LLM Interpretability

LLM: Large Language Model

Follow this topic Ask about this topic

Momentum

57 papers in the last four weeks, up 73% on the four weeks before. 0.6% of all new papers.

Jul 13Week of Sep 28

Latest papers 546

Jun 4, 2026cs.CL

Interpreting Style Representations via Style-Eliciting Prompts

Style representation learning is a powerful tool for authorship analysis and modeling writing style, yet the latent nature of learned representations makes them difficult to interpret. Recent work has attempted to explain these representations by generating natural language descriptions with large language models (LLMs) conditioned on input text. However, such descriptions are often prone to the LLM's biases and hallucinations, and they lack an explicit objective and practical utility. In this work, we propose a novel framework for interpreting style representations through style-eliciting prompts: natural language instructions designed to steer LLMs to generate text that reflects specific stylistic attributes. We curate 1,010 distinct style features spanning 26 stylistic categories and construct a dataset by prompting an LLM to generate text conditioned on these features. Using this data, we train a decoder to generate a style prompt from the style representation of the generated text. We evaluate our approach on three tasks: (1) recovering original style prompts from generated text, (2) generating text in the same style using the recovered prompts, and (3) steering LLM outputs to match the style of human-written texts. Experiments demonstrate that our method consistently outperforms strong baselines that directly prompt LLMs with target text, achieving superior performance in both style description and style imitation. These results highlight that style-eliciting prompts can provide a practical and interpretable interface to stylistic information encoded in style representations.

Junghwan Kim, David Jurgens

University of Michigan

LLM Prompting LLM Interpretability Controllable Text Generation Neural Network Interpretability Text Generation

Jun 4, 2026cs.LG

A Navigable Manifold of Hypothesized Consciousness-Spectrum States in Language Model Representations

Across contemplative, philosophical, and psychological accounts, human consciousness is often described along a similar spectrum, ranging from reactive and self-focused patterns to more integrative and coherent ones. Understanding whether language models encode such a structured, human-interpretable consciousness spectrum in representation space is important for model guidance, evaluation and alignment. In this work, we study the geometric structure and dynamics of patterns along this spectrum in transformer embedding spaces. We show that embeddings exhibit a globally organized geometry aligned with this spectrum: sentences associated with similar states cluster into locally coherent regions, forming a structured manifold. In particular, higher-level and lower-level regions exhibit convexity-like stability, while intermediate regions form a transition corridor. Dynamically, both utility-guided and geometry-only greedy trajectories consistently traverse from lower- to higher-level regions, passing through intermediate tiers, indicating that navigability is an intrinsic property of the representation space, guided but not dictated by a global directional signal. These results suggest that embedding spaces encode structured and navigable geometry aligned with a hypothesized consciousness-spectrum taxonomy, broadly inspired by recurring structural descriptions of human consciousness across contemplative traditions, philosophy, and modern psychology, providing a representation-level perspective for analyzing and guiding model behavior.

Sophie Zhao

School of Computer Science Georgia Institute of Technology

Neural Representation Geometry LLM Interpretability Representation Geometry in Language Models

Jun 3, 2026cs.CL

Localizing Prompt Ambiguity in Large Language Models with Probe-Targeted Attribution

Prompt ambiguity is a common source of failure in large language models, but is difficult to localize because it is a latent property of the prompt, while existing attribution methods are designed to explain observable outputs such as logits or generated tokens. We introduce PRIG, a gradient attribution method that uses a probe logit to attribute latent ambiguity to token positions. Specifically, PRIG trains a linear probe to distinguish clear prompts from ambiguous prompts and attributes the probe score to earlier token representations in the residual stream. To enable token-level evaluation, we construct synthetic ambiguity datasets across coding, math, and writing by rewriting one task-critical sentence per prompt, and complement them with a human-written gold benchmark. In this setting, PRIG localizes ambiguous spans substantially better than gradient attribution baselines, achieving 0.840 AUROC on the combined synthetic benchmark and 0.891 AUROC on the gold set. It also outperforms GPT-5.4 on sentence-level ambiguity identification and retains useful signal out-of-domain. These results establish PRIG as a practical tool for identifying which parts of a prompt are ambiguous. More broadly, they suggest that latent prompt properties can be localized through intermediate representations, rather than through output-level attribution.

