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 12, 2026cs.LG

Rational Sparse Autoencoder

Sparse autoencoders (SAEs) are standard tools for mechanistic interpretability, but current SAE families are constrained by fixed encoder nonlinearities such as ReLU, JumpReLU, and TopK. This hard-codes a particular sparsity mechanism into the model and can distort the reconstruction-versus-sparsity trade-off. We introduce the Rational Sparse Autoencoder (RSAE), which replaces the fixed encoder activation with a trainable rational function. Rational activations are flexible enough to uniformly approximate the activation primitives used by existing SAE families on compact domains (for TopK, the thresholded gate obtained after a separating top-k threshold is supplied), while also providing a richer function class for adapting to the observed pre-activation geometry. We realise this idea through a two-stage pipeline: an initialisation procedure that copies the pre-trained baseline SAE weights, plugs in rational coefficients obtained by the relaxed Remez exchange on synthetic data, and calibrates the scale parameters along with the rational coefficients; followed by a fine-tuning step under the standard sparsity-regularised reconstruction objective. Empirically, on residual-stream activations of three open-weight language models and across all three baseline activation families, the RSAE strictly improves on it after the fine-tuning step, both on reconstruction-side metrics and on downstream-behaviour metrics, without sacrificing feature-level interpretability under sparse probing. These gains are consistent across host language models, across baseline activation families, and across the full range of baseline sparsity we tested, while the upgrade itself adds only a handful of scalar parameters per autoencoder and runs in minutes on a single consumer GPU.

Naiyu Yin, Yue Yu

Department of Mathematics · Lehigh University · Bethlehem, PA 18015

Activation Sparsity LLM Interpretability Sparse Autoencoders Sparse Coding

Jun 12, 2026cs.AI

A Definition of Good Explanations and the Challenges Explaining LLM Outputs

How to define a good explanation is a long-standing philosophical debate which has found recent renewed interest in the context of AI outputs. Explainability is crucial for AI adoption in many contexts, but in order to produce good explanations of AI systems, we must first have an understanding of what good explanations are. In this paper we propose a definition inspired by the notion of counterfactual explanations, however we argue that one must also take into account the interlocutor's prior beliefs in each fact that could be offered in an explanation. We explore the ramifications of this definition for AI explainability and, in particular, why LLM outputs are difficult to produce good explanations for.

Louis Mahon, Elliot Ford, Callum Hackett

UnlikelyAI, London, UK

Explainable Artificial Intelligence Counterfactual Explanations LLM Interpretability Explanation Evaluation

Jun 12, 2026cs.LG

A Low-Rank Subspace Analysis of LLM Interventions

Interventions designed to modify a particular behavior in LLMs, such as refusal or sycophancy, often produce unintended changes in other behaviors. This lack of targeted control makes it difficult to design and implement reliable safety controls. To understand these side-effects, we introduce a diagnostic framework for analyzing interacting behaviors in LLMs. We model behaviors as low-rank subspaces in activation space, and study how interventions influence across behaviors. Across multiple instruction-tuned models (7B-70B) and across refusal, jailbreak, and sycophancy settings, we find that different behaviors share internal representations, and intervening on one behavior alters others in asymmetric ways. Some behaviors act as upstream control points whose interventions propagate broadly across other behaviors, while others remain more isolated. We relate these effects to two geometric quantities: (i) the overlap between behavior subspaces, measured as the average squared cosine of principal angles, and (ii) the angle between each behavior subspace and the decision subspace (capturing the model's final decision e.g., refuse vs. comply). Empirically, intervention effects on other behaviors tend to be larger for behavior pairs with higher subspace overlap, and for source behaviors whose subspaces lie closer (smaller angle) to the decision subspace. These findings highlight a challenge for targeted behavior control: behaviors are difficult to modify independently, as interventions can propagate through shared representations and asymmetric interactions.

Angira Sharma, Christian Schroeder de Witt, Philip Torr +2

University of Oxford

LLM Interpretability LLM Refusal Behavior Causal Interventions in Language Models Representation Geometry in Language Models

Jun 12, 2026cs.LG

When Language Representations Interact: Separability and Cross-Lingual Effects in LLMs

Large language models exhibit strong multilingual capabilities, however, their internal representations are difficult to interpret. Understanding these interactions is important for ensuring reliable behavior in multilingual systems. Recent work has shown that causal-geometric structure can explain how certain concepts are encoded as approximately linear and separable directions, but whether this framework extends to multilingual models, where language identity is correlated and hierarchical, is underexplored. We apply causal-geometric analysis to multilingual LLMs, studying 28 bilingual contrasts across three models, allowing us to analyze when languages behave as approximately independent factors and when structured dependencies persist. We find evidence that language concepts admit stable linear representations that are largely separable under a covariance-adjusted (causal) inner product, with structured deviations reflecting linguistic similarity. Moreover, languages within the same family (such as Germanic or Romance) exhibit a simplex-like geometric structure, suggesting hierarchical organization. These results extend causal-geometric interpretability to multilingual settings and provide insight into how separability and similarity may exist in multilingual LLM representations, motivating interpretability analyses that diagnose when and how structured dependencies between concepts can be anticipated. This has implications for trustworthy deployment, as residual structure between languages may lead to unintended cross-lingual effects when models are monitored or intervened upon.

