters biased completions toward more neutral alternatives. We present our work as a robust and reproducible system that offers a more direct and interpretable approach to building safer and more accountable LLMs.

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

This study investigates the potential of large language models (LLMs) to provide accurate estimates of concreteness, valence and arousal for multi-word expressions. Unlike previous artificial intelligence (AI) methods, LLMs can capture the nuanced meanings of multi-word expressions. We systematically evaluated ChatGPT-4o's ability to predict concreteness, valence and arousal. In Study 1, ChatGPT-4o showed strong correlations with human concreteness ratings (r = .8) for multi-word expressions. In Study 2, these findings were repeated for valence and arousal ratings of individual words, matching or outperforming previous AI models. Study 3 extended the prevalence and arousal analysis to multi-word expressions and showed promising results despite the lack of large-scale human benchmarks. These findings highlight the potential of LLMs for generating valuable psycholinguistic data related to multiword expressions. To help researchers with stimulus selection, we provide datasets with AI norms of concreteness, valence and arousal for 126,397 English single words and 63,680 multi-word expressions

Large language models are prone to hallucinating factually incorrect statements. A key source of these errors is exposure to new factual information through supervised fine-tuning (SFT), which can increase hallucinations w.r.t. knowledge acquired during pre-training. In this work, we explore whether SFT-induced hallucinations can be mitigated using established tools from the continual learning literature, since they arise as a by-product of knowledge degradation during training. We propose a self-distillation-based SFT method that facilitates effective factual learning while minimizing hallucinations w.r.t. pre-existing knowledge by regularizing output-distribution drift. We also show that, in settings where new knowledge acquisition is unnecessary, suppressing factual plasticity by freezing parameter groups, can preserve task performance while reducing hallucinations. Lastly, we investigate the mechanism behind SFT-induced hallucinations through three hypotheses: capacity limitations, behavior cloning, and localized interference. Our experiments show that a main driver is interference among overlapping semantic representations, and that self-distillation succeeds by mitigating this interference.

Sycophancy and hallucination are persistent failure modes of Large Language Models (LLMs) across domains. However, it becomes particularly consequential in clinical question answering, where responses must remain grounded in the provided context and robust to user pressure. Hallucination can introduce information that is unsupported by the context, while sycophancy can cause a model to abandon a previously correct answer when challenged by the user. Existing approaches, such as prompt-based safeguards and always-on activation steering, often address these behaviors separately or apply interventions broadly across turns, which can unnecessarily deteriorate responses that were already correct. To address these limitations within a single framework, we employ Inference Time Intervention (ITI) to jointly control both behaviors by learning separate steering directions for hallucination and sycophancy from contrastive clinical pairs and applying them to causally verified attention heads. During runtime, behavior-specific gates then determine when intervention is needed: the hallucination component mitigates unsupported claims, while the sycophancy component mitigates answer shifts caused by user pressure. We evaluate this framework on clinical questions grounded in EHR data while keeping the model weights frozen. Across all evaluation settings, we conducted 15,900 model-response runs. Across 600 pressure trajectories for the 4-billion-parameter model, the unsteered model caved in 570 cases. At the same time, gated steering helped it last longer in 551 of them. It held its ground under pressure at levels comparable to those of models with more than 100 billion parameters, showing that targeted inference-time steering can improve robustness without intervening at every turn.

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.

Controlling undesirable Large Language Model (LLM) behaviors, such as the generation of unsafe content or failing to adhere to safety guidelines, often relies on costly fine-tuning. Activation steering provides an alternative for inference-time control, but existing methods typically lack fine-grained, adaptive mechanisms. We introduce a novel approach using a lightweight, trainable controller network integrated during inference. This controller network observes specific intermediate LLM activations and predicts both a global scaling factor and layer-specific weights. The predicted global scaling factor and layer-specific weights then dynamically modulate the intensity of a steering patch, derived from a pre-computed "refusal direction" vector, applied across the LLM's layers during generation. Trained on activations from both harmful and benign prompts, our controller learns to discriminatively apply nuanced, layer-aware interventions, activating steering primarily for harmful inputs. Experiments using safety benchmarks like ToxicChat & In-The-Wild Jailbreak Prompts demonstrate that our weighted steering controller significantly increases refusal rates compared to the base LLM, achieving targeted behavioral modification without altering the original model parameters. Our experiments with Llama-3.1-8B, Llama-3.2-1B & Mistral-7B show our approach outperforms existing methods, presenting an efficient and adaptive method for fine-grained control over LLM behavior at inference time.

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.

