I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders

• URL: http://arxiv.org/abs/2503.18878v2
• Date: Tue, 05 Aug 2025 20:14:47 GMT
• Title: I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders
• Authors: Andrey Galichin, Alexey Dontsov, Polina Druzhinina, Anton Razzhigaev, Oleg Y. Rogov, Elena Tutubalina, Ivan Oseledets,
• Abstract summary: Internal mechanisms behind reasoning in LLMs remain unexplored.<n>We employ Sparse Autoencoders to test our hypothesis.<n>Our work provides the first step towards a mechanistic understanding of reasoning in LLMs.
• Score: 8.1201445044499
• License: http://creativecommons.org/licenses/by-nc-sa/4.0/
• Abstract: Recent LLMs like DeepSeek-R1 have demonstrated state-of-the-art performance by integrating deep thinking and complex reasoning during generation. However, the internal mechanisms behind these reasoning processes remain unexplored. We observe reasoning LLMs consistently use vocabulary associated with human reasoning processes. We hypothesize these words correspond to specific reasoning moments within the models' internal mechanisms. To test this hypothesis, we employ Sparse Autoencoders (SAEs), a technique for sparse decomposition of neural network activations into human-interpretable features. We introduce ReasonScore, an automatic metric to identify active SAE features during these reasoning moments. We perform manual and automatic interpretation of the features detected by our metric, and find those with activation patterns matching uncertainty, exploratory thinking, and reflection. Through steering experiments, we demonstrate that amplifying these features increases performance on reasoning-intensive benchmarks (+2.2%) while producing longer reasoning traces (+20.5%). Using the model diffing technique, we provide evidence that these features are present only in models with reasoning capabilities. Our work provides the first step towards a mechanistic understanding of reasoning in LLMs. Code available at https://github.com/AIRI-Institute/SAE-Reasoning

