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

Your Models Have Thought Enough: Training Large Reasoning Models to Stop Overthinking

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

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.

Proposes Just-Enough Thinking (JET), a DAPO-based RL method that mixes truncated-and-completed reasoning rollouts with a quality-controlled length reward to train LRMs to stop overthinking.

Problem

Large reasoning models produce overly long chains of thought, and RL for concise reasoning struggles because LRMs rarely generate short correct rollouts, leaving biased training signal.

Approach

Truncate the model's own full reasoning traces at several points, insert a stop-thinking cue to complete answers (keeping samples in the model's distribution), and assign a length reward only to correct responses normalized by the shortest and longest correct answer in each group.

Key results

• On Olympiad with DeepSeek-Distill-Qwen-1.5B, JET gains 4.6 accuracy points while reducing output length by 46.3%.
• Averaged over five math benchmarks, JET on the 1.5B model reaches 63.3 accuracy (+4.4) with 29.8% length reduction versus base.
• On GPQA-Diamond with the 1.5B model, JET improves accuracy by +11.1 points while cutting length by 25.6%.

Surprising findings

• Truncating reasoning sometimes turns initially incorrect answers into correct ones, suggesting extra reasoning can mislead the model.
• Without a length reward the model quickly drifts back to much longer answers, even when trained on truncated rollouts.

Limitations

• Baseline results for AdaptThink, LCR1 and Laser were taken from released checkpoints because reproduction did not match reported numbers.
• Length reward weights and reward design are sensitive; mismatched accuracy/length weights cause reward hacking.
• Evaluated only on two distilled Qwen models plus one Llama 8B model, with mostly math-focused training data.

Experimental setup

Datasets: MATH DAPO-Math-17k

Benchmarks: GSM8K MATH500 AIME24 AMC Olympiad CommonsenseQA GPQA-Diamond MMLU

Metrics: accuracy output length token compression ratio accuracy change

Baselines: SFT DPO DAPO AdaptThink Laser-D Laser-DE LCR1

Builds on: DAPO Evidence Accumulation Models Laser AdaptThink OVERTHINK

Closest papers by meaning

• Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.89)
• Do NOT Think That Much for 2+3=? On the Overthinking of Long Reasoning Models (ICML 2025 · similarity 0.88)
• How Far Are We from Optimal Reasoning Efficiency? (NeurIPS 2025 · similarity 0.88)
• Promoting Efficient Reasoning with Verifiable Stepwise Reward (AAAI 2026 · similarity 0.88)
• Does Your Reasoning Model Implicitly Know When to Stop Thinking? (ICML 2026 · similarity 0.88)
• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.88)