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

Parsa Hosseini, Akasha Tigalappanavara, Sumit Nawathe, Chenrui Fan, Sourya Basu et al.|September 25, 2026arXiv

Key Takeaway

Training language models to predict their confidence in intermediate reasoning steps—using only self-supervised learning—makes them generate shorter reasoning traces at inference time without any explicit length penalties or early-stopping mechanisms.

Summary

This paper shows that reasoning models can generate shorter, more efficient reasoning traces by learning to predict their own confidence in answers—without explicitly optimizing for length.

training efficiency reasoning

Key Terms

reasoning-trace self-supervised-learning metacognitive-signals inference-time-compute