Govind Ramesh, Yao Dou, Wei Xu

Georgia Institute of Technology

Gradient-Based Attribution LLM Interpretability

Jun 3, 2026cs.LG

Trust, but Don't Verify: Epistemic Blind Spots in LLM Source Evaluation

Language models increasingly act as epistemic proxies, synthesizing evidence from multiple sources to inform decisions. Whether they evaluate the quality of that evidence, or merely aggregate it based on surface presentation, remains poorly understood. We show that models possess the capability to detect fabricated statistics in isolation but do not recruit this capability during multi-source synthesis, producing similar numeric estimates whether the statistics are fabricated or valid. Specifically, source influence is governed by a methodology-register gate that responds to the distributional register of analytical text but not to numeric validity: for example, statistically impossible confidence intervals receive the same weight as valid ones. The behavioral dissociation replicates across six models from four families (Anthropic Claude, Qwen, OLMo, and OpenAI GPT-5.4) and three professional domains. Mechanistic analyses, including causal tracing, linear probes, and component-level attribution, converge on the same account: the model encodes and causally uses a methodology-register representation that transfers across domains, while numeric-validity signals, decodable in isolation, are suppressed to chance during multi-source synthesis. Prompting-based mitigations, even an oracle checklist naming the exact statistical checks, produce blanket skepticism rather than selective discernment, and the post-training pipelines we examine reinforce the shortcut without building numeric verification. Unlike sycophancy, which tracks user preference, this failure tracks whether a source presents as analytically credible, not whether its claims are consistent. We term this \textit{epistemic alignment}: like preference and safety alignment, the question is not capability but deployment.

Rohan N. Pradhan, Steve Goley

Amazon

LLM Interpretability LLM Reliability

Jun 3, 2026cs.LG

Pattern Selectivity is Not Task-Causal Structure: A Cross-Architecture Mechanistic Study of Composed-Task Circuits in 1B-Class Language Models

We test whether a single screen-and-ablate recipe -- identify attention-head circuits by task-pattern selectivity, then verify by causal ablation against a matched-random null -- produces consistent mechanistic claims across model families. The recipe ports across pipelines; the specific circuit it identifies does not. Across four composed tasks (indirect-object identification, greater-than, successor sequences, variable binding) and three 1B-class language models from distinct training pipelines (Pythia 1B / Pile / dense; OLMo 1B / DCLM / dense; OLMoE 1B-7B / DCLM / mixture-of-experts), we run a unified protocol with the matched-random null sampled across ten seeds per cell. The resulting 12 (task, model) cells contain no two that share the same primary causal screen at comparable effect size: the same task, with the same behavioral capability, is implemented through different attention-pattern types across models. We introduce a five-category screen-outcome taxonomy -- primary cause, secondary cause, correlate, interferer, null -- with quantitative thresholds, and show that all five outcomes appear in the panel. We propose a falsifiable hypothesis: the MoE model in our panel builds composed-task circuits on top of a foundational previous-token positional substrate (the prev-token-circuit ablation is the strongest causal screen on 3 of 4 tasks for OLMoE 1B-7B), with the IOI exception consistent with IOI being a final-position name-copying task whose structure directly probes a different pattern. The hypothesis comes with explicit predictions for other MoE language models. We frame the methodology honestly: the spectral participation-ratio signal from the companion methodology paper is a general indicator of specialized computation; what makes a finding task-specific is the task-pattern screen plus a per-model causal verification.

Yongzhong Xu

Attention Head Analysis Decoder-Only Language Models LLM Interpretability Circuits in Language Models Causal Interventions in Language Models Language Model Probing

Jun 3, 2026cs.CL

Fast & Faithful Function Vectors

Function vectors (FVs) are task representations elicited during in-context learning that can be used to steer Large Language Models (LLMs). However, design choices in their formulation remain underexplored. In this work, we study the impact of varying FV definitions for instructions along two degrees of freedom: attention head selection and steering. For head selection, using gradient-based attributions with Layer-wise Relevance Propagation (LRP) substantially improves efficiency as well as accuracy. For FV steering, applying it in a distributed manner yields a higher accuracy compared to simple aggregation. Our code is publicly available.

Minh An Pham, Anton Segeler, Thomas Wiegand +4

Fraunhofer Heinrich-Hertz-Institute, Berlin, Germany · Technische Universität Berlin, Berlin, Germany · BIFOLD – Berlin Institute for the Foundations of Learning and Data, Berlin, Germany +1

Task Vectors LLM Interpretability In-Context Learning Activation Steering

Jun 3, 2026cs.LG

Data Attribution in Large Language Models via Bidirectional Gradient Optimization

Large Language Models (LLMs) are increasingly deployed across diverse applications, raising critical questions for governance, accountability, and data provenance. Understanding which training data most influenced a model's output remains a fundamental open problem. We address this challenge through training data attribution (TDA) for auto-regressive LLMs by expanding upon the inverse formulation: How would training data be affected if the model had seen the generated output during training? Our method perturbs the base model using bidirectional gradient optimization (gradient ascent and descent) on a generated text sample and measures the resulting change in loss across training samples. Our framework supports attribution at arbitrary data granularity, enabling both factual and stylistic attribution. We evaluate our method against baselines on pretrained models with known datasets, and show that it outperforms previous work on influence metrics, thereby enhancing model interpretability, an essential requirement for accountable AI systems.

Frédéric Berdoz, Luca A. Lanzendörfer, Kaan Bayraktar +1

ETH Zurich

Gradient-Based Attribution LLM Interpretability Training Data Attribution

Jun 3, 2026q-bio.NC

Do Large Language Models Have Emotions?