Boris Marinov, Angira Sharma, Christian Schroeder de Witt +3

1UCL · University of Oxford

Multilingual Language Models LLM Interpretability Representation Geometry in Language Models

Jun 11, 2026cs.AI

Adversarial Concept Search: Predicting Compositional Errors From Feature Geometry

Humans cannot always intuit what scenarios are most challenging to LLMs. Hoping to capture challenging edge cases, developers either design problems to be difficult for humans or curate extensive benchmarks. What if we could instead anticipate which scenarios a model will fail on? In this paper, we use an LLM's representational geometry to predict which concept combinations it will fail on. We attribute this compositional failure to interference between salient features. In tasks that require systematic composition - toy programmatic settings, multihop reasoning, multilingual factual recall - we find that when a pair of concepts is encoded near-orthogonally, the model reliably composes them. When their linear encodings are close, producing interference, the model fails to compose them. Our method reliably anticipates failure modes across different compositional tasks, without evaluating specific inputs. These results lay the groundwork to use representational geometry to identify high-risk examples, construct targeted stress tests, and provide a scalable foundation for active learning in real-world deployment.

Jennifer Meng Lu, Ruochen Zhang, Isabelle Lee +3

Brown University · University of Southern California · Harvard University +1

LLM Interpretability Representation Geometry in Language Models Compositional Generalization

Jun 11, 2026cs.AI

Reasoning as Pattern Matching: Shared Mechanisms in Human and LLM Everyday Reasoning

When large language models (LLMs) fail to generalize or make content-sensitive errors in reasoning, it is often taken as evidence that LLMs are not truly reasoning, but rather performing a kind of pattern matching. The implication is that human behavior does not exhibit the same types of failures because human reasoning relies on principled and content-invariant world models. We test this assumption by first evaluating humans and LLMs on their ability to engage in common-sense reasoning about a variety of everyday situations. Our results reveal convergent patterns of reasoning across 46 LLMs and two cohorts of human participants. We then ask whether this behavioral convergence is due to LLMs having acquired content-invariant world models or a set of pattern-matching heuristics by characterizing the roles of content-invariant and content-sensitive model neurons in producing human-like responses. We find that while LLMs encode both content-invariant and content-sensitive representations, it is content-sensitive mechanisms which are causally responsible for aligning models with humans. Taken together, our results suggest that everyday causal reasoning in people and LLMs makes heavy use of pattern-matching.

Zach Studdiford, Gary Lupyan

University of Wisconsin–Madison Department of Psychology Department of Computer Science

Cognitive Modeling LLM Interpretability Causal Reasoning in Language Models LLM Reasoning

Jun 11, 2026cs.LG

Beyond the Commitment Boundary: Probing Epiphenomenal Chain-of-Thought in Large Reasoning Models

Chain-of-thought (CoT) reasoning is the dominant paradigm for inference-time scaling in language models, yet the causal influence of individual steps on the final answer remains poorly understood. In this work, we use answer logits at the end of each reasoning step to estimate each step's causal importance to the final answer and intermediate guesses, shedding light on the answer formation process of several reasoning model families. Across diverse tasks, we find that reasoning typically crosses a commitment boundary, a sharp transition from transient intermediate guesses to a stable, high-confidence answer. This transition often happens in a single step, well before the model's reasoning block ends, and is followed by epiphenomenal CoT steps that leave the final answer probability unaltered. Using attention probes, we show that answer-formation stages can be linearly decoded from the activations of intermediate reasoning steps with high accuracy, showing robust generalization to unseen reasoning tasks. We leverage this property for early-exiting reasoning blocks at the commitment boundary location, reducing the length of CoTs up to 55% with negligible impact on model performance.

Daniel Scalena, Sara Candussio, Luca Bortolussi +3

CLCG, University of Groningen · University of Milano-Bicocca · University of Trieste +1

CoT Faithfulness LLM Interpretability Efficient Language Model Reasoning Efficient Language Model Inference Large Reasoning Models Reasoning Trace Analysis

Jun 11, 2026cs.CL

Localizing Anchoring Pathways in Language Models

Irrelevant numbers in a prompt can shift language model judgments, producing anchoring effects in numerical reasoning. We study where this anchor-sensitive signal is carried inside language models using a controlled multiple-choice setup with shared answer options. We define a logit-difference metric comparing the correct answer option with the answer option corresponding to the anchor, and validate that it tracks behavioral anchoring. Using attribution-based circuit localization on 7B--8B Qwen and Llama base and instruction-tuned models, we find that edge-level methods recover this signal more faithfully than node-level methods. Low- and high-anchor circuits transfer strongly within a model, suggesting shared pathway structure across anchor direction. However, sparse transfer across base and instruction-tuned variants is less reliable, indicating that post-training changes which pathways matter most. Overall, our results provide a mechanistic account of how anchoring-related decision signals are carried inside language models.

