Unsupervised Confidence Calibration for Reasoning LLMs from a Single Generation
For practitioners deploying reasoning LLMs, this method provides practical confidence calibration without requiring labels or multiple samples, addressing a key reliability bottleneck.
The paper introduces an unsupervised method for calibrating confidence estimates of reasoning LLMs using only a single generation at inference time, outperforming baselines across 5 tasks and 9 models, and improving downstream selective prediction and decision-making.
Reasoning language models can solve increasingly complex tasks, but struggle to produce the calibrated confidence estimates necessary for reliable deployment. Existing calibration methods usually depend on labels or repeated sampling at inference time, making them impractical in many settings. We introduce a method for unsupervised confidence calibration of reasoning LLMs when only a single generation is available at inference time. Our approach uses offline sampling on unlabeled data to derive a self-consistency-based proxy target, then distills this signal into a lightweight deployment-time confidence predictor. In a broad evaluation across 5 math and question-answering tasks using 9 reasoning models, our method substantially outperforms baselines, including under distribution shift, and improves downstream performance in selective prediction and simulated downstream decision-making.