Do LLMs have emotions? A recent paper from Anthropic reports finding internal representations of emotion concepts in Claude Sonnet 4.5, concluding that the LLM has 'functional emotions.' We evaluate this claim against what is known about how emotions actually function in biological systems. We argue that emotions serve two core functions: the context-sensitive interpretation of situations, and the reorganization of processing across multiple systems in response to those interpretations. The Anthropic findings offer partial support for the first function, though the consistent, discrete emotional representations identified in Claude sit uneasily with affective neuroscience findings that human emotion is characterized by variable rather than uniform neural signatures. On the second function, the evidence is mixed: Claude's representations modulate output without producing the dynamic reorganization of attention, decision speed, and motivational state that defines emotion in biological systems. We close by proposing what it would take for an LLM to have emotions.

Amit Goldenberg, James J. Gross

Harvard Business School · Department of Psychology, Harvard University · Harvard University, Digital, Data and Design Institute +1

LLM Evaluation LLM Interpretability Emotion Representation in Language Models Affective Computing

Jun 2, 2026cs.LG

Sparse Mixture-of-Experts Reward Models Learn Interpretable and Specialized Experts for Personalized Preference Modeling

Preference modeling plays a central role in reinforcement learning from human feedback (RLHF), enabling large language models (LLMs) to align with human values. However, most existing approaches assume a universal reward function, neglecting the diversity and heterogeneity of human preferences. To address this limitation without additional annotation costs, recent work has proposed learning multiple preference components from binary data and combining them to model individual preferences. Nevertheless, these components often fail to capture coherent and disentangled patterns, limiting their interpretability and effectiveness for personalization. In this work, we propose a sparse Mixture-of-Experts (MoE) reward model that encourages sparse routing and expert diversity during training on binary preference data. Across controlled and real-world experiments, sparse MoE learns interpretable routing patterns and specialized experts. It also improves test-time personalization, and post-adaptation shifts in expert weights provide a qualitative lens for analyzing how the model adapts to personalized preferences.

Yifan Wang, Jinyi Mu, Mayank Jobanputra +5

Saarland University · Independent Researcher · Bielefeld University +2

Pairwise Preference Learning Reward Modeling LLM Interpretability Sparse Mixture-of-Experts User Preference Modeling

Jun 2, 2026cs.CL

Language Models Compare Quantities Using Number-specific and Unit-specific Heuristics

Quantities with measurement units, such as 110 cm and 1.2 m, require language models (LMs) to combine a numeral with a symbolic unit scale. Here, we study how LMs compare such quantities in controlled settings spanning several unit systems. We find that accuracy degrades near the comparison boundary, where small changes in value determine the correct answer. The resulting errors are systematic: linear surrogate models predict LM preferences from numerical-difference and unit-scale-difference cues, and causal interventions on subspaces aligned with these variables shift model's output. The results suggest that LMs compare quantities through a bag of heuristics over numerals and units, rather than first converting both expressions to an exact shared-scale representation.

Mutsumi Sasaki, Go kamoda, Ryosuke Takahashi +4

Tohoku University · SOKENDAI · NINJAL +2

Numerical Reasoning in Language Models LLM Evaluation Pairwise Comparison LLM Interpretability

Jun 2, 2026cs.CL

Framing Migration News with LLMs: Structured CoT as a Support for Human Interpretation

Frame analysis of migration news is a socially consequential task: media scholars and researchers who study how migration is narrated need tools that are not only accurate, but transparent, auditable, and accessible within the resource constraints typical of academic research groups. Existing LLM-based approaches rely on proprietary APIs and large models that raise concerns about data privacy, reproducibility and equitable access among media researchers. This work studies how a locally deployable open-source LLM can support interpretable frame analysis as an assistive tool. We introduce a Structured Chain-of-Thought (SCoT) prompting approach using Llama3-8B, enabling step-by-step justifications grounded in predefined framing categories. This structured design allows users to audit model outputs and examine alternative interpretations in a task that is inherently subjective. We evaluate our approach on a dataset of migration-related news and show that SCoT improves classification performance over zero-shot and few-shot baselines while remaining feasible on a single GPU. Then, we conduct a human-centered evaluation in which annotators assess the coherence and influence of "the model's reasoning". Results indicate that SCoT explanations are generally perceived as logical (mean score 4.1/5, though with notable variation across texts) and can prompt reflection on initial interpretations, even when disagreement persists. Our findings highlight both the potential and risks of LLM-assisted frame analysis. While structured reasoning can increase the traceability of model outputs and support critical interpretation, it can also influence human judgment in subtle ways. By enabling local deployment and emphasizing human-in-the-loop interaction, this work contributes to discussions on responsible and accessible computational tools for the study of socially impactful media narratives.