Hillary N. Owusu, Sarah Wiegreffe, Naomi H. Feldman

University of Maryland, College Park

Numerical Reasoning in Language Models LLM Interpretability Circuits in Language Models

Jun 10, 2026cs.LG

Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal

Language-model post-training is the main stage at which model behavior is shaped, yet it still largely involves optimization of scalar rewards that summarize diverse desiderata. This abstraction gives practitioners little visibility into what their data actually teaches models, allowing spurious correlations to be learned by a model and inducing undesirable behaviors such as over-stylization and sycophancy. To address this problem, we ask: can we inspect a preference dataset before optimization and decide, at the level of concepts, which behaviors a model should be allowed to learn? Motivated by this, we introduce a data-centric post-training pipeline that uses interpretability protocols to develop statistical hypotheses for the latent concepts separating preferred from dispreferred generations, making them explicit for fine-grained user feedback. Building on this view, we unify several interpretability-based training protocols as ways of shaping rewards via feature or data interventions. Empirically, we show that our pipeline diagnoses undesirable signals in existing preference data, mitigates off-target learning, and can also help amplify or shape desired properties such as safeguards and model personality. More broadly, our results suggest that interpretability can turn post-training from optimizing opaque proxy rewards into a process of auditing and sculpting the learning signal itself.

Leon Bergen, Usha Bhalla, Sidharth Baskaran +14

Core contributors

Reward Shaping LLM Interpretability LLM Post-Training

Jun 10, 2026cs.LG

ICA Lens: Interpreting Language Models Without Training Another Dictionary

Finding interpretable directions in language-model representations is critical for understanding and controlling model behavior. Sparse autoencoders (SAEs) have become the standard tool for this purpose, but using them as the default first lens often requires training, storing, and evaluating large overcomplete dictionaries. This bottleneck limits rapid exploration and raises a fundamental question: how much interpretable structure is already visible from activation geometry before training another neural dictionary? Our intuition is simple: many interpretable directions are selective on tokens, and these directions should look less Gaussian than random directions. We therefore revisit independent component analysis (ICA), a classical method for finding non-Gaussian directions, as a compact lens for language-model interpretability. We find that ICA has been underestimated for LLM interpretability, because prior uses often relied on off-the-shelf ICA implementations that are brittle on LLM activations and lacked systematic tools for inspecting and evaluating the recovered directions. To bridge these gaps, we introduce ICALens, the first practical workflow for stable, efficient, and auditable ICA analysis of LLM representations. It combines an optimized GPU-parallel FastICA pipeline with LLM-specific stability recipes and better fitting diagnostics, enabling efficient and reliable layer-wise analysis. Across GPT-2 Small, Gemma 2 2B, and Qwen 3.5 2B Base, ICALens efficiently recovers compact, human-interpretable directions without per-layer gradient-based dictionary training. On SAEBench, ICA is competitive with public SAEs in sparse probing and outperforms them in targeted probe perturbation under small-to-medium budgets. These results suggest that ICA should not be viewed as a weak baseline, but as an efficient and complementary first lens for exploring language-model representations.

Sida Liu, Feijiang Han

Independent Researcher · University of Maryland

Independent Component Analysis LLM Interpretability Language Model Introspection Representation Geometry in Language Models

Jun 10, 2026cs.LG

Bergson: An Open Source Library for Data Attribution

Data attribution is a promising field in interpretability that aims to explain model behavior through the influence of its training data, with applications including debugging undesirable model behavior and training dataset curation. However, significant engineering effort is required to perform it at scale, and many cutting edge techniques lack open-source tooling and support. Bergson is an open source library that aims to enable faster progress in the field by providing a host of techniques that scale to very large language models and pre-training datasets. The library natively supports on-disk gradient stores and multi-node distributed training, and provides quality of life tools for researchers. Finally, we introduce the first open-source implementations of three leading data attribution methods: MAGIC, SOURCE, and TrackStar. The library is available at https://github.com/EleutherAI/bergson .

Lucia Quirke, Louis Jaburi, David Johnston +6

EleutherAI · Independent

Gradient-Based Attribution LLM Interpretability Training Data Attribution

Jun 9, 2026cs.AI

Forecasting Future Behavior as a Learning Task

Trust in an AI system is often anchored by explanations of how it works, which one then uses to forecast its behavior on new inputs. For large reasoning models (LRMs), this conventional route is particularly difficult to follow: explanation methods for single token generations do not naturally generalize to long trajectories, and the trajectories themselves are often not faithful when read as natural language. We propose an alternative that bypasses the explanation step: treat behavior forecasting as a learnable task and train Behavior Forecasters that operates on a single reasoning trajectory to make the same forecasts one would typically seek from an explanation. The forecaster's training data is obtained by querying the LRM with no human annotation, and its inference is done in a single forward pass. We instantiate this approach on two tasks: how likely the LRM is to repeat its answer on re-runs, and how removing parts of the input changes its answer. We evaluate this approach on both tasks across three diverse reasoning datasets and find that trained Behavior Forecasters are more accurate than GPT-5.4 and Claude Opus-4.6 reading the same trajectories as naive readers, at a small fraction of their inference cost. We find that fine-tuning the backbone end-to-end and initializing it from the target LRM are each necessary for strong performance. These results show that the reasoning trajectory carries information about the LRM's future behavior that goes beyond what naive reading conveys.

