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

Beware of Reasoning Overconfidence: Pitfalls in the Reasoning Process for Multi-solution Tasks

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency

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.

Introduces MuSoBench and the concept of reasoning overconfidence, showing Short-CoT prompting is overconfident on multi-solution tasks while Long-CoT reduces it, and proposes a cognitive-rigidity hypothesis supported by attention-entropy analysis.

Problem

LLMs perform poorly on multi-solution tasks requiring complete and diverse answer sets; they stop early yet report high confidence in incomplete solution sets (reasoning overconfidence).

Approach

Defines reasoning overconfidence as the gap between verbalized confidence and recall of the full solution set, builds a controlled benchmark (TimeTabling, SubsetSum) with exhaustively enumerated solutions, compares Short-CoT vs Long-CoT, and explains the effect via a cognitive-rigidity hypothesis tied to attention entropy.

Key results

• Short-CoT yields ECE above 78.22% for every model series, indicating pronounced reasoning overconfidence.
• Long-CoT lowers ECE by at least 18.28% across model families and reaches single-digit ECE for closed-source models, though open-source models still exceed 56.46% ECE.
• On Qwen3-8B, w/ Voting over 32 paths raises TimeTabling recall from 9.99% to 49.55% and lowers ECE (r) from 77.23% to 64.32%.

Surprising findings

• Decoding temperature has little effect on recall or calibration, suggesting overconfidence is not a sampling artifact.
• Median-confidence aggregation raises precision and recall but worsens calibration, while voting improves both.

Limitations

• Open-source models remain strongly overconfident even under Long-CoT (ECE above 56.46%).
• Benchmark covers only combinatorial tasks (TimeTabling, SubsetSum) plus a single creative generation dataset.
• Cognitive-rigidity hypothesis is supported only by preliminary attention-entropy analysis.

Experimental setup

Datasets: TimeTabling SubsetSum NoveltyBench

Benchmarks: MuSoBench (TimeTabling, SubsetSum) NoveltyBench

Metrics: precision recall expected calibration error correct solution retention rate error solution correction rate new solution discovery rate utility distinct

Baselines: Short-CoT Long-CoT w/ Median Conf w/ Voting

Builds on: chain-of-thought zero-shot CoT self-consistency NoveltyBench

Closest papers by meaning

• 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)
• Measuring the Faithfulness of Thinking Drafts in Large Reasoning Models (NeurIPS 2025 · similarity 0.89)
• Reasoning or Retrieval? A Study of Answer Attribution on Large Reasoning Models (ICLR 2026 · similarity 0.89)
• Rethinking External Slow-Thinking: From Snowball Errors to Probability of Correct Reasoning (ICML 2025 · similarity 0.89)
• Demystifying Reasoning Dynamics with Mutual Information: Thinking Tokens are Information Peaks in LLM Reasoning (NeurIPS 2025 · similarity 0.88)