David Alonso del Barrio, Jing Wen, Daniel Gatica-Perez

Idiap Research Institute and EPFL Switzerland

LLM Interpretability CoT Reasoning LLM-Assisted Annotation Human-in-the-Loop Annotation

Jun 2, 2026cs.LG

When Graph Tokens Sink: A Mechanistic Analysis of Graph Language Models

Graph Language Models (GLMs) have become a promising direction for adapting Large Language Models (LLMs) to graph learning tasks. By transforming graph topology and node information into graph tokens, GLMs allow LLMs to jointly process structured graph inputs and textual instructions. Yet, it remains unclear how LLMs internally interpret these graph tokens and whether graph tokens act as meaningful carriers of graph structure. In this work, we analyze how LLMs process graph information through graph-token behavior in representative GLM architectures. Findings. We find that the internal saliency of graph tokens in GLMs is not equivalent to graph information utilization. Graph sink tokens consistently emerge as activation-level outliers: they can be identified by massive activation values along a small set of hidden-state dimensions and are biased toward early graph-token positions. However, this activation-level saliency does not imply that these tokens are the main carriers of graph information. Unlike classical attention sinks in language and vision-language models, graph sink tokens do not necessarily attract the largest attention weights from query tokens. Through pruning, repositioning, and swapping interventions, we show that graph sink tokens are not the most important semantic or structural tokens for downstream prediction. Implications. Together, these results suggest that after current GLMs map graph structure into the LLM token space, the resulting graph-token representations do not naturally form a fully usable topology-aware internal representation; instead, they exhibit a decoupling between activation-level saliency and graph-semantic utility. This decoupling points to limitations in existing graph-token construction, placement, and alignment mechanisms.

Ding Zhang, Runtao Zhou, Wenqing Zheng +3

University of Virginia · Capital One

LLM Interpretability Mechanistic Interpretability Language Modeling Attention Sinks LLMs for Graph Learning

Jun 2, 2026cs.LG

A Close Look At World Model Recovery In Supervised Fine-Tuned LLM Planners

Supervised fine-tuning (SFT) improves end-to-end classical planning in large language models (LLMs), but do these models also learn to represent and reason about the planning problems they are solving? Due to the relative complexity of classical planning problems and the challenge that end-to-end plan generation poses for LLMs, it has been difficult to explore this question. In our work, we devise and perform a series of interpretability experiments that holistically interrogate world model recovery by examining both internal representations and generative capabilities of fine-tuned LLMs. We find that: a) Supervised fine-tuning on valid action sequences enables LLMs to linearly encode action validity and some state predicates. b) Models that struggle to use output probabilities for classifying action validity may still learn internal representations that separate valid from invalid actions. c) Broader state space coverage during fine-tuning, such as from random walk data, yields more accurate recovery of the underlying world model. In summary, this work contributes a recipe for applying interpretability techniques to planning LLMs and generates insights that shed light on open questions about how knowledge is represented in LLMs.

Patrick Emami, Nan Qiang, Peter Graf

National Laboratory of the Rockies · Epic Systems

Supervised Fine-Tuning LLM Planning LLM Interpretability

Jun 2, 2026cs.AI

Decomposing how prompting steers behavior

Prompting steers large language models (LLMs) and vision-language models (VLMs) without weight updates, but it remains unclear how instruction changes reshape internal representations to produce behavior. We introduce a nested geometric decomposition framework that treats prompting as a transformation of the representational geometry of the content following the prompt. For each prompt pair, we align representations of the same stimuli under two prompts using increasingly expressive stimulus-invariant maps: translation, rigid transformation with uniform scaling, sequential axis scaling, affine transformation, and nonlinear transformation. We then causally test each map by replacing a single layer's prompt-A hidden state for held-out stimuli with its mapped counterpart and measuring recovery of prompt-B representational geometry and behavior. Across three LLMs, three VLMs, and six text or image datasets spanning style, emotion, scene content, and number, prompts consistently reshape representations toward the instructed task structure. Cross-validated variance decomposition shows that much prompt-induced activation change is captured by shape-preserving maps, especially translation and rigid transformation with uniform scaling, while tier profiles reveal model- and task-specific routing strategies across layers. Crucially, although translation and rigid tiers already improve behavioral agreement, affine transformation is the first tier to nearly recover target-prompt task geometry and yields corresponding behavioral gains. This suggests that cross-dimensional linear mixing is a key mechanism by which prompts reorganize representations toward instructed task structure. Our framework decomposes prompt-induced representational change into interpretable geometric components and reveals how models route task-relevant structure to produce prompt-driven behavior.

Fan L. Cheng, Nikolaus Kriegeskorte

Columbia University

LLM Prompting LLM Interpretability VLM Interpretability Mechanistic Interpretability Causal Interventions in Language Models Representation Geometry in Language Models

Jun 2, 2026cs.LG

Multi-component Causal Tracing in Large Language Models

Causal tracing systematically intervenes on a large language model's (LLM's) internal representations to uncover and quantify the causal pathways linking specific inputs or computations to specific metrics of interest, quantifying the LLM's behavior. Building on previous single-component or single-layer studies, this paper presents a unified framework for causally tracing multiple components simultaneously. This framework systematically identifies the subsets of components (e.g., attention heads and multi-layer perceptron neurons) most critical to a desired target performance metric (e.g., accuracy and fairness). This is achieved by incorporating flexible interventions applied to a wide range of desired metrics. To address the combinatorial complexity of the multi-component problem, an efficient algorithm is designed that leverages soft interventions and a carefully designed metric transformation, converting the combinatorial search problem into a continuous one that can be solved efficiently under proper constraints, thereby generating proper binary decisions for selecting components. Experimental results demonstrate that the proposed method efficiently identifies subsets of the model's components that have a high impact on the target metric, outperforming existing baseline approaches. Our code is available at https://github.com/ZiruiYan/multi-component-causal-tracing.

