Learning to Stop without Learning to Stop: Self-Supervised Confidence Training Improves Reasoning Efficiency

TL;DR

A self-supervised method that trains models to predict their confidence during reasoning leads to shorter, more efficient reasoning traces without explicit length penalties or early-stopping mechanisms. The approach improves inference efficiency across multiple models and benchmarks, reducing token count by up to 25% while maintaining accuracy. Efficiency gains emerge naturally from learning metacognitive signals rather than direct optimization for brevity.

Why It Matters

This approach offers a simple, plug-and-play way to enhance reasoning efficiency without modifying inference procedures or training objectives. It reduces computational costs in real-world deployments and suggests that metacognitive awareness can naturally improve reasoning quality and speed. The method is broadly applicable across model architectures and domains.

Key Findings

• Self-supervised confidence training on just 600 problems significantly reduces reasoning length during inference without any explicit efficiency objective.
• Efficiency improvements occur without altering the generation process or adding early-stopping logic, indicating emergent behavior from confidence learning.
• The method preserves the original reasoning structure and composition, suggesting that efficiency gains stem from better internal monitoring rather than behavioral suppression.