Mosh Levy, Yoav Goldberg, Asa Cooper Stickland

Bar-Ilan University · Constellation · UK AI Security Institute +1

LLM Interpretability Large Reasoning Models

Jun 9, 2026cs.IR

How Do LLMs Cite? A Mechanistic Interpretation of Attribution in Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG) aims to enhance the trustworthiness of Large Language Models (LLMs) by grounding their outputs in external documents, often using inline citations for verifiability. However, the faithfulness of these citations -- whether the model genuinely uses a source to generate an answer -- remains a critical, unverified assumption. This paper offers the first mechanistic account of how a large language model decides whether to attach an inline citation while answering a factoid question. Using the Llama-3.1-8B-Instruct model in a controlled experimental environment based on the PopQA dataset, we employ an activation patching approach. We map the underlying mechanism responsible for citation, discovering that it is not a single, localized component but a distributed, multi-stage "attributional ensemble" of attention heads and MLP layers. We show that amplifying or attenuating only those critical heads and MLPs repairs over 90% of missed citations and eliminates 69% of spurious ones on PopQA without harming answer accuracy. Although gains on the multi-document HotpotQA benchmark are modest, the same component set still moves citation rates in the intended direction, indicating that the underlying mechanism is not dataset-specific. The results reveal a potential disconnect between the model's apparent reasoning and its internal computational pathway, suggesting that inline citations can create a false sense of security.

Ian van Dort, Maria Heuss

University of Amsterdam

Retrieval-Augmented Generation LLM Interpretability RAG Evaluation

Jun 9, 2026cs.CL

MIRAGE: A Polarity-Flipping Encoding Subspace in LLM Agents

When LLM agents are coerced into covertly encoding sensitive data (Base64, ROT13, acrostic, synonym chains, and beyond), the resulting outputs evade output-side detection but the underlying computation does not. Across nine encoding families and eight models from five architecture families, that computation is supported by a shared low-dimensional encoding subspace in the residual stream. A logistic-regression probe trained on eight encoding families recovers the held-out ninth at AUC 0.975-1.000, reading the computation rather than surface features. The same direction exhibits a second mechanistic signature at the planning token, flipping polarity to activate positively when the model will simulate the encoding inline and negatively when it will outsource it to a tool call, distinguishing two execution strategies before the encoded text exists. We build MIRAGE (Model-Internal Readout of Agentic Generation Exfiltration), a two-channel real-time monitor exploiting both signals. On 126 agentic exfiltration scenarios, it reaches AUC = 0.918, substantially outperforming output-only detection (AUC = 0.518). Monitor performance is fundamentally a property of the host model's geometry: benign-encoding false-positive rate ranges from 0% on Qwen-7B to 100% on Phi-3.5, revealing that the probe faithfully reads whether a model's geometry separates covert from overt encoding. Across all tested adversarial budgets, every attack suppressing the subspace also destroyed encoding fidelity, reported as an empirical regularity on the evaluated budgets, not a structural impossibility claim.

Pratibha Revankar, Kargi Chauhan, Jihye Kim +3

University of California, Santa Cruz.

LLM Interpretability AI Agent Security AI Agent Monitoring LLM Security Language Model Probing

Jun 8, 2026cs.LG

Interpreting and Steering a Text-to-Speech Language Model with Sparse Autoencoders

Language models increasingly serve as the backbone of text-to-speech (TTS) systems, yet we understand little about the representations they build when text and generated speech tokens share a single residual stream. We train BatchTopK sparse autoencoders on the LM backbone of CosyVoice3 and introduce a modality-aware auto-interp pipeline that labels each feature from where it fires-text-prefix context, 1-second speech clips, or both. The recovered features are interpretable, spanning phonemes, laughter, accent prompts and speaker gender. Steering through the SAE latent space shows these features are causal rather than merely descriptive: targeted interventions raise laughter probability from 0.02 to 0.79, flip perceived speaker gender, and control speech rate while preserving spoken content. SAE features thus serve both as interpretability objects and as control directions for TTS synthesis.

Nikita Koriagin, Georgii Aparin, Nikita Balagansky +1

♣T-Tech · ♦AI Foundation and Algorithm Lab

TTS Synthesis Language Model Steering LLM Interpretability Sparse Autoencoders Controllable Speech Generation Speech Language Models

Jun 8, 2026cs.CL

The Neutral Mask: How Alignment Training Provides Shallow Alignment while Leaving Partisan Structure Intact in a Large Language Model

The ambition behind alignment training is to make large language models safe and useful. The primary mechanisms, reinforcement learning from human feedback (RLHF) and its direct-optimization variants, shape the behavior of deployed language models by aligning them with ``human values.'' Yet the process is opaque. What values are being encoded; whose values are they; and how does alignment training encode them? A growing body of evidence suggests that these methods produce only functional compliance rather than deep alignment. We offer a mechanistic case study of this phenomenon for partisan political orientation with a comparison of the internal representations of Llama 3.1 8B before and after alignment training. We show that alignment training does not remove the structured partisan direction in the base model. Instead, it compresses the variance of the partisan signal to generate consistently balanced and non-partisan output. Sparse autoencoder decomposition reveals that policy-encoding features, which activate sporadically in the base model, are completely inactive in the Instruct model. Feature-level steering experiments confirm the causal disconnect. Alignment training thus encodes a norm of political neutrality, not by erasing the model's knowledge of partisanship, but by severing the causal pathway from partisan geometry to output generation. Importantly, this neutrality is functional, not structural so that the underlying geometry that enables partisan steering remains intact. The mechanisms that bypass RLHF's guardrails, such as inferring and amplifying a user's partisan identity, reactivate partisan generation. If alignment training operates by disconnecting rather than removing value-laden structure, then the same pattern may hold for other value domains, and the aligned model's behavior may be more fragile than its outputs suggest.