Zirui Yan, Dennis Wei, Dmitriy A. Katz +2

Rensselaer Polytechnic Institute · IBM Research

LLM Interpretability Causal Reasoning in Language Models Causal Interventions in Language Models

Jun 2, 2026cs.LG

Perplexity Can Miss SAE Feature Damage Under Quantization

Quantization is a standard path to deploying large language models, and quantized models are typically judged acceptable when perplexity or downstream accuracy remains close to the full-precision original. But behavioral parity need not imply feature fidelity: the sparse-autoencoder (SAE) features used to interpret a full-precision model may change after weight rounding. We test this directly by using a frozen SAE as a fixed measurement basis, encoding full-precision and round-to-nearest (RTN) quantized activations on identical tokens, and measuring per-feature survival by Pearson correlation across bit-widths from INT8 to INT4 on Pythia-70M and Gemma-2-2B. Our central finding is that perplexity can miss feature damage: on Gemma-2-2B, INT7 improves perplexity while degrading 18.7% of active SAE features, and under sliding-window evaluation INT6 also improves perplexity while only 51.3% of active features survive. Feature survival is graded rather than cliff-like, with 62.4% of active Pythia features and 51.3% of active Gemma features surviving at INT6; most non-surviving features are blurred rather than fully damaged. Survival is also predictable from full-precision feature statistics alone, with cross-validated AUC 0.92--0.97 and peak activation as the strongest marginal predictor. Finally, RTN quantization and matched-perplexity magnitude pruning damage strongly overlapping feature sets, with Jaccard overlap 0.79--0.86 and damage-score Spearman correlation 0.98. These results show that behavioral metrics alone are insufficient evidence that full-precision interpretability findings transfer to quantized models, motivating feature-level audits of compression.

Evan Duan

University of Michigan

LLM Quantization LLM Interpretability Sparse Autoencoders Post-Training Quantization

Jun 1, 2026cs.CL

Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States

Linear probing of large language model (LLM) hidden states is widely used to claim that models learn distinct representations for different reasoning types. We test this by probing Qwen3-14B on three benchmarks spanning the classical trichotomy: LogiQA 2.0 (deductive), ARC-Challenge (inductive), and

α

NLI (abductive). At layer 32 of 40, linear probes achieve 100% cross-validated accuracy with well-separated geometry (intrinsic dimensionalities: 20.6, 28.5, 33.6; convex hull contamination

\leq

1.5%). However, this separation is entirely driven by format confounds. Residualizing source identity, option count, and response length reduces accuracy to chance. Trace-anchor similarity indicates largely shared reasoning across tasks (42.5% agreement vs.\ 33.3% chance), and causal steering with random controls (

n=20

) shows no functional link between geometry and reasoning mode (

p=0.286

). Thus, high probe accuracy reflects task format rather than computational structure, motivating routine format deconfounding in mechanistic interpretability.

Subramanyam Sahoo, Vinija Jain, Aman Chadha +1

Horizon Research · Meta · Apple +1

LLM Interpretability Representation Geometry in Language Models Language Model Probing Hidden-State Probing LLM Reasoning

Jun 1, 2026cs.CL

On the Persistent Effects of Lexicality in Large Language Models

Representations extracted from large language models (LLMs) play an important role in many downstream applications. However, the structure of these representations is often influenced by lexical overlap rather than semantic content. Our understanding of the relationship between this lexical influence and semantic content, and its implications for downstream tasks, remains limited. In this work, we investigate representations to quantify the effect of lexical overlap relative to semantic content. We consider several adversarial semantic stress tests and further connect our findings to the information theory perspective. We find that lexical influence extends across the depth of models, consistently across architectures, training regimes, and objective functions, including the models trained for semantic similarity. Moreover, we observe a mid-depth region in which both lexical and semantic signals degrade simultaneously, indicating a transitional regime where representations are poor for both surface form and meaning. We further demonstrate the effect of lexical influence on downstream uses of LLMs using summarization and model editing as a case study.