Humans can attribute beliefs to others. However, it is unknown to what extent this ability results from an innate biological endowment or from experience accrued through child development, particularly exposure to language describing others' mental states. We test the viability of the language exposure hypothesis by assessing whether models exposed to large quantities of human language display sensitivity to the implied knowledge states of characters in written passages. In pre-registered analyses, we present a linguistic version of the False Belief Task to both human participants and a Large Language Model, GPT-3. Both are sensitive to others' beliefs, but while the language model significantly exceeds chance behavior, it does not perform as well as the humans, nor does it explain the full extent of their behavior -- despite being exposed to more language than a human would in a lifetime. This suggests that while statistical learning from language exposure may in part explain how humans develop the ability to reason about the mental states of others, other mechanisms are also responsible.

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

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

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.

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.

Recent research shows that Large Language Models (LLMs) exhibit a compelling level of proficiency in Theory of Mind (ToM) tasks. This ability to impute unobservable mental states to others is vital to human social cognition and may prove equally important in principal-agent relations between individual humans and Artificial Intelligences (AIs). In this paper, we explore how a mechanism studied in developmental psychology known as Violation of Expectation (VoE) can be implemented to reduce errors in LLM prediction about users by leveraging emergent ToM affordances. And we introduce a metacognitive prompting framework to apply VoE in the context of an AI tutor. By storing and retrieving facts derived in cases where LLM expectation about the user was violated, we find that LLMs are able to learn about users in ways that echo theories of human learning. Finally, we discuss latent hazards and augmentative opportunities associated with modeling user psychology and propose ways to mitigate risk along with possible directions for future inquiry.

Multimodal large language models can exhibit text dominance, over-relying on linguistic priors instead of grounding predictions in non-text inputs. One example is large audio-language models (LALMs) where decisive audio evidence can be under-utilized even when it contains important information. To address this issue we use mechanistic interpretability to identify a small set of audio-specialist attention heads whose audio attention yields a ``listening'' signal. We show that this signal increases when audio evidence affects the model's output, providing an indicator of audio engagement under standard prompting. Leveraging this localization, we construct an audio--silence steering direction and apply an inference-time activation intervention to the final representation, amplifying the model's audio effect. To demonstrate the utility of this intervention, we show on MMAU that this improves accuracy by up to +8.0 percentage points on two Qwen-based LALMs, without any parameter updates.

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

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.

According to the Stimulus Organism Response (SOR) theory, all human behavioral reactions are stimulated by context, where people will process the received stimulus and produce an appropriate reaction. This implies that in a specific context for a given input stimulus, a person can react differently according to their internal state and other contextual factors. Analogously, in dyadic interactions, humans communicate using verbal and nonverbal cues, where a broad spectrum of listeners' non-verbal reactions might be appropriate for responding to a specific speaker behaviour. There already exists a body of work that investigated the problem of automatically generating an appropriate reaction for a given input. However, none attempted to automatically generate multiple appropriate reactions in the context of dyadic interactions and evaluate the appropriateness of those reactions using objective measures. This paper starts by defining the facial Multiple Appropriate Reaction Generation (fMARG) task for the first time in the literature and proposes a new set of objective evaluation metrics to evaluate the appropriateness of the generated reactions. The paper subsequently introduces a framework to predict, generate, and evaluate multiple appropriate facial reactions.

With the continuous development of pre-trained language models, prompt-based training becomes a well-adopted paradigm that drastically improves the exploitation of models for many natural language processing tasks. Prompting also shows great performance compared to traditional fine-tuning when adapted to zero-shot or few-shot scenarios where the number of annotated data is limited. In this framework, the role of verbalizers is essential, as an interpretation from masked word distributions into output predictions. In this work, we propose mave, an approach for verbalizer construction by enrichment of class labels using neighborhood relation in the embedding space of words for the text classification task. In addition, we elaborate a benchmarking procedure to evaluate typical baselines of verbalizers for document classification in few-shot learning contexts. Our model achieves state-of-the-art results while using significantly fewer resources. We show that our approach is particularly effective in cases with extremely limited supervision data.

Mechanistic interpretability seeks to understand the internal mechanisms of machine learning models, where localization -- identifying the important model components -- is a key step. Activation patching, also known as causal tracing or interchange intervention, is a standard technique for this task (Vig et al., 2020), but the literature contains many variants with little consensus on the choice of hyperparameters or methodology. In this work, we systematically examine the impact of methodological details in activation patching, including evaluation metrics and corruption methods. In several settings of localization and circuit discovery in language models, we find that varying these hyperparameters could lead to disparate interpretability results. Backed by empirical observations, we give conceptual arguments for why certain metrics or methods may be preferred. Finally, we provide recommendations for the best practices of activation patching going forwards.

