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

The First Impression Problem: Internal Bias Triggers Overthinking in Reasoning Models

ICLR 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.

Shows that a model's preliminary (direct-answer) guess, or "internal bias," conflicts with its later reasoning and triggers overthinking, supported by length correlations, question-removal and bias-injection interventions, and attention analysis.

Problem

Reasoning models overthink, producing redundant reflection steps, but the root cause of this behavior is underexplored. The paper asks what drives this tendency.

Approach

Models form a preliminary answer ("internal bias") from the question alone before reasoning; when this guess deviates from the final reasoned answer, the model reflects more. Removing the question after the first answer, and injecting bias via fine-tuning, causally shift reasoning length, and attention to the question during reflection appears to reactivate the bias.

Key results

• Removing the question after the first answer reduces reasoning length by a ratio r of 31.1% to 53.5% across datasets, with accuracy maintained or improved on complex tasks (e.g., AIME 2025 36.7% to 46.7%).
• In bias injection on CharCount (zh), Low2Wrong raised length from 355.5 to 454.9 tokens, while High2Correct cut length from 600.8 to 412.1 tokens.
• Mitigation methods (FCS SFT, FCS DPO, SEAL, PROBE) shorten reasoning but keep a comparable or larger R_Delta; FCS SFT drops AIME 2024 accuracy from 63.3% to 50.0%.

Surprising findings

• Removing the input question mid-reasoning improved accuracy on complex tasks such as AIME 2025, suggesting the removed redundant reasoning did not contribute to answers.
• Shortening methods lowered average length but did not eliminate the length gap driven by internal bias, and FCS SFT reduced accuracy on AIME 2024.

Limitations

• Analysis restricted to tasks with numerical, range, or multiple-choice answers; open-ended bias is undefined.
• Deviation degree is approximated by MAE or inconsistency rate from 64 direct answers, which is an approximation of the bias distribution.
• Mitigation baselines reproduced on a single model and mostly one dataset; the paper does not establish a mitigation method that removes the bias effect.

Experimental setup

Datasets: CharCount (zh) CharCount (en) KnowLogic AIME 2024 AIME 2025

Benchmarks: CharCount (zh) CharCount (en) KnowLogic AIME 2024 AIME 2025

Metrics: accuracy reasoning length (tokens) relative length increase R_Delta length reduction ratio r inconsistency rate MAE reflection keyword count

Baselines: First-Correct Solutions (FCS) SFT FCS DPO SEAL PROBE

Builds on: DeepSeek-R1 chain-of-thought SEAL PROBE

Closest papers by meaning

• Do NOT Think That Much for 2+3=? On the Overthinking of Long Reasoning Models (ICML 2025 · similarity 0.89)
• Let LRMs Break Free from Overthinking via Self-Braking Tuning (NeurIPS 2025 · similarity 0.88)
• Beware of Reasoning Overconfidence: Pitfalls in the Reasoning Process for Multi-solution Tasks (AAAI 2026 · similarity 0.88)
• Decoding the Critique Mechanism in Large Reasoning Models (NeurIPS 2026 · similarity 0.88)
• Efficient Reasoning with Balanced Thinking (ICLR 2026 · similarity 0.88)
• Mitigating Overthinking in Large Reasoning Models via Manifold Steering (NeurIPS 2025 · similarity 0.87)