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

Let LRMs Break Free from Overthinking via Self-Braking Tuning

NeurIPS 2025 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 1 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.

Proposes Self-Braking Tuning (SBT), which builds SFT data from OpenR1-Math by detecting overthinking via a step-level overthink score and masking redundant reasoning, so that Qwen2.5-Math and Llama models learn to stop reasoning on their own, cutting tokens by up to ~60% with comparable accuracy.

Problem

Large reasoning models generate excessively long, redundant chains of thought (overthinking), raising compute and latency; most existing fixes rely on external controls such as token budgets or inference-time interventions.

Approach

Identify overthinking via an overthink score combining reasoning efficiency ratio (first-correct-step position) and overthinking marker ratio, then build SFT trajectories that keep the foundation solution plus part of an evolution solution, mask a small redundant segment from loss, and add natural-language braking phrases so the model learns to terminate reasoning internally.

Key results

• SBT-E on Llama-3.1-8B-Instruct reduces average tokens by 62.8% (8576 to 3193) with average accuracy 45.73% vs 48.59% baseline.
• SBT-E on Qwen2.5-Math-7B-Instruct cuts average tokens from 6029 to 4178 (30.7%) with accuracy 75.54% vs 78.19%.
• SBT-E on Qwen2.5-Math-1.5B-Instruct gives 57.83% average accuracy vs 59.36% baseline with 1673 vs 3277 tokens.

Measured effects reported in the paper

Surprising findings

• For Llama-3.1-8B, SBT-D improves MATH500 accuracy by 2.62 points while cutting tokens by 58.7%.
• Removing masked redundant thinking gives only +0.19 points accuracy but raises tokens by 37.8%.

Limitations

• Evaluated mainly on math reasoning; generalization limited to a small out-of-domain appendix.
• Overthinking thresholds and beta weighting are hand-tuned.
• Training limited to a 92K-example dataset; scaling unexplored.

Experimental setup

Datasets: OpenR1-Math

Benchmarks: AIME24 AIME25 AMC23 MATH500 GSM8K MMLU-Redux GPQA-Diamond

Metrics: accuracy average tokens

Baselines: OpenR1-Math SFT baseline (unmodified data) Token-Budget CoD DEER NoThinking LightThinker

Builds on: OpenR1-Math DeepSeek-R1 Megatron-LM

Closest papers by meaning

• Decoding the Critique Mechanism in Large Reasoning Models (NeurIPS 2026 · similarity 0.91)
• Do NOT Think That Much for 2+3=? On the Overthinking of Long Reasoning Models (ICML 2025 · similarity 0.91)
• Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring (NeurIPS 2026 · similarity 0.91)
• Efficient Reasoning with Balanced Thinking (ICLR 2026 · similarity 0.90)
• Mitigating Overthinking in Large Reasoning Models via Manifold Steering (NeurIPS 2025 · similarity 0.90)
• Stop When Further Reasoning Won’t Help: Attention-State Adaptive Generation in Reasoning Models (ICML 2026 · similarity 0.89)