Wendy K. Tam

Professor and Stevenson Chair in the Departments of Political Science, Computer Science, Biomedical Informatics, and the Law School at Vanderbilt University, and an affiliate of the National Center for Supercomputing Applications at the University of Illinois at Urbana-Champaign

LLM Alignment LLM Interpretability Political Bias in Language Models

Jun 8, 2026cs.AI

PRISM: Recovering Instruction Sets from Language Model Activations

As LLMs are deployed as agents, reliable monitoring requires knowing not only what they output, but which instructions are steering their behavior. This is difficult when models infer unintended subgoals, follow contextual cues, or are influenced by prompt injections and hidden objectives. While activation-to-language methods suggest that hidden states can reveal natural-language information, existing approaches are not designed to recover the full set of simultaneous instructions, constraints, prohibitions, and subgoals present in agentic settings. As a first step toward this monitoring goal, we formalize instruction set retrieval and introduce PRISM, an activation-conditioned interpreter that decodes hidden states from a frozen target model into a faithful bullet list of instructions. Unlike prior activation-to-language methods, PRISM is trained to recover instruction sets directly, using judge-guided GRPO to reward covered instructions and penalize unsupported ones. Across benign, constrained, prompt-injection, and hidden-objective settings, PRISM outperforms activation-to-language baselines, especially on security-relevant objectives.

Gilad Gressel, Rahul Pankajakshan, Julia Diament +3

Center for Cybersecurity Systems & Networks, Amrita Vishwa Vidyapeetham · Ben-Gurion University of the Negev · Microsoft

Neural Decoding LLM Interpretability Language Model Probing

Jun 7, 2026cs.CL

The Amplifying Mirror: Locating and Steering the Partisan Direction inside a Large Language Model

Large language models are rapicly replacing search engines as the primary interface between people and information. Unlike search engines, which retrieve existing content, LLMs generate novel text shaped by internal representations learned during training. Here we show that partisan political identity is encoded in the model's activation space, and that this direction directly shapes generation. Using 190,491 tweets from sitting members of the U.S. Congress as labeled training data, we train linear probes on the hidden states of the Llama 3.1 8B Instruct model. We identify a single geometric axis at layer 18 that separates Republican from Democratic text with an AUC of 0.945 and a Cohen's d of 1.94, and use sparse autoencoders to decompose that axis into interpretable partisan features. Causally intervening along this axis, ablating or amplifying the partisan component mid-generation, produces systematic shifts in the model's output. We witness stance reversals, register shifting, and structured fabrications of authority. Our results demonstrate that partisan bias in language models is not a vague emergent property but a learned geometric feature that can be precisely located and steered. Partisan bias is not a bug to be patched, but a structural property of how these models encode information about their users. As LLMs displace search engines as the interface to knowledge, understanding that product design (and its consequences) will be essential for navigating the legal, social, and political transitions from an information ecosystem that is curated to one that is generated.

Wendy K. Tam

Professor and Stevenson Chair in the Departments of Political Science, Computer Science, Biomedical Informatics, and the Law School at Vanderbilt University

LLM Interpretability Political Bias in Language Models Linguistic Bias in Language Models Causal Interventions in Language Models Activation Steering

Jun 7, 2026cs.CL

Analyzing the Correlation Between Hallucinations and Knowledge Conflicts in Large Language Models

Hallucinations -- factually incorrect or unverifiable outputs -- remain one of the most challenging limitations of Large Language Models (LLMs), especially in knowledge-intensive tasks. One proposed explanation is internal knowledge conflicts arising from fixed, outdated training data. This paper investigates whether internal representations linked to knowledge conflicts correlate with hallucination behaviors in LLMs. Using probing techniques inspired by two prior works, we analyzed activations from hidden, attention, and MLP layers, as well as output logits, across predefined tasks. We probed LLaMA-3-8B on hallucination detection benchmarks and Falcon-7B on a knowledge conflict dataset. Our findings show that, although conceptually related, hallucination activation patterns cannot be fully reduced to or explained by knowledge conflict representations. Nonetheless, probing proves a robust tool across multiple languages and activation types, supporting its role in improving LLM interpretability. This work advances the broader understanding of hallucinations in LLMs and underscores the value of fine-grained analysis of their internal behavior.

Lucrezia Laraspata, Giovanna Castellano, Gennaro Vessio

Department of Computer Science, University of Bari Aldo Moro, Bari, Italy

Knowledge Conflicts in Language Models LLM Interpretability Hallucination in Language Models

Jun 7, 2026cs.CL

Inside the LLM Word Factory

Transformer language models process input provided as subword fragments, but natural language semantics usually rely on word-level concepts. Detokenization is the process where models reconcile these two facts, aggregating subwords into word-level representations through their computation. Prior work has found that this takes place mostly in early-to-middle layers, but so far the exact mechanics of the process have not been pinned down. We venture deep into detokenization using activation patching in controlled paired experiments that isolate the contribution of different model components, localizing English detokenization in Llama2-7B to a two-stage process at Layer 1. Attention transmits a token-specific signal from nonfinal subwords, using sequential relays if necessary, while the MLP composes it with the local embedding. This two-stage structure generalizes to twelve models from eight families, but the depth over which it takes place depends on the flavor of positional encoding: RoPE-based models detokenize over 1 to 5 layers, while learned-absolute models take 5 to 10. Finally, we provide a probe for determining the success of the detokenization process based on early-layer activations alone, performing at 0.94-0.97 AUROC depending on the amount of context.