LLMs often exhibit Aha moments during reasoning, such as apparent self-correction following tokens like "Wait," yet their underlying mechanisms remain unclear. We introduce an information-theoretic framework that decomposes reasoning into procedural information and epistemic verbalization - the explicit externalization of uncertainty that supports downstream control actions. We show that purely procedural reasoning can become informationally stagnant, whereas epistemic verbalization enables continued information acquisition and is critical for achieving information sufficiency. Empirical results demonstrate that strong reasoning performance is driven by uncertainty externalization rather than specific surface tokens. Our framework unifies prior findings on Aha moments and post-training experiments, and offers insights for future reasoning model design.

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 modeling of turn-taking in dialog can be viewed as the modeling of the dynamics of voice activity of the interlocutors. We extend prior work and define the predictive task of Voice Activity Projection, a general, self-supervised objective, as a way to train turn-taking models without the need of labeled data. We highlight a theoretical weakness with prior approaches, arguing for the need of modeling the dependency of voice activity events in the projection window. We propose four zero-shot tasks, related to the prediction of upcoming turn-shifts and backchannels, and show that the proposed model outperforms prior work.

Verbal confidence elicitation is widely used to extract uncertainty estimates from LLMs. We tested whether seven instruction-tuned open-weight models (3-9B parameters, four families) produce verbalised confidence that meets minimal validity criteria for item-level Type-2 discrimination under minimal numeric elicitation with greedy decoding. In a pre-registered study (OSF: osf.io/azbvx), 524 TriviaQA items were administered under numeric (0-100) and categorical (10-class) elicitation to eight models at Q5_K_M quantisation on consumer hardware, yielding 8,384 deterministic trials. A psychometric validity screen was applied to each model-format cell. All seven instruct models were classified Invalid on numeric confidence (H2 confirmed, 7/7 vs. predicted >=4/7), with a mean ceiling rate of 91.7% (H1 confirmed). Categorical elicitation did not rescue validity. Instead, it disrupted task performance in six of seven models, producing accuracy below 5% (H4 not confirmed). Token-level logprobability did not usefully predict verbalised confidence under the observed variance regime (H5 confirmed, mean cross-validated R^2 < 0.01). Within the reasoning-distilled model, reasoning-trace length showed a strong negative partial correlation with confidence (rho = -0.36, p < .001), consistent with the Reasoning Contamination Effect. These results do not imply that internal uncertainty representations are absent. They show that minimal verbal elicitation fails to preserve internal signals at the output interface in this model-size regime. Psychometric screening should precede any downstream use of such signals.

Multimodal large language models (MLLMs) are broadly empowering various fields. Despite their advancements, the explainability of MLLMs remains less explored, hindering deeper understanding, model credibility, and effective visualization. Unlike conventional vision models (e.g., CNNs, ViTs, CLIP) that produce a single output, MLLMs generate sequences of tokens progressively, where each generated token depends on the previous context. Therefore, earlier context tokens can introduce redundant activations that interfere with the explanation of later tokens beyond their original information. Existing studies often overlook this issue, but our observations reveal that these redundant correlations can significantly hurt the reliability of explanations. To address this, we propose an estimated causal inference method to mitigate the interference of context to achieve high-quality MLLM explanation, with a novel rank Gaussian filter to further reduce activation noises. We term this method Token Activation Map (TAM) to highlight the consideration of interactions between tokens. TAM also indicates that it excels at explaining multiple tokens of MLLM, which is different from the Class Activation Map (CAM) for a single prediction. Our TAM method significantly outperforms existing SoTA methods, showcasing high-quality visualization results that can be utilized for various scenarios, such as object localization, failure case analysis, video visualization, MLLMs visual comparison, and model understanding (e.g., color, shape, action, location, visual reasoning, multi-turn conversation, etc). The code is available atgithub.com/xmed-lab/TAM.

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

Large Language Models (LLMs) are claimed to be capable of Natural Language Inference (NLI), necessary for applied tasks like question answering and summarization. We present a series of behavioral studies on several LLM families (LLaMA, GPT-3.5, and PaLM) which probe their behavior using controlled experiments. We establish two biases originating from pretraining which predict much of their behavior, and show that these are major sources of hallucination in generative LLMs. First, memorization at the level of sentences: we show that, regardless of the premise, models falsely label NLI test samples as entailing when the hypothesis is attested in training data, and that entities are used as ``indices'' to access the memorized data. Second, statistical patterns of usage learned at the level of corpora: we further show a similar effect when the premise predicate is less frequent than that of the hypothesis in the training data, a bias following from previous studies. We demonstrate that LLMs perform significantly worse on NLI test samples which do not conform to these biases than those which do, and we offer these as valuable controls for future LLM evaluation.

