Research map / Large language models / Reasoning and chain of thought

Decoding the Critique Mechanism in Large Reasoning Models

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Injects arithmetic errors into LRM reasoning traces, finds models often recover in the final answer without verbalizing it ("hidden critique"), extracts a difference-in-means critique vector, and steers LRMs to improve error detection and test-time scaling.

Problem

Understanding how large reasoning models internally detect and recover from errors in their chain-of-thought, and whether this self-critique ability can be identified and controlled without training.

Approach

Injecting arithmetic errors into the thinking trace reveals hidden self-correction in final answers; a difference-in-means activation vector between recovered and standard runs (the critique vector) is linearly separable, interpretable via logit lens, and steers error detection and test-time scaling when added to the residual stream.

Key results

• On GSM8K-Error, R1 models recover implicitly in 55%-78% of cases while explicit corrections stay below 6%.
• Positive steering (alpha=1.0) raises Qwen3-4B GSM8K-Error test-time scaling accuracy from 77% to nearly 90%, while negative steering drops it to slightly above 50%.
• Positive steering improves ProcessBench error detection accuracy by about 5% on all models; Qwen3-4B BIG-Bench Mistake error accuracy rises from 52% to 56%.

Surprising findings

• Models reach correct final answers despite corrupted, unverbalized reasoning traces in most error-injected cases.
• Recovery persists when the question is masked, suggesting reliance on the corrupted CoT rather than re-solving.

Limitations

• Steering is unconditional, adding a fixed vector at every step, which can over-revise correct traces and harm correct accuracy in some cases (e.g., Qwen3-4B).
• Transfer to harder benchmarks such as AIME gains are modest.
• Experiments limited to four open-source models in two families and mainly arithmetic error injection.

Experimental setup

Datasets: GSM8K MATH500 ProcessBench BIG-Bench Mistake AIME24 AIME25

Benchmarks: ProcessBench BIG-Bench Mistake GSM8K-Error MATH500-Error AIME24 AIME25 GSM8K MATH500

Metrics: AUROC ECE error accuracy correct accuracy F1 accuracy

Baselines: baseline without steering (alpha=0) budget forcing with Wait tokens

Builds on: budget forcing (s1) representation engineering steering vectors logit lens

Closest papers by meaning

• Let LRMs Break Free from Overthinking via Self-Braking Tuning (NeurIPS 2025 · similarity 0.91)
• Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring (NeurIPS 2026 · similarity 0.90)
• Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens (NeurIPS 2026 · similarity 0.90)
• RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Reasoning Intervention in Large Reasoning Models (ICLR 2026 · similarity 0.90)
• Do NOT Think That Much for 2+3=? On the Overthinking of Long Reasoning Models (ICML 2025 · similarity 0.90)
• The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · similarity 0.90)