The researchers propose a self-supervised fine-tuning procedure called ConfSFT that trains reasoning models to predict their confidence at intermediate points along their own reasoning trajectories. The confidence labels are derived entirely from the model's own token probabilities, requiring no gold answers or external judges. The loss contains no objective for reasoning length or efficiency, and at inference the model uses standard generation with no confidence elicitation or stopping.
Despite this, the fine-tuned models generate up to 25% fewer tokens at matched accuracy across Gemma, Qwen, Nemotron, and GPT-OSS on mathematical, scientific, and coding reasoning benchmarks. This efficiency gain is comparable to training methods that explicitly optimize for shorter reasoning. The authors also show that confidence prediction improves over training while accuracy remains stable, and that reasoning composition is largely preserved.
This result is interesting because it suggests that efficient reasoning can emerge as a downstream consequence of learning metacognitive signals, without being directly optimized. It offers a simple approach that requires only 600 training problems and no change to inference.