Theory of Mind (ToM) is a critical component of intelligence, yet accurately measuring it continues to be a subject of debate. Prior research has attempted to apply human ToM assessments to natural language processing models using either human-created standardized tests or rule-based templates. However, these methods primarily focus on simplistic reasoning and require further validation. In this study, we utilize dynamic epistemic logic, which has established overlaps with ToM, to generate more intricate problems. We also introduce novel verbalization techniques to express these problems using natural language. Our findings indicate that certain language model scaling (from 70M to 6B and 350M to 174B) does not consistently yield results better than random chance. While GPT-4 demonstrates improved epistemic reasoning capabilities, there is still room for enhancement. Our code and datasets are publicly available https://github.com/antoinelrnld/modlog https://huggingface.co/datasets/sileod/mindgames

We present our submission to the EgoProactive track of the ECCV 2026 Wearable AI Challenge, which ranked first in the large-model division and second in the <=2B division. The task requires a wearable assistant to decide after each eight-second segment of egocentric video whether to intervene or remain silent. Our approach has two main components. First, we reformulate intervention timing as single-token classification. Rather than generating either interrupt<utterance> or silent, the model predicts yes or no, and we derive the decision from the renormalised probabilities of these two tokens. This formulation improved macro-F1 by 0.249 and G-mean by 0.30 over free-form generation. Second, because labelled data were limited to the released validation set, we generated additional supervision using a tool-calling video agent that inspects each clip and assigns intervention timestamps. A narration-only alternative was four times larger and ten times cheaper, but transferred worse than supervision from an unrelated real corpus, suggesting that visual grounding is more important than annotation volume for this task.

While Large Language Models (LLMs) demonstrate impressive reasoning capabilities, understanding and validating their knowledge utilization remains challenging. Chain-of-thought (CoT) prompting partially addresses this by revealing intermediate reasoning steps, but the knowledge flow and application remain implicit. We introduce IAO (Input-Action-Output) prompting, a structured template-based method that explicitly models how LLMs access and apply their knowledge during complex reasoning tasks. IAO decomposes problems into sequential steps, each clearly identifying the input knowledge being used, the action being performed, and the resulting output. This structured decomposition enables us to trace knowledge flow, verify factual consistency, and identify potential knowledge gaps or misapplications. Through experiments across diverse reasoning tasks, we demonstrate that IAO not only improves zero-shot performance but also provides transparency in how LLMs leverage their stored knowledge. Human evaluation confirms that this structured approach enhances our ability to verify knowledge utilization and detect potential hallucinations or reasoning errors. Our findings provide insights into both knowledge representation within LLMs and methods for more reliable knowledge application.

The transition of Large Language Models (LLMs) from exploratory tools to active "silicon subjects" in social science lacks extensive validation of operational validity. This study introduces Conditioned Comment Prediction (CCP), a task in which a model predicts how a user would comment on a given stimulus by comparing generated outputs with authentic digital traces. This framework enables a rigorous evaluation of current LLM capabilities with respect to the simulation of social media user behavior. We evaluated open-weight 8B models (Llama3.1, Qwen3, Ministral) in English, German, and Luxembourgish language scenarios. By systematically comparing prompting strategies (explicit vs. implicit) and the impact of Supervised Fine-Tuning (SFT), we identify a critical form vs. content decoupling in low-resource settings: while SFT aligns the surface structure of the text output (length and syntax), it degrades semantic grounding. Furthermore, we demonstrate that explicit conditioning (generated biographies) becomes redundant under fine-tuning, as models successfully perform latent inference directly from behavioral histories. Our findings challenge current "naive prompting" paradigms and offer operational guidelines prioritizing authentic behavioral traces over descriptive personas for high-fidelity simulation.

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.

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}.

Deployed large reasoning models (LRMs) often behave unexpectedly. Test-time steering controls LRM outputs by intervening on their hidden representations, but it can degrade output quality. We argue that prior steering work implicitly relies on internal features that detect behavior in already generated text. We show that these detection features are poor predictors of future behavioral outcomes, and thus not the natural intervention target. Instead, we train activation probes to predict future behavior likelihoods from intermediate reasoning steps. These probes predict the most likely behavior with 64%-91% accuracy, revealing a separate type of internal prediction features. Building on these prediction features, we introduce a text-level steering method, Future Probe Controlled Generation. FPCG samples multiple candidate sentences and chooses the best one according to a probe predicting the future behavior likelihood. This enables steering with almost no output quality degradation. FPCG also enables steering in several evaluations where activation steering fails. These results show that distinguishing detection and prediction features enables a more nuanced approach to controlling LRM behaviors.

