Settle: Learning When to Stop Reasoning

2026-09-30

AI Summary

This study addresses the computational inefficiency of reasoning models that continue generating tokens after their answers have already converged. We propose a novel answer-stability-based stopping mechanism that learns optimal termination timing from stability signals within complete trajectories. Through constrained optimization, only the end-of-sequence (EOS) token is fine-tuned, strictly preserving the base model’s predictive distribution without introducing inference overhead or requiring non-standard decoding strategies. This approach accurately identifies termination points while effectively predicting answer correctness, substantially expanding the accuracy-length Pareto frontier. On MATH-500, our method reduces generated tokens by 40% with merely a 0.5% accuracy drop, outperforming an equivalent-data supervised fine-tuning baseline by 6.16 percentage points.

Abstract

Reasoning models often continue generating after their answers have settled. Settle learns when to stop from answer stability in completed traces. It trains the existing end-of-reasoning token while keeping other predictions close to the base model, and requires only ordinary decoding at inference. On MATH-500 with Qwen3-4B, Settle reduces token count by 40% with a 0.5-percentage-point decrease in accuracy. It gains 6.16 percentage points over supervised fine-tuning on the same traces shortened at their first stable answer, at nearly identical token counts. Its stopping score predicts whether a correct answer will remain correct. Settle extends the accuracy-token-count Pareto frontier of the evaluated stopping methods.

Problem

• Reasoning models
• Early stopping
• Token efficiency
• Answer stability
• Inference optimization

Innovation

• Reasoning Models
• Early Stopping
• Answer Stability
• Token Efficiency
• Pareto Frontier