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

Reasoning Can Be Restored by Correcting a Few Decision Tokens

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 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.

Shows that base-vs-reasoning token disagreement is sparse, early, and planning-linked, and proposes a training-free, disagreement-gated one-token takeover from a reasoning model that recovers much of the reasoning gap at small budgets.

Problem

Large reasoning models substantially outperform their base LLMs on hard reasoning benchmarks, but it is unclear where base models diverge token by token and how to close the gap with few interventions.

Approach

Quantify base-vs-reasoning token-level cross-entropy disagreement, find it is sparse, front-loaded, uncertainty-aligned and planning-enriched, then intervene with a one-token takeover by the reasoning model only at calibrated high-disagreement positions, switching back to the base model afterward.

Key results

• Qwen3-0.6B guided decoding at realized budget ~0.13 raises average Acc/Pass@8 from 13.0/36.0 to 52.4/80.0, surpassing Qwen3-0.6B Thinking (43.4/64.1).
• Qwen3-1.7B guided at ~0.16 reaches 62.1/83.8 vs. its Thinking variant 64.1/80.3.
• Matched-budget random and early-only baselines at 25% budget (26.4/55.2 and 25.7/58.4) are beaten by guided intervention at ~4% budget (29.1/61.4).

Surprising findings

• Disagreement-guided takeover at ~4% of tokens already outperforms random and early-only replacement at 25% budget.
• Planning tokens are 17.6x enriched among takeover tokens, suggesting interventions mostly target planning rather than arithmetic.

Limitations

• Primarily Qwen3 family and math benchmarks; LLaMA and GPQA evidence is only preliminary.
• Gate queries both base and reasoning models at every step, so it is a diagnostic rather than an efficient system.
• Planning/execution classification uses a heuristic keyword classifier.

Experimental setup

Datasets: GSM8K

Benchmarks: GSM8K MATH500 AMC23 OlympiadBench AIME24 AIME25 GPQA-Diamond

Metrics: Accuracy Pass@8 Recovery AUROC Gini coefficient

Baselines: Qwen3-0.6B (Thinking) Qwen3-1.7B (Thinking) Random token replacement Early-only replacement Qwen3-8B (Thinking)

Builds on: chain-of-thought RLVR reverse KL divergence

Closest papers by meaning

• On Reasoning Strength Planning in Large Reasoning Models (NeurIPS 2025 · similarity 0.90)
• On the Token Value Inequality in Efficient Reasoning (NeurIPS 2026 · similarity 0.90)
• RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Reasoning Intervention in Large Reasoning Models (ICLR 2026 · similarity 0.89)
• How Far Are We from Optimal Reasoning Efficiency? (NeurIPS 2025 · similarity 0.89)
• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.88)
• Demystifying Reasoning Dynamics with Mutual Information: Thinking Tokens are Information Peaks in LLM Reasoning (NeurIPS 2025 · similarity 0.88)