Rethinking confidence in reasoning models

Cohere introduces CALIBER, a training method intended to improve how reasoning models estimate uncertainty by producing two distinct confidence scores: an initial estimate of whether the model can solve a question before reasoning and a final estimate of whether its specific completed answer is correct. The approach trains initial confidence using a model’s success rate across repeated attempts at the same problem, while final confidence is trained against the correctness of each individual response, alongside rewards for solving problems accurately. Cohere argues that useful confidence requires both calibration, measured by Expected Calibration Error, and the ability to distinguish correct from incorrect answers, measured by AUROC. In tests on math and question-answering benchmarks, CALIBER-trained 7B and 30B models generally reduced calibration error relative to compared single-confidence methods, with the 7B model reducing ECE by 52.5% on BigMathDigits and showing improvements on GPQA, TriviaQA, and SimpleQA without additional task-specific training. The findings also show that reasoning can raise confidence for correct answers and lower it for incorrect ones, although the direction of confidence changes varies by dataset; Cohere suggests that such updates could eventually support decisions to verify, retrieve information, or abstain, though those applications were not tested.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.