Every act of language generation compresses a rich internal state into a single token sequence. We call this process intention collapse: a many-to-one projection from a high dimensional intention space I into an external language space L. We formalize intention collapse for contemporary language models, define three simple, model agnostic intention metrics (intention entropy Hint, effective dimensionality dimeff, and latent knowledge recoverability Recov), and propose an empirical agenda for studying how inference time computation shapes internal intentions before they are verbalized. We also report a first small scale experiment. Using a 4 bit Mistral 7B model on 200 GSM8K problems, we compare a direct answer baseline, a chain of thought (CoT) regime, and a babble control. CoT raises accuracy from 5.5 percent to 53 percent, sharply reduces pre collapse intention entropy (from 1.42 to 0.37 bits), and shows higher global effective dimensionality than the other regimes despite producing fewer tokens than babble. At the same time, Hint has little item level predictive power, and a linear probe on I achieves AUROC 0.65 in the CoT regime but only about chance in the baseline regime, where it collapses to the majority class. These preliminary results indicate that intention level metrics can distinguish inference regimes and expose latent information that is partly lost during collapse, while also revealing important limitations of our current proxies

Saliency maps can explain a neural model's predictions by identifying important input features. They are difficult to interpret for laypeople, especially for instances with many features. In order to make them more accessible, we formalize the underexplored task of translating saliency maps into natural language and compare methods that address two key challenges of this approach -- what and how to verbalize. In both automatic and human evaluation setups, using token-level attributions from text classification tasks, we compare two novel methods (search-based and instruction-based verbalizations) against conventional feature importance representations (heatmap visualizations and extractive rationales), measuring simulatability, faithfulness, helpfulness and ease of understanding. Instructing GPT-3.5 to generate saliency map verbalizations yields plausible explanations which include associations, abstractive summarization and commonsense reasoning, achieving by far the highest human ratings, but they are not faithfully capturing numeric information and are inconsistent in their interpretation of the task. In comparison, our search-based, model-free verbalization approach efficiently completes templated verbalizations, is faithful by design, but falls short in helpfulness and simulatability. Our results suggest that saliency map verbalization makes feature attribution explanations more comprehensible and less cognitively challenging to humans than conventional representations.

Understanding how language supports emotion inference remains a topic of debate in emotion science. The present study investigated whether language-derived emotion-concept knowledge would causally support emotion inference by manipulating the language-specific knowledge representations in large language models. Using the prompt technique, 14 attributes of emotion concepts were found to be represented by distinct artificial neuron populations. By manipulating these attribute-related neurons, the majority of the emotion inference tasks showed performance deterioration compared to random manipulations. The attribute-specific performance deterioration was related to the importance of different attributes in human mental space. Our findings provide causal evidence in support of a language-based mechanism for emotion inference and highlight the contributions of emotion-concept knowledge.

When communicating with elders with cognitive impairment, cognitive stimulation (CS) help to maintain the cognitive health of elders. Data sparsity is the main challenge in building CS-based dialogue systems, particularly in the Chinese language. To fill this gap, we construct a Chinese CS conversation (CSConv) dataset, which contains about 2.6K groups of dialogues with CS principles and emotional support strategy labels. Making chit chat while providing emotional support is overlooked by the majority of existing cognitive dialogue systems. In this paper, we propose a multi-source knowledge fusion method for CS dialogue (CSD), to generate open-ended responses guided by the CS principle and emotional support strategy. We first use a progressive mask method based on external knowledge to learn encoders as effective classifiers, which is the prerequisite to predict the CS principle and emotional support strategy of the target response. Then a decoder interacts with the perceived CS principle and emotional support strategy to generate responses. Extensive experiments conducted on the CSConv dataset demonstrate the effectiveness of the proposed method, while there is still a large space for improvement compared to human performance.

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

We present an automated, contrastive evaluation pipeline for auditing the behavioral impact of interventions on large language models. Given a base model M_1 and an intervention model M_2, our method compares their free-form, multi-token generations across aligned prompt contexts and produces human-readable, statistically validated natural-language hypotheses describing how the models differ, along with recurring themes that summarize patterns across validated hypotheses. We evaluate the approach in synthetic setting by injecting known behavioral changes and showing that the pipeline reliably recovers them. We then apply it to three real-world interventions, reasoning distillation, knowledge editing and unlearning, demonstrating that the method surfaces both intended and unexpected behavioral shifts, distinguishes large from subtle interventions, and does not hallucinate differences when effects are absent or misaligned with the prompt bank. Overall, the pipeline provides a statistically grounded and interpretable tool for post-hoc auditing of intervention-induced changes in model behavior.

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.

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.

Humans invent new words when there is a rising demand for a new useful concept (e.g., doomscrolling). We explore and validate a similar idea in our communication with LLMs: introducing new words to better understand and control the models, expanding on the recently introduced neologism learning. This method introduces a new word by adding a new word embedding and training with examples that exhibit the concept with no other changes in model parameters. We show that adding a new word allows for control of concepts such as flattery, incorrect answers, text length, as well as more complex concepts in AxBench. We discover that neologisms can also further our understanding of the model via self-verbalization: models can describe what each new word means to them in natural language, like explaining that a word that represents a concept of incorrect answers means ``a lack of complete, coherent, or meaningful answers...'' To validate self-verbalizations, we introduce plug-in evaluation: we insert the verbalization into the context of a model and measure whether it controls the target concept. In some self-verbalizations, we find machine-only synonyms: words that seem unrelated to humans but cause similar behavior in machines. Finally, we show how neologism learning can jointly learn multiple concepts in multiple words.