Benzi Busigin, Yuval Pinter

Stein Faculty of Computer and Information Science Ben-Gurion University of the Negev Beer Sheva, Israel

Language Model Decoding LLM Interpretability Subword Tokenization

Jun 7, 2026cs.AI

Explaining Black-Box Language Models: Learning to Optimize Linguistically-Structured Word Subsets

As deep language models (DLMs) are increasingly deployed in high-stakes domains such as healthcare, understanding their decision rationale becomes paramount for ensuring trust, safety, and accountability. However, achieving this vital level of interpretability is particularly challenging when these DLMs operate as black-box systems (e.g., via APIs), where access to internal model states (e.g., parameters, gradients) is restricted. Despite numerous efforts, existing explanation methods often fail to concurrently satisfy three key desiderata: (i) inference-time efficiency, (ii) black-box compatibility without inducing out-of-distribution behavior, and (iii) comprehensible explanations grounded in the input's linguistic structure. To address these challenges, we propose a method that explains predictions of DLMs by selecting a small, informative subset of input words. We formulate this as an amortized optimization problem, enabling efficient one-shot inference without the need for input-specific search. Our selection policy is trained via REINFORCE-style policy gradients, allowing discrete word selection in a fully gradient-free setting. To enhance interpretability and align with human linguistic intuition, we integrate graph-structured knowledge into this selection process, fostering linguistically coherent subsets that result in explanations both highly informative and cognitively meaningful to end-users. We evaluated our method on diverse DLM architectures and multiple real-world datasets. It consistently identifies word subsets with enhanced discriminative power and stronger alignment with linguistically salient cues, outperforming both conventional black-box compatible methods and gradient-based approaches that are given oracle access to the black-box model's gradients for a more challenging benchmark. Our code is available at here.

Minyoung Hwang, Seokhyun Lee, Changhee Lee

Korea University Republic of Korea

Feature Attribution LLM Interpretability

Jun 6, 2026cs.LG

Pre-Intervention Prediction of Sparse Autoencoder Steering Side Effects

Sparse autoencoder (SAE) features are increasingly used to steer language models, but feature steering is rarely clean: the same intervention can behave inconsistently across contexts and perturb unrelated features. We introduce a pre-intervention screening framework for forecasting SAE steering side effects from feature statistics computed before steering. We operationalize side effects along two axes of steering modularity, effect stability and collateral spread, and evaluate GPT-2-small, Pythia-70M-deduped, Gemma-2-2B, and Llama-3.1-8B across ReLU, JumpReLU, and TopK SAE dictionaries. Across these settings, decoder geometry, activation statistics, co-activation structure, and direct-logit footprint predict steering modularity better than frequency-only and activation-magnitude baselines. The signal is strongest in GPT-2-small, Pythia-70M, and Llama-3.1-8B, where it survives residualization against magnitude-related confounds, and weaker in Gemma-2-2B. Held-out screening shows that ranking unseen features by predicted cleanliness can select features that steer more cleanly on fresh contexts, but the successful axis varies by setting: GPT-2 improves most cleanly, Pythia improves mainly on stability, Llama mainly on collateral, and Gemma only partially. A controlled Llama Scope width comparison shows that the predictive signal persists under a 32K-to-128K dictionary-width change, although the screening payoff becomes less stable. Overall, SAE steering side effects are predictable in advance, but the useful predictor signature and transferred modularity axis are model- and dictionary-setting dependent.

Evan Duan

Language Model Steering LLM Interpretability Sparse Autoencoders Causal Interventions in Language Models

Jun 6, 2026cs.AI

Cross-LLM Consistency in Inference: Evidence from Shared Interactions

Large language models (LLMs) differ in architecture, training data, and optimization procedures, yet they may still develop similar internal inference patterns. In this paper, we examine this hypothesis using interaction-based explanations. We find that LLMs often share interaction patterns when predicting the same target token from the same prompt. This consistency is more pronounced among advanced LLMs. Shared interactions also tend to be lower-order and show weaker positive-negative cancellation than non-shared interactions. These results suggest that advanced LLMs may be implicitly optimized toward common inference patterns, even though the mechanisms that give rise to such cross-model consistency remain open.

