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

Do NOT Think That Much for 2+3=? On the Overthinking of Long Reasoning Models

ICML 2025 · 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.

Studies overthinking in long reasoning models (QwQ-32B-Preview, DeepSeek-R1), proposes outcome and process efficiency metrics, and mitigates it via self-generated shorter traces trained with SimPO, reducing tokens while keeping accuracy.

Problem

Long reasoning models spend excessive tokens on simple problems, generating redundant solutions that add little accuracy or diversity, so test-time compute is used inefficiently.

Approach

Define outcome efficiency (tokens up to first correct answer over total tokens) and process efficiency (tokens in distinct reasoning strategies over total tokens), then build self-generated concise traces (First-Correct Solutions plus one reflection) and train with SimPO on (short, longest) pairs to reduce overthinking.

Key results

• SimPO with FCS+Reflection on QwQ-32B-Preview reduces MATH500 tokens from 2398.5 to 1330.3 (about 45%) with accuracy 91.4 vs 91.2.
• On GSM8K, FCS+Reflection raises accuracy from 94.7 to 95.9 while cutting tokens from 756.1 to 416.5.
• On AIME24, tokens drop from 9480.6 to 5154.4 with accuracy falling from 46.7 to 43.3.

Measured effects reported in the paper

Surprising findings

• In over 80% of cases the first solution already yields the correct answer, so most generated tokens come after the answer is found.
• Long reasoning models like QwQ-32B-Preview produce about 20x more tokens than conventional LLMs on trivial questions such as 2+3.

Limitations

• Gains on the hardest benchmark (AIME24) come with an accuracy drop
• Solution segmentation and strategy clustering rely on LLM judges (Llama-3.3-70B, GPT-4o)
• Evaluated on a single base model family (QwQ-32B-Preview)

Experimental setup

Datasets: PRM12K ASDIV GSM8K MATH500 GPQA Diamond AIME24

Benchmarks: ASDIV GSM8K MATH500 GPQA Diamond AIME24

Metrics: accuracy outcome efficiency process efficiency number of solutions number of tokens

Baselines: QwQ-32B-Preview Llama-3.3-70B-Instruct Qwen2.5-Math-72B-Instruct SFT on shortest response DPO on shortest response RPO on shortest response

Builds on: SimPO DPO RPO self-taught reasoner

Closest papers by meaning

• Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.92)
• Efficient Reasoning with Balanced Thinking (ICLR 2026 · similarity 0.92)
• Let LRMs Break Free from Overthinking via Self-Braking Tuning (NeurIPS 2025 · similarity 0.91)
• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.91)
• Mitigating Overthinking in Large Reasoning Models via Manifold Steering (NeurIPS 2025 · similarity 0.90)
• Beware of Reasoning Overconfidence: Pitfalls in the Reasoning Process for Multi-solution Tasks (AAAI 2026 · similarity 0.90)