The self-rationalising capabilities of large language models (LLMs) have been explored in restricted settings, using task/specific data sets. However, current LLMs do not (only) rely on specifically annotated data; nonetheless, they frequently explain their outputs. The properties of the generated explanations are influenced by the pre-training corpus and by the target data used for instruction fine-tuning. As the pre-training corpus includes a large amount of human-written explanations "in the wild", we hypothesise that LLMs adopt common properties of human explanations. By analysing the outputs for a multi-domain instruction fine-tuning data set, we find that generated explanations show selectivity and contain illustrative elements, but less frequently are subjective or misleading. We discuss reasons and consequences of the properties' presence or absence. In particular, we outline positive and negative implications depending on the goals and user groups of the self-rationalising system.

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.

In this paper, we show that when spoken language models (SLMs) are instructed to speak in a specific speaking style at the beginning of a multi-turn conversation, they cannot maintain the required speaking styles after several turns of interaction; we refer to this as the style amnesia of SLMs. We focus on paralinguistic speaking styles, including emotion, accent, volume, and speaking speed. We evaluate three proprietary and two open-source SLMs, demonstrating that none of these models can maintain a consistent speaking style when instructed to do so. We further show that when SLMs are asked to recall the style instruction in later turns, they can recall the style instruction, but they fail to express it throughout the conversation. We also show that explicitly asking the model to recall the style instruction can partially mitigate style amnesia. In addition, we examine various prompting strategies and find that SLMs struggle to follow the required style when the instruction is placed in system messages rather than user messages, which contradicts the intended function of system prompts.

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.

Large Language Models (LLMs) exhibit impressive capabilities but often hallucinate, confidently providing incorrect answers instead of admitting ignorance. Prior work has shown that models encode linear representations of their own knowledge and that activation steering can reduce hallucinations. These approaches, however, require real-time monitoring and intervention during inference. We introduce Contrastive Activation Steering for Amortized Learning (CASAL), an efficient algorithm that connects interpretability with amortized optimization. CASAL directly bakes the benefits of activation steering into model's weights. Once trained, LLMs answer questions they know while abstaining from answering those they do not. CASAL's light-weight design requires training only a submodule of a single transformer layer and yet reduces hallucination by 30%-40% across multiple short-form QA benchmarks. CASAL is 30x more compute-efficient and 20x more data-efficient than strong LoRA-based baselines such as SFT and DPO, boosting its practical applicability in data scarce domains. Importantly, CASAL also generalizes effectively to out-of-distribution (OOD) domains. We showcase CASAL's flexibility in mitigating hallucinations in both text-only and vision-language models. To our knowledge, CASAL is the first steering-based training method that has been shown to be effective for both dense and Mixture-of-Experts (MoE) models. CASAL represents a promising step forward for applying interpretability-inspired method for practical deployment in production systems.

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

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

Non-verbal communication often comprises of semantically rich gestures that help convey the meaning of an utterance. Producing such semantic co-speech gestures has been a major challenge for the existing neural systems that can generate rhythmic beat gestures, but struggle to produce semantically meaningful gestures. Therefore, we present RAG-Gesture, a diffusion-based gesture generation approach that leverages Retrieval Augmented Generation (RAG) to produce natural-looking and semantically rich gestures. Our neuro-explicit gesture generation approach is designed to produce semantic gestures grounded in interpretable linguistic knowledge. We achieve this by using explicit domain knowledge to retrieve exemplar motions from a database of co-speech gestures. Once retrieved, we then inject these semantic exemplar gestures into our diffusion-based gesture generation pipeline using DDIM inversion and retrieval guidance at the inference time without any need of training. Further, we propose a control paradigm for guidance, that allows the users to modulate the amount of influence each retrieval insertion has over the generated sequence. Our comparative evaluations demonstrate the validity of our approach against recent gesture generation approaches. The reader is urged to explore the results on our project page.

Language models (LMs) can be directed to perform target tasks by using labeled examples or natural language prompts. But selecting examples or writing prompts for can be challenging--especially in tasks that involve unusual edge cases, demand precise articulation of nebulous preferences, or require an accurate mental model of LM behavior. We propose to use *LMs themselves* to guide the task specification process. In this paper, we introduce **Generative Active Task Elicitation (GATE)**: a learning framework in which models elicit and infer intended behavior through free-form, language-based interaction with users. We study GATE in three domains: email validation, content recommendation, and moral reasoning. In preregistered experiments, we show that LMs prompted to perform GATE (e.g., by generating open-ended questions or synthesizing informative edge cases) elicit responses that are often more informative than user-written prompts or labels. Users report that interactive task elicitation requires less effort than prompting or example labeling and surfaces novel considerations not initially anticipated by users. Our findings suggest that LM-driven elicitation can be a powerful tool for aligning models to complex human preferences and values.