Hammad Rizwan, Muhammad Umair Haider, Nishant Subramani +3

Dalhousie University Halifax, NS · University of Kentucky Lexington, KY · Carnegie Mellon University Pittsburgh, PA

LLM Interpretability Language Modeling Linguistic Bias in Language Models

May 31, 2026cs.AI

An Enigma of Artificial Reason: Investigating the Production-Evaluation Gap in Large Reasoning Models

Studies of human reasoning have shown that people are typically stronger at evaluating reasoning than producing it from scratch. In contrast, large reasoning models (LRMs) are trained to excel at producing long chains of reasoning to solve complex problems. How then do LRMs perform at evaluating reasons? We investigate this with the Valid-Answer-Invalid-Reasoning (VAIR) dataset: math problems and solutions with trivial reasoning flaws but valid answers, designed to isolate reasoning evaluation from the confound of reasoning production. Unlike humans, who we find are only 6% worse at grading than solving such problems, we find a substantial production-evaluation gap in LRMs: frontier models score as low as 48% when evaluating VAIR solutions, despite near-perfect solution production. Why this enigma? Through chain-of-thought (CoT) analysis, we find evidence of an answer confirmation bias: LRMs often produce then check for the correct answer instead of carefully verifying each step, fabricating rationalizations even when noticing anomalous reasoning. Linear probes corroborate this, showing that while LRM activations encode some representation of valid reasoning, they fail to robustly represent VAIR solutions as invalid. Causal patching of the final answer's representations causes LRM verdicts and activations to flip, demonstrating that answer validity is responsible for models' confirmation biases. These findings indicate an outstanding limitation in dominant approaches to reasoning training, which incentivize LRMs to produce and confirm reasoning towards correct answers, but not to robustly evaluate the underlying reasons.

Mingzhong Sun, Teresa Yeo, Armando Solar-Lezama +1

1NUS Department of Computer Science · 4Singapore-MIT Alliance for Research and Technology (SMART) · 2MIT EECS +1

Reasoning Evaluation LLM Interpretability Large Reasoning Models Reasoning Verification

May 31, 2026cs.CL

Unlocking the Black Box of Latent Reasoning: An Interpretability-Guided Approach to Intervention

Latent reasoning enables Large Language Models (LLMs) to perform multi-step inference within continuous hidden states, offering efficiency gains over explicit Chain-of-Thought (CoT). However, the opacity of these continuous thought vectors hinders their reliability and controllability. This paper bridges the gap between mechanistic interpretability and actionable control. We first present a systematic analysis using structural, causal, and geometric probes, revealing that latent vectors encode compressed, faithful representations of reasoning steps, with early vectors acting as critical causal hubs. Building on this, we operationalize these interpretability insights into a suite of training-free, decode-time interventions that refine the latent reasoning process by imposing the identified geometric and semantic priors. Extensive experiments across multiple model scales and diverse task domains demonstrate that our approaches consistently improve reasoning accuracy. Our interpretability-guided interventions consistently unlock latent capabilities and improve reasoning accuracy without any parameter updates.

Shuochen Chang, Tong Bai, Xiaofeng Zhang +5

1Shanghai Jiao Tong University · 2Fudan University

LLM Interpretability Causal Interventions in Language Models LLM Reasoning Latent Reasoning

May 31, 2026cs.AI

The Shape of Wisdom: Decision Trajectories in Language Models

Language models do not simply choose an answer at the output layer. In a 9,000-trajectory MMLU study across Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct, and Mistral-7B-Instruct-v0.3, the score of the answer moves across depth in structured ways. We describe each trajectory with three quantities: the current answer margin, the next-layer change in that margin, and the distance from a decision flip. The main empirical picture is that correctness and stability are different: the largest group is unstable-correct, not stable-correct. A traced subset then asks what moves the margin. In stable-correct cases, the average attention scalar points in the correct direction, while the average MLP scalar does not; span deletion shows that removing answer-supporting text hurts the margin and removing distractor-like text helps it. The result is not a full circuit explanation. It is a reproducible way to see which answers are settled, which remain fragile, and which measured sources move them.

Shailesh Rana

Independent Researcher

LLM Interpretability

May 31, 2026cs.AI

The Case for Model Science: Verify, Explore, Steer, Refine

We argue that the AI community is now ready to move beyond benchmarking and consolidate scattered efforts in model analysis into a systematic discipline, a direction we term Model Science. Complex AI models now serve billions of users, yet our understanding of how they work lags far behind our ability to deploy them. Decades of benchmark-driven research have delivered remarkable progress: extensive leaderboards, a wide range of performance metrics, tracking capability gains across diverse tasks; yet this success has also revealed the limits of benchmarks as they tell us whether models perform but not why they succeed or fail, they miss critical failure modes, such as hallucinations or shortcuts. Precedents from established sciences point the way forward: cognitive science shows that understanding complex systems requires complementary levels of analysis; neuroscience demonstrates that deep study of single cases reveals what population studies miss; medicine teaches that specialised training must develop alongside research practice; and agriculture models how shared infrastructure and principles enable cumulative progress. These lessons inform three foundations for Model Science. First, we propose to consolidate research around four functional perspectives: Verify, Explore, Steer, and Refine that address complementary questions about model behaviour. Second, we discuss the required infrastructure for cumulative knowledge: catalogues of datasets, models and findings. Third, we highlight the need for deep analysis of individual model instances, not just model families, because single cases can reveal what population studies miss.