Related papers

• Eliciting Reasoning in Language Models with Cognitive Tools [9.68459632251626]
  We build on the long-standing literature in cognitive psychology and cognitive architectures.<n>We endow an LLM with a small set of "cognitive tools" encapsulating specific reasoning operations.<n>Surprisingly, this simple strategy results in considerable gains in performance on standard mathematical reasoning benchmarks.
  arXiv  Detail & Related papers  (2025-06-13T13:56:52Z)
• Decomposing MLP Activations into Interpretable Features via Semi-Nonnegative Matrix Factorization [17.101290138120564]
  Current methods rely on dictionary learning with sparse autoencoders (SAEs)<n>Here, we tackle these limitations by directly decomposing activations with semi-nonnegative matrix factorization (SNMF)<n>Experiments on Llama 3.1, Gemma 2 and GPT-2 show that SNMF derived features outperform SAEs and a strong supervised baseline (difference-in-means) on causal steering.
  arXiv  Detail & Related papers  (2025-06-12T17:33:29Z)
• The Reasoning-Memorization Interplay in Language Models Is Mediated by a Single Direction [34.86855316803838]
  We identify a set of linear features in the model's residual stream that govern the balance between genuine reasoning and memory recall.<n>We show that intervening in these reasoning features helps the model more accurately activate the most relevant problem-solving capabilities during answer generation.
  arXiv  Detail & Related papers  (2025-03-29T14:00:44Z)
• OpenVLThinker: An Early Exploration to Complex Vision-Language Reasoning via Iterative Self-Improvement [91.88062410741833]
  This study investigates whether similar reasoning capabilities can be successfully integrated into large vision-language models (LVLMs)<n>We consider an approach that iteratively leverages supervised fine-tuning (SFT) on lightweight training data and Reinforcement Learning (RL) to further improve model generalization.<n>OpenVLThinker, a LVLM exhibiting consistently improved reasoning performance on challenging benchmarks such as MathVista, MathVerse, and MathVision, demonstrates the potential of our strategy for robust vision-language reasoning.
  arXiv  Detail & Related papers  (2025-03-21T17:52:43Z)
• Computation Mechanism Behind LLM Position Generalization [59.013857707250814]
  Large language models (LLMs) exhibit flexibility in handling textual positions.<n>They can understand texts with position perturbations and generalize to longer texts.<n>This work connects the linguistic phenomenon with LLMs' computational mechanisms.
  arXiv  Detail & Related papers  (2025-03-17T15:47:37Z)
• Interpreting and Steering LLMs with Mutual Information-based Explanations on Sparse Autoencoders [29.356200147371275]
  Large language models (LLMs) excel at handling human queries, but they can occasionally generate flawed or unexpected responses.<n>We propose using a fixed vocabulary set for feature interpretations and designing a mutual information-based objective.<n>We propose two runtime steering strategies that adjust the learned feature activations based on their corresponding explanations.
  arXiv  Detail & Related papers  (2025-02-21T16:36:42Z)
• AlphaMaze: Enhancing Large Language Models' Spatial Intelligence via GRPO [0.0]
  Large Language Models (LLMs) have demonstrated impressive capabilities in language processing, yet they often struggle with tasks requiring visual spatial reasoning. We introduce a novel two-stage training framework designed to equip standard LLMs with visual reasoning abilities for maze navigation.
  arXiv  Detail & Related papers  (2025-02-20T16:05:18Z)
• Unveiling the Magic of Code Reasoning through Hypothesis Decomposition and Amendment [54.62926010621013]
  We introduce a novel task, code reasoning, to provide a new perspective for the reasoning abilities of large language models.<n>We summarize three meta-benchmarks based on established forms of logical reasoning, and instantiate these into eight specific benchmark tasks.<n>We present a new pathway exploration pipeline inspired by human intricate problem-solving methods.
  arXiv  Detail & Related papers  (2025-02-17T10:39:58Z)
• Satori: Reinforcement Learning with Chain-of-Action-Thought Enhances LLM Reasoning via Autoregressive Search [57.28671084993782]
  Large language models (LLMs) have demonstrated remarkable reasoning capabilities across diverse domains.<n>Recent studies have shown that increasing test-time computation enhances LLMs' reasoning capabilities.<n>We propose a two-stage training paradigm: 1) a small-scale format tuning stage to internalize the COAT reasoning format and 2) a large-scale self-improvement stage leveraging reinforcement learning.
  arXiv  Detail & Related papers  (2025-02-04T17:26:58Z)
• Retrieval-Augmented Semantic Parsing: Using Large Language Models to Improve Generalization [6.948555996661213]
  We introduce Retrieval-Augmented Semantic Parsing (RASP), a simple yet effective approach that integrates external lexical knowledge into the parsing process.<n>Our experiments show that LLMs outperform previous encoder-decoder baselines for semantic parsing.
  arXiv  Detail & Related papers  (2024-12-13T15:30:20Z)
• Automatically Interpreting Millions of Features in Large Language Models [1.8035046415192353]
  sparse autoencoders (SAEs) can be used to transform activations into a higher-dimensional latent space.<n>We build an open-source pipeline to generate and evaluate natural language explanations for SAE features.<n>Our large-scale analysis confirms that SAE latents are indeed much more interpretable than neurons.
  arXiv  Detail & Related papers  (2024-10-17T17:56:01Z)
• Thought-Like-Pro: Enhancing Reasoning of Large Language Models through Self-Driven Prolog-based Chain-of-Thought [31.964412924094656]
  Large language models (LLMs) have shown exceptional performance as general-purpose assistants. We introduce a novel learning framework, THOUGHT-LIKE-PRO, to facilitate learning and generalization across diverse reasoning tasks. Our empirical findings indicate that our proposed approach substantially enhances the reasoning abilities of LLMs.
  arXiv  Detail & Related papers  (2024-07-18T18:52:10Z)
• CLOMO: Counterfactual Logical Modification with Large Language Models [109.60793869938534]
  We introduce a novel task, Counterfactual Logical Modification (CLOMO), and a high-quality human-annotated benchmark. In this task, LLMs must adeptly alter a given argumentative text to uphold a predetermined logical relationship. We propose an innovative evaluation metric, the Self-Evaluation Score (SES), to directly evaluate the natural language output of LLMs.
  arXiv  Detail & Related papers  (2023-11-29T08:29:54Z)
• A Closer Look at the Self-Verification Abilities of Large Language Models in Logical Reasoning [73.77088902676306]
  We take a closer look at the self-verification abilities of large language models (LLMs) in the context of logical reasoning. Our main findings suggest that existing LLMs could struggle to identify fallacious reasoning steps accurately and may fall short of guaranteeing the validity of self-verification methods.
  arXiv  Detail & Related papers  (2023-11-14T07:13:10Z)
• Towards LogiGLUE: A Brief Survey and A Benchmark for Analyzing Logical Reasoning Capabilities of Language Models [56.34029644009297]
  Large language models (LLMs) have demonstrated the ability to overcome various limitations of formal Knowledge Representation (KR) systems. LLMs excel most in abductive reasoning, followed by deductive reasoning, while they are least effective at inductive reasoning. We study single-task training, multi-task training, and "chain-of-thought" knowledge distillation fine-tuning technique to assess the performance of model.
  arXiv  Detail & Related papers  (2023-10-02T01:00:50Z)
• Evaluating and Explaining Large Language Models for Code Using Syntactic Structures [74.93762031957883]
  This paper introduces ASTxplainer, an explainability method specific to Large Language Models for code. At its core, ASTxplainer provides an automated method for aligning token predictions with AST nodes. We perform an empirical evaluation on 12 popular LLMs for code using a curated dataset of the most popular GitHub projects.
  arXiv  Detail & Related papers  (2023-08-07T18:50:57Z)
• Large Language Models Are Not Strong Abstract Reasoners [12.354660792999269]
  Large Language Models have shown tremendous performance on a variety of natural language processing tasks. It is unclear whether LLMs can achieve human-like cognitive capabilities or whether these models are still fundamentally circumscribed. We introduce a new benchmark for evaluating language models beyond memorization on abstract reasoning tasks.
  arXiv  Detail & Related papers  (2023-05-31T04:50:29Z)
• Large Language Models are In-Context Semantic Reasoners rather than Symbolic Reasoners [75.85554779782048]
  Large Language Models (LLMs) have excited the natural language and machine learning community over recent years. Despite of numerous successful applications, the underlying mechanism of such in-context capabilities still remains unclear. In this work, we hypothesize that the learned textitsemantics of language tokens do the most heavy lifting during the reasoning process.
  arXiv  Detail & Related papers  (2023-05-24T07:33:34Z)
• Post Hoc Explanations of Language Models Can Improve Language Models [43.2109029463221]
  We present a novel framework, Amplifying Model Performance by Leveraging In-Context Learning with Post Hoc Explanations (AMPLIFY) We leverage post hoc explanation methods which output attribution scores (explanations) capturing the influence of each of the input features on model predictions. Our framework, AMPLIFY, leads to prediction accuracy improvements of about 10-25% over a wide range of tasks.
  arXiv  Detail & Related papers  (2023-05-19T04:46:04Z)
• Interpretability at Scale: Identifying Causal Mechanisms in Alpaca [62.65877150123775]
  We use Boundless DAS to efficiently search for interpretable causal structure in large language models while they follow instructions. Our findings mark a first step toward faithfully understanding the inner-workings of our ever-growing and most widely deployed language models.
  arXiv  Detail & Related papers  (2023-05-15T17:15:40Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.