Methods for controlling large language models (LLMs), including local weight fine-tuning, LoRA-based adaptation, and activation-based interventions, are often studied in isolation, obscuring their connections and making comparison difficult. In this work, we present a unified view that frames these interventions as dynamic weight updates induced by a control signal, placing them within a single conceptual framework. Building on this view, we propose a unified preference-utility analysis that separates control effects into preference, defined as the tendency toward a target concept, and utility, defined as coherent and task-valid generation, and measures both on a shared log-odds scale using polarity-paired contrastive examples. Across methods, we observe a consistent trade-off between preference and utility: stronger control increases preference while predictably reducing utility. We further explain this behavior through an activation manifold perspective, in which control shifts representations along target-concept directions to enhance preference, while utility declines primarily when interventions push representations off the model's valid-generation manifold. Finally, we introduce a new steering approach SPLIT guided by this analysis that improves preference while better preserving utility. Code is available at https://github.com/zjunlp/EasyEdit/blob/main/examples/SPLIT.md.

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

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

The learning trajectories of linguistic phenomena in humans provide insight into linguistic representation, beyond what can be gleaned from inspecting the behavior of an adult speaker. To apply a similar approach to analyze neural language models (NLM), it is first necessary to establish that different models are similar enough in the generalizations they make. In this paper, we show that NLMs with different initialization, architecture, and training data acquire linguistic phenomena in a similar order, despite their different end performance. These findings suggest that there is some mutual inductive bias that underlies these models' learning of linguistic phenomena. Taking inspiration from psycholinguistics, we argue that studying this inductive bias is an opportunity to study the linguistic representation implicit in NLMs. Leveraging these findings, we compare the relative performance on different phenomena at varying learning stages with simpler reference models. Results suggest that NLMs exhibit consistent "developmental" stages. Moreover, we find the learning trajectory to be approximately one-dimensional: given an NLM with a certain overall performance, it is possible to predict what linguistic generalizations it has already acquired. Initial analysis of these stages presents phenomena clusters (notably morphological ones), whose performance progresses in unison, suggesting a potential link between the generalizations behind them.

Current large language models have dangerous capabilities, which are likely to become more problematic in the future. Activation steering techniques can be used to reduce risks from these capabilities. In this paper, we investigate the efficacy of activation steering for broad skills and multiple behaviours. First, by comparing the effects of reducing performance on general coding ability and Python-specific ability, we find that steering broader skills is competitive to steering narrower skills. Second, we steer models to become more or less myopic and wealth-seeking, among other behaviours. In our experiments, combining steering vectors for multiple different behaviours into one steering vector is largely unsuccessful. On the other hand, injecting individual steering vectors at different places in a model simultaneously is promising.

One of the most central aspects of language processing is the ability to trade off between stored representations and abstract knowledge: one must retrieve stored representations, but also generate novel ones by applying productive rules. While recent work has examined abstract knowledge in language models, holistic storage has received far less attention. We probe internal representations in both text-based LLMs and an ASR model, testing whether V+up phrasal verbs develop distinct representations as a function of frequency and predictability. All models show evidence of holistic storage driven by frequency and predictability, further supporting usage-based theories of language.

Large language models (LLMs) often produce confident yet incorrect answers, which can lead to risky failures in real-world applications. We study whether post-training can make a model's self-assessment explicit: when the model is uncertain, can it be trained to signal so within its own response? A central design question is where in the response this signal should be exposed -- during reasoning, while the answer is still being formed, or at the end, once the answer has been produced. We study both. For end-of-reasoning self-assessment, we train the model to verbalize a confidence score for its response, with the aim of high confidence on correct answers and low confidence on incorrect ones. For during-reasoning self-assessment, we train the model to emit the marker <uncertain> whenever its current reasoning state appears unreliable. Across factual reasoning tasks, both forms sharply reduce overconfident errors while improving answer quality, and both can be used as triggers for retrieval augmented generation (RAG) to improve the final response. We further analyze their internal mechanisms: end-of-reasoning verbalized confidence sharpens a confidence-related structure already present in the pretrained model, whereas during-reasoning <uncertain> emission teaches the model to mark high-risk reasoning steps, with parameter changes concentrated in the model's late layers.