Przemyslaw Biecek, Luca Longo, Jianlong Zhou +3

Center for Credible AI · University of Warsaw · Warsaw University of Technology +7

LLM Evaluation LLM Interpretability

May 31, 2026cs.CL

Not All Explanations Simulate Equally: Comparing Verbalized Feature Attributions and Self-Generated Rationales

Natural-language explanations are often treated as a unified interface for understanding model behavior, but different explanation sources may support simulation in different ways. This paper compares two families of explanations for question answering models: verbalized feature attributions and self-generated rationales. We evaluate them under a shared counterfactual simulation setting, using an LLM judge as predictor and measuring whether it can better predict a model's answers to follow-up questions when given its explanation. Across multiple instruction-tuned models, we analyze how explanation source, verbalization strategy, and feature granularity affect the simulatability of explanations. Our results show that explanation format and granularity affect simulatability: attribution-based explanations and self-generated rationales differ in how much they improve counterfactual prediction, with effects that vary across models and formats.

Pingjun Hong, Benjamin Roth

Faculty of Computer Science, University of Vienna, Vienna, Austria · UniVie Doctoral School Computer Science, University of Vienna, Vienna, Austria · Faculty of Philological and Cultural Studies, University of Vienna, Vienna, Austria

Feature Attribution LLM Interpretability Explainability Evaluation

May 31, 2026cs.CL

MENTIS: What Belief Changes Under Alignment? Measuring Multi-Scale Latent Torsion in Language Models

Preference alignment has substantially improved the observable behavior of large language models, yet it remains unclear what alignment changes internally. Aligned systems still fail under jailbreaks, prompt injection, and retrieval-time corruption, suggesting behavior-level evaluation alone is incomplete. Post-training should leave measurable traces in internal computation. We ask: when an instruction-tuned (IT) model becomes a preference-aligned (PA) model, what geometric structure changes, where do those changes concentrate, and how selectively do they vary across concepts, prompts, and model families? We introduce MENTIS, a geometry-first framework for measuring alignment-induced internal reorganization in paired checkpoints. MENTIS compares IT and PA models using a primary layerwise covariance-based torsion norm (T1), a secondary spectral torsion diagnostic (T2), and an Energy-Radiance-Activation measure (ERA) for depth localization. Across four 7-8B model pairs on LITMUS, our study reveals that alignment-induced change is selective rather than uniform: normative concepts exhibit larger torsion shifts than factual concepts on average; torsion is negatively correlated with contextual entropy; and peak effects localize to architecture-specific mid-to-late layers. The same pattern appears across word-level, prompt-level, and model-level analyses. These results suggest preference alignment leaves structured, depth-localized geometric signatures in internal computation beyond what behavior-level evaluation alone can reveal.

Partha Pratim Saha, Samarth Raina, Mayur Parvatikar +4

Pragya Lab, BITS Pilani Goa, India · IIIT Delhi, India · Amazon, USA +2

LLM Interpretability Preference Alignment Representation Geometry in Language Models

May 30, 2026cs.LG

Query Lens: Interpreting Sparse Key-Value Features with Indirect Effects

While sparse autoencoders provide features more interpretable than individual neurons, reliably characterizing them remains challenging. We propose Query Lens, which extends Logit Lens to enable more comprehensive and faithful interpretations of sparse features. By jointly considering encoder-side key features and decoder-side value features, we identify both the inputs that activate a feature and the outputs it promotes. We also account for indirect, module-mediated effects that arise when the feature is processed by downstream modules, going beyond the direct effect captured by Logit Lens. In experiments, we find that Query Lens yields coherent token signatures for features that remain uninterpretable under Logit Lens. Finally, we propose the Subspace Channel Hypothesis, suggesting that downstream modules read features through layer-specific subspaces.

Hwiyeong Lee, Ingyu Bang, Uiji Hwang +2

1Hanyang University, Seoul, Republic of Korea.

Representation Learning LLM Interpretability Sparse Autoencoders Mechanistic Interpretability

May 29, 2026cs.CL

How Far Do Auto-Interpretation Labels Generalize: A Controlled Study Across Languages, Scripts, and Rewordings

Sparse autoencoder (SAE) features are increasingly used to interpret language models, with auto-generated natural-language labels serving as the primary interface for understanding what each feature represents. We ask whether these labels generalize: does a feature labeled for a concept actually track that concept across languages and scripts? Using Serbian digraphia as a controlled testbed -- the same language written in both Latin and Cyrillic via deterministic transliteration -- we first find that SAE feature sets activated by the same content in different languages, scripts, and wordings share substantial overlap (mean Jaccard 0.39 vs 0.13 random baseline, peaking at 0.57), suggesting genuine cross-lingual semantic features. We then test whether auto-interpretation labels keep pace. They often do not: features whose labels describe semantic content miss the same meaning in Serbian up to 4

\times

more often than within English, and miss Serbian Cyrillic more than Serbian Latin -- two scripts that are deterministic transliterations of each other. The gap grows with network depth, yet the labels give no indication that they fail. These results suggest that auto-interpretation labels reflect a feature's behavior on the languages and scripts a model has seen most in training, rather than the concept itself.