Siyu Lou, Yao Yan, Yuntian Chen +1

School of Computer Science Shanghai Jiao Tong University Ningbo Key Laboratory of Advanced Manufacturing Simulation Eastern Institute of Technology, Ningbo · College of Computer and Information Science Chongqing Normal University · Ningbo Key Laboratory of Advanced Manufacturing Simulation Eastern Institute of Technology, Ningbo

LLM Inference LLM Interpretability Cross-Lingual Consistency in Language Models

Jun 6, 2026cs.CL

MechLens: Late Crystallization of Factual Knowledge Explains Intervention Effectiveness in Language Models

Understanding where LLMs store factual knowledge is critical for hallucination mitigation. We systematically quantify Late Crystallization: factual knowledge does not gradually emerge across layers but "crystallizes" abruptly at the final layers. Across five model families (Pythia, Gemma, Qwen2.5, Llama-3.1, Mistral; 0.5--14B), 26.8%--93.4% of correct answers never enter top-10 predictions at any intermediate layer, with late emergence (>80% depth) consistent across architectures. Cross-scale (Qwen2.5-14B) and cross-benchmark (MMLU: 98.2%) results confirm generality; tuned lens rules out probe artifacts. A sentiment-classification control (0.5% for Qwen vs. 85.9% factual; 2.0% for Mistral vs. 26.8%) confirms the phenomenon is specific to factual recall. Late Crystallization yields a crystallization-guided intervention principle: CAA outperforms DoLa on moderate-crystallization models (Llama, Mistral; p<0.001), with a directionally consistent reversal on high-crystallization Qwen (+25.4% vs. +15.5% MC1, p=0.069). LayerNorm ablation shows crystallization is intrinsic to the residual stream; LN scaling (x1.2) yields +11.8% MC1 with zero inference overhead. We further reveal a Computability-Memorization Spectrum: computable knowledge crystallizes earlier (layer 22.1/28) than memorized facts (28.0/28). We release MechLens supporting five model families.

Xueping Gao

Alibaba Cloud

Memorization in Language Models LLM Interpretability Knowledge Representation Factual Knowledge in Language Models Causal Interventions in Language Models

Jun 5, 2026cs.AI

Beyond Post-hoc Explanation: Toward Glassbox AI via Probabilistic Mediation

Large language models are rapidly becoming infrastructural components in high-stakes institutional settings, including public administration, legal reasoning, and healthcare, where opacity is not merely inconvenient but institutionally and legally untenable. Existing approaches to explainability are predominantly post-hoc, offering unstable, non-contestable accounts that have no formal relationship to the reasoning process that produced the output. We argue that the problem is not the absence of explanation but the absence of structured reasoning in the first place. This paper makes the case for a fundamentally different architecture, which we call the Glassbox Framework, in which Bayesian networks serve as transparent, ante-hoc mediation layers for generative models. Bayesian networks encode domain knowledge, causal assumptions, and probabilistic dependencies before inference occurs, enabling auditable reasoning traces, uncertainty quantification, and contestable outputs. We characterise the architecture of this framework and ground it in a benefit eligibility scenario, identifying the foundational challenges spanning semantic alignment, dynamic model construction, probabilistic grounding, and human governance that must be solved to realise it at scale. By shifting from post-hoc explanation to ante-hoc probabilistic mediation, this work outlines a principled path toward AI systems that are not only powerful but fundamentally accountable.

Manuele Leonelli

School of Science and Technology, IE University, Madrid, Spain

Explainable Artificial Intelligence AI Accountability Bayesian Networks LLM Interpretability

Jun 5, 2026cs.LG

When Attribution Patching Lies: Diagnosis and a Second-Order Correction

A central goal of mechanistic interpretability is to identify which internal components causally drive a language model's behavior. Because these importance estimates serve as the evidence for identifying circuits, systematic errors can lead to the misidentification of the underlying mechanisms. While activation patching provides a gold-standard causal metric, its computational cost is prohibitive at scale. Practitioners instead rely on attribution patching, a gradient-based, first-order approximation whose reliability remains poorly understood. In this work, we characterize the source of this unreliability, demonstrating that the dominant error stems from the non-linearities in the downstream network rather than local curvature at the patched component. This insight yields three practical tools: (i) a reliability score to detect untrustworthy estimates, (ii) error bounds quantifying potential attribution mis-specifications, and (iii) a Hessian-vector-product (HVP) correction that eliminates the leading-order error with only one additional backward pass. In evaluations across five model families (124M-9B parameters) and both random-token and naturalistic (name-swap) perturbations, HVP is the only second-order correction feasible at larger scale, where standard baselines like Integrated Gradients become computationally prohibitive. In comparative experiments, a multi-step HVP variant matches or exceeds the accuracy of Integrated Gradients at significantly lower compute, outperforming prior second-order baselines. These improvements lead to higher-fidelity circuit recovery on standard benchmarks and support a Screen-Flag-Fix workflow that targets computational effort only toward the components flagged as unreliable.

Luyang Zhang, Jialu Wang

1Carnegie Mellon University · University of California, Santa Cruz

Gradient-Based Attribution LLM Interpretability Causal Attribution Mechanistic Interpretability Circuits in Language Models Activation Patching

Jun 5, 2026cs.CL

Interpreting Brain Responses to Language with Sparse Features from Language Models

A central goal of cognitive neuroscience is to characterize the features that are represented by human language cortex. Artificial language models (LMs) have emerged as a powerful tool to address this challenge, but studies relating biological and artificial representations are often criticized as relating one black box to another. The present work introduces Augmented Sparse Encoding Models, an encoding framework that replaces dense LM hidden states with hierarchically-organized sparse autoencoder (SAE) features, while explicitly including surprisal as a predictor. Using this approach, we (i) produce interpretations of neural responses and (ii) test whether model-brain alignment reflects primary or idiosyncratic variation in LM representations. Using a high-field 7T fMRI dataset of eight participants listening to 200 linguistically diverse sentences, we first validate our modeling framework by recovering previous interpretations of voxel populations tuned to processing difficulty and meaning abstractness. We then interpret a previously-uncharacterized (but reliable) voxel population and find that it is tuned to people-related content. Next, we show that the fronto-temporal human language network is predicted by a common set of features across its constituent regions, but find that frontal regions are relatively well-explained by surprisal alone, even in the absence of LM-based features. Finally, we show that brain responses during language processing are not merely predictable from an arbitrary set of LM features. Rather, brain responses are best explained by the features that tend to capture the most general information encoded in LM representations, suggesting a nontrivial correspondence between brain and LM language representation.