Large language models should follow hierarchical instructions where system prompts override user inputs, yet recent work shows they often ignore this rule while strongly obeying social cues such as authority or consensus. We extend these behavioral findings with mechanistic interpretations on a large-scale dataset. Linear probing shows conflict-decision signals are encoded early, with system-user and social conflicts forming distinct subspaces. Direct Logit Attribution reveals stronger internal conflict detection in system-user cases but consistent resolution only for social cues. Steering experiments show that, despite using social cues, the vectors surprisingly amplify instruction following in a role-agnostic way. Together, these results explain fragile system obedience and underscore the need for lightweight hierarchy-sensitive alignment methods.

A vision--language--action (VLA) policy can complete complex tasks while ignoring the evidence that should determine its actions. An object held near the wrist camera can displace the instructed target. Language and action show the same pattern: a familiar noun can trigger the operation it was paired with in training even after the verb changes, and a gripper that closed on nothing may lift anyway. We call these dependencies modality shortcuts: regularities in successful demonstrations make visual, lexical, or motor cues sufficient to predict expert actions without the task evidence needed for the underlying decision. More demonstrations of the same kind can raise task success while leaving these shortcuts intact. We propose Perturbot which makes task-relevant evidence easier to use and shortcuts insufficient on their own: it applies task-preserving wrist-view perturbations, enriches instructions with decision-relevant captions, and adds random and failed trajectory segments relabeled with the behavior they contain. It complements scaling by changing what is scaled, and leaves inference unchanged. Moreover, we propose GroundingFscore, an offline score that diagnoses how severely a policy relies on modality shortcuts. Task success rate shows whether a policy improves, while GroundingFscore reveals whether the policy scales healthily, relying on task evidence rather than shortcuts. Together, Perturbot and GroundingFscore provide a training-and-evaluation framework for disentangling VLA decisions from shortcut priors while preserving responsiveness to task-relevant evidence.

With the development in cognitive science and Large Language Models (LLMs), increasing connections have come to light between these two distinct fields. Building upon these connections, we propose a conjecture suggesting the existence of a duality between LLMs and Tulving's theory of memory. We identify a potential correspondence between Tulving's synergistic ecphory model (SEM) of retrieval and the emergent abilities observed in LLMs, serving as supporting evidence for our conjecture. Furthermore, we speculate that consciousness may be considered a form of emergent ability based on this duality. We also discuss how other theories of consciousness intersect with our research.

A critical component of a successful language generation pipeline is the decoding algorithm. However, the general principles that should guide the choice of decoding algorithm remain unclear. Previous works only compare decoding algorithms in narrow scenarios and their findings do not generalize across tasks. To better structure the discussion, we introduce a taxonomy that groups decoding strategies based on their implicit assumptions about how well the model's likelihood is aligned with the task-specific notion of utility. We argue that this taxonomy allows a broader view of the decoding problem and can lead to generalizable statements because it is grounded on the interplay between the decoding algorithms and the likelihood-utility misalignment. Specifically, by analyzing the correlation between the likelihood and the utility of predictions across a diverse set of tasks, we provide the first empirical evidence supporting the proposed taxonomy, and a set of principles to structure reasoning when choosing a decoding algorithm. Crucially, our analysis is the first one to relate likelihood-based decoding strategies with strategies that rely on external information such as value-guided methods and prompting, and covers the most diverse set of tasks up-to-date.

Large language models (LLMs) have achieved impressive results in natural language processing but are prone to memorizing portions of their training data, which can compromise evaluation metrics, raise privacy concerns, and limit generalization. Traditional methods for detecting memorization rely on output probabilities or loss functions, often lacking precision due to confounding factors like common language patterns. In this paper, we introduce an analytical method that precisely detects memorization by examining neuron activations within the LLM. By identifying specific activation patterns that differentiate between memorized and not memorized tokens, we train classification probes that achieve near-perfect accuracy. The approach can also be applied to other mechanisms, such as repetition, as demonstrated in this study, highlighting its versatility. Intervening on these activations allows us to suppress memorization without degrading overall performance, enhancing evaluation integrity by ensuring metrics reflect genuine generalization. Additionally, our method supports large-scale labeling of tokens and sequences, crucial for next-generation AI models, improving training efficiency and results. Our findings contribute to model interpretability and offer practical tools for analyzing and controlling internal mechanisms in LLMs.

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

Language understanding entails not just extracting the surface-level meaning of the linguistic input, but constructing rich mental models of the situation it describes. Here we propose that because processing within the brain's core language system is fundamentally limited, deeply understanding language requires exporting information from the language system to other brain regions that compute perceptual and motor representations, construct mental models, and store our world knowledge and autobiographical memories. We review the existing evidence for this hypothesis, and argue that recent progress in cognitive neuroscience provides both the conceptual foundation and the methods to directly test it, thus opening up a new strategy to reveal what it means, cognitively and neurally, to understand language.