Sripad Karne

Columbia University

LLM Interpretability Sparse Autoencoders Automated Evaluation

May 29, 2026cs.CL

The Latin Substrate: How Language Models Represent and Mediate Script Choice

Many languages are written in multiple scripts, requiring large language models (LLMs) to generate equivalent linguistic content in distinct orthographic forms. While prior work suggests that LLMs route information through shared latent representations, how they internally mediate script variation remains poorly understood. We study this question by first examining per-layer output distributions with the logit lens, which reveals consistent latent romanization during transliteration, and then through representational and mechanistic analyses of script generation. At the representational level, we show that scripts of the same language become increasingly separable across layers and that a simple linear steering direction can flip a model's output script while largely preserving semantic content. The vector generalizes to writing systems unseen during construction: it flips non-Latin output to Latin, but maps Latin output into varied non-Latin scripts. At the mechanistic level, we localize a small set of late-layer attention heads that causally mediate script choice. These heads transfer across unrelated languages and writing systems, suggesting that script routing is implemented by language-agnostic components. Across both analyses, we observe a consistent directional asymmetry: non-Latin output is produced by a compact, identifiable gate, while Latin-script output emerges from diffuse contributions across the model. Overall, our findings hint that LLMs organize script variation around shared latent representations while exhibiting a privileged substrate toward Latin script.

Daniil Gurgurov, Alan Saji, Katharina Trinley +2

D German Research Center for AI (DFKI) · I Indian Institute of Technology, Madras · H University of Hamburg +1

Multilingual Language Models LLM Interpretability

May 29, 2026cs.CL

Steering LLMs? Actually, Sparse Autoencoders can outperform simple baselines

Sparse Autoencoders (SAEs) have been seen as a promising avenue for exploring the internals of Large Language Models (LLMs) and for steering model output generation. When AxBench - a model steering benchmark - was introduced in Wu et al. (2025), SAEs did not seem to live up to their original hype due to poor steering performance relative to a set of simple baselines. This work serves as a partial rebuttal for Sparse Autoencoders and suggests that the results of Wu et al. (2025) did not do them full justice. We find that Sparse Autoencoders can, in fact, perform close to on par with the reference LoRA performance on the AxBench benchmark, when features are selected and labelled with our supervised pipeline. We also find that our pipeline selects features that are surprisingly causal of their identified labels when using only its interpretability-based components. Lastly, we present evidence that high sparsity (low l0) may not be crucial for successful steering based on interpretability, which is in contrast to the earlier findings in Wang et al. (2025).

Mikkel Godsk Jørgensen, Lars Kai Hansen

DTU Compute · Technical University of Denmark

Language Model Steering LLM Interpretability Sparse Autoencoders

May 29, 2026cs.LG

The Shape of Addition: Geometric Structures of Arithmetic in Large Language Models

Large Language Models exhibit paradoxical fragility in fundamental arithmetic, implying a disconnect between internal computation and discrete output. By analyzing the residual stream geometry during multi-operand addition, we identify the Iso-Raw-Sum Trajectory (IRST), a geometric structure where representations are anchored by semantic digits and modulated by continuous carry fibers. We propose the Noisy Quantization Model to explain this geometry, framing arithmetic errors as Geometric Slippages caused by internal neural noise pushing a continuous, latent Carry Potential across quantization thresholds. This geometric framework further elucidates Probe Versatility, explaining how lightweight probes can disentangle coexisting latent signals (such as ground truth versus hallucination) from a single activation vector. Finally, we validate these insights through a geometric consistency check method that effectively detects and corrects these quantization failures during inference. Our code is available at https://github.com/RL-MIND/Shape-of-Addition.

Liuyuan Wen, Xun Zhu, Lihao Huang +2

State Key Laboratory of Novel Software Technology, Nanjing University, Nanjing 210023, China.

Neural Representation Geometry LLM Interpretability LLM Mathematical Reasoning

May 29, 2026cs.CL

Toxic HallucinAItions: Perturbing Prompts and Tracing LLM Circuits

Large language models (LLMs) are increasingly deployed in conversational settings where user tone ranges from polite to adversarial or toxic, yet less is known about whether toxic language in otherwise semantically equivalent prompts can degrade factual reliability. We study how lexical and tone-based prompt perturbations affect the factual reliability of LLMs. Using controlled prompt variations across polite, random, and three toxicity levels, we evaluate five LLMs on ARC-Easy, GSM8K, and MMLU. We find that toxic lexical perturbations consistently reduce factual accuracy and increase uncertainty, while polite phrasing yields limited and inconsistent changes. To examine whether these answer inconsistencies correspond to internal changes, we conduct attribution-graph analyses of model activations and influences. We find that increasing toxicity selectively amplifies perturbation-sensitive variant nodes while relatively stable core reasoning nodes remain more invariant. These findings position prompt tone as a critical dimension of LLM reliability and provide behavioral and mechanistic evidence that surface-level lexical variation can alter factual outputs and internal computation.

Soorya Ram Shimgekar, Agam Goyal, Amruta Parulekar +6

University of Illinois Urbana-Champaign · Nimblemind

Prompt Sensitivity LLM Interpretability LLM Reliability Language Model Robustness