Decoding the Critique Mechanism in Large Reasoning Models
Summary
This paper investigates how large reasoning models can detect and correct their own errors internally, identifying a highly interpretable critique vector that enhances error detection without additional training, improving test-time scaling performance.
View Cached Full Text
Cached at: 05/26/26, 10:43 AM
Paper page - Decoding the Critique Mechanism in Large Reasoning Models
Source: https://huggingface.co/papers/2603.16331
Abstract
Large Reasoning Models demonstrate hidden critique abilities that allow error recovery through internal mechanisms, identified via interpretable critique vectors that enhance error detection without additional training.
Large Reasoning Models(LRMs) exhibit backtracking andself-verificationmechanisms that enable them to revise intermediate steps and reach correct solutions, yielding strong performance on complex logical benchmarks. We hypothesize that such behaviors are beneficial only when the model has sufficiently strong ``critique’’ ability to detect its own mistakes. This work systematically investigates how current LRMs recover from errors by inserting arithmetic mistakes in their intermediate reasoning steps. Notably, we discover a peculiar yet important phenomenon: despite the error propagating throughout the entirechain-of-thought(CoT) without any verbalized correction, the model still reaches the correct final answer after the thinking process finishes. This recovery implies the existence of an internal mechanism helping the model to detect errors and trigger self-correction, which we refer to as the hiddencritique ability. Building onfeature space analysis, we identify a highly interpretablecritique vectorrepresenting this behavior. Extensive experiments across multiple model scales and families demonstrate that steeringlatent representationswith this vector improves the model’s error detection capability and enhances the performance oftest-time scalingat no extra training cost. Our findings provide a valuable understanding of LRMs’ critique behavior, suggesting a promising direction to control and improve theirself-verificationmechanism. Our code is available at: https://github.com/mail-research/lrm-critique-vectors.
View arXiv pageView PDFGitHub0Add to collection
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2603.16331 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2603.16331 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2603.16331 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
An Enigma of Artificial Reason: Investigating the Production-Evaluation Gap in Large Reasoning Models
This paper investigates the production-evaluation gap in large reasoning models (LRMs), finding that they fail to robustly evaluate reasoning despite near-perfect solution production, due to an answer confirmation bias.
Enhanced and Efficient Reasoning in Large Learning Models
This paper proposes a method for improving reasoning in large language models by recoding data to explicitly represent relationships, enabling efficient principled reasoning with polynomial-time learnability for relational rules, which addresses hallucinations and supports sound reasoning across multiple calls.
Reasoning Can Be Restored by Correcting a Few Decision Tokens
This paper shows that the reasoning gap between base LLMs and large reasoning models is concentrated on a small set of early planning tokens. It introduces disagreement-guided token intervention, where replacing only those critical tokens with a reasoning model's outputs allows a base model to nearly match the reasoning model's performance.
Can We Understand How Large Language Models Reason?
This article explores the ongoing efforts and challenges in understanding how large language models reason, focusing on interpretability research.
Mathematical Reasoning in Large Language Models: Benchmarks, Architectures, Evaluation, and Open Challenges
This survey synthesizes recent advancements in mathematical reasoning with large language models, covering benchmarks, architectures, training strategies, and evaluation protocols. It identifies key challenges such as reasoning faithfulness and benchmark biases.