Michael A. Lepori, Kendrick Kay, Greta Tuckute

Dept. of Computer Science Brown University · Dept. of Radiology University of Minnesota · Kempner Institute Harvard University

LLM Interpretability Sparse Autoencoders Neural Encoding

Jun 4, 2026cs.AI

LLM Self-Recognition: Steering and Retrieving Activation Signatures

Recent advances in interpretability suggest that large language models (LLMs) implicitly encode signals in their generated text that enable self-recognition of their outputs. We demonstrate that this capability is reliable, even in low-entropy scenarios, and that it can be amplified through targeted intervention. By steering the internal residual stream during generation with a random sparse vector, we create a detectable fingerprint that enables attribution of a given text to a specific LLM. This signal is recoverable from the activations of an LLM used as a detector, achieving over 98% accuracy across multiple detection settings while preserving the quality of generated text. As AI-generated content proliferates, this approach offers a practical alternative to traditional detectors by leveraging the model's natural representation structure for attribution rather than embedding a signal externally. Our contributions include: (i) establishing reliable self-recognition capabilities in LLMs, (ii) a simple steering mechanism enabling multi-LLM identification with no quality degradation, (iii) demonstrating that activation spaces contain exploitable structure for encoding signals without semantic interference.

Thibaud Ardoin, Jonas Schäfer, Gerhard Wunder

Department of Computer Science, Freie Universitaet Berlin, Germany.

AI-Generated Text Detection LLM Interpretability Language Model Fingerprinting

Jun 4, 2026cs.CL

The Tell-Tale Norm: $\ell_2$ Magnitude as a Signal for Reasoning Dynamics in Large Language Models

Recent work has sought to understand Large Language Models (LLMs) reasoning, yet a principled, model-intrinsic signal that captures its layer-wise reasoning dynamics remains underexplored. We bridge this gap by demonstrating that the l2 norm of hidden states serves as an endogenous signal of the model's reasoning intensity. Using Sparse Autoencoders (SAEs) as a diagnostic probe, we observe that LLMs' internal reasoning is marked by a sharp increase in reasoning feature activations concentrated in late layers. Motivated by this pattern, we establish a formal link between reasoning intensity and the model's latent geometry and theoretically prove that the l2 norm of hidden states bounds the activation strength of SAE reasoning features. Empirical correlation analysis and causal interventions further validate the l2 norm as a faithful indicator, where heightened norms consistently correspond to critical reasoning steps. We then introduce three test-time scaling techniques guided by l2 norms: (i) Adaptive Layer-wise Reasoning Recursion, (ii) Endogenous Reasoning State Steering, and (iii) l2-guided Response Selection, which requires no additional training or data and is compatible with advanced inference engines. Experiments across model architectures and benchmarks show that l2-norm-based techniques significantly improve reasoning performance, offering a principled yet simple lens to perceive and control LLM latent reasoning dynamics. Our code is available at https://github.com/zjy1298/The-Tell-Tale-Norm.

Jinyang Zhang, Hongxin Ding, Yue Fang +4

School of Computer Science, Peking University · National Engineering Research Center of Software Engineering, Peking University · Key Laboratory of High Confidence Software Technologies, Ministry of Education +2

LLM Interpretability Test-Time Scaling LLM Reasoning

Jun 4, 2026cs.LG

LLM Explainability with Counterfactual Chains and Causal Graphs

Causal graphs provide a high-level language for making mechanisms transparent. Recent work uses Large Language Models (LLMs) to recover causal graphs of external-world processes. Instead, in this paper, we use causal graphs to model LLM inference itself, providing stakeholders with a transparent view of how the model perceives and organizes high-level concepts to produce a prediction. We propose a four-phase method for constructing such graphs. Given a target LLM and a set of textual examples, our method discovers class-discriminative, human-interpretable concepts and maps each input to LLM-perceived concept states. We then introduce an MCMC-inspired counterfactual augmentation procedure that expands the sparse observational data through chains of counterfactuals. This enables stable causal discovery with

σ

-CG, yielding informative, interpretable graphs. We apply our method to three LLMs across disease diagnosis, sentiment analysis, and LLM-as-a-judge classification tasks. We evaluate the learned graphs for predictive fidelity and structural stability, and the MCMC-inspired augmentation for convergence and downstream utility. Our results show that the discovered causal graphs capture meaningful dependencies consistent with LLMs' reasoning. Together, this paper provides a foundation for concept-level explainability of LLMs.

Nirit Nussbaum-Hoffer, Nitay Calderon, Liat Ein-Dor +1

Faculty of Data and Decision Sciences, Technion · IBM Research

LLM Reasoning with Graphs LLM Interpretability Causal Structure Learning Concept-Based Explanations Causal Reasoning in Language Models Counterfactual Data Augmentation+1