Cohere trains reasoning models to reassess confidence after answering

Cohere introduced CALIBER, a training framework that makes a reasoning model state its confidence before working through a problem and reassess a specific answer after it has reasoned. The method trains the two scores to represent different probabilities rather than repeated estimates of the same thing.

In Cohere’s company-run tests, a 7-billion-parameter model trained with CALIBER cut Expected Calibration Error on the held-out BigMathDigits math set from 0.059 for the strongest single-confidence baseline to 0.028, a 52.5% reduction. The same model recorded a Brier score of 0.130 and an AUROC of 0.806, the best results among the evaluated methods on all three confidence metrics for that setting, according to the Cohere Labs preprint.

CALIBER trains the initial score against the empirical success rate from multiple attempts at the same prompt. It trains the final score against whether that individual attempt was correct. The first estimate asks how often the current model can solve the question; the second can use the generated reasoning and answer to assess that attempt. An answer-correctness reward remains in the objective so a model cannot earn a strong calibration reward merely by being confidently pessimistic about wrong answers.

Calibration and discrimination are not interchangeable. ECE measures how closely stated confidence matches observed accuracy, while AUROC measures how well confidence ranks correct answers above incorrect ones. On BigMathDigits, the 30-billion-parameter CALIBER model had ECE of 0.022 and AUROC of 0.797. CoCA-lite produced nearly the same ECE, 0.023, but AUROC of 0.508. DCPO-lite and RLCR posted slightly higher AUROC than CALIBER, at 0.811 and 0.807, while recording higher ECE.

The 7B model also changed its estimates in different directions after reasoning on BigMathDigits: mean confidence rose from 78.5% to 83.5% on correct responses and fell from 62.6% to 54.5% on incorrect responses. Outside the math training distribution, Cohere reported ECE reductions against the strongest single-confidence baseline of 26.8% on GPQA, 32.2% on TriviaQA and 36.2% on SimpleQA. CALIBER did not have the best AUROC on GPQA or TriviaQA, and all tested methods were only about 5% accurate on SimpleQA, limiting what its confidence metrics show there.

The experiments trained two Cohere model architectures on 30,000 filtered numerical-answer math problems and evaluated them on 1,000 held-out math problems plus GPQA, TriviaQA and SimpleQA. The work addresses a different question from Cohere’s separate research on multilingual reasoning training. The authors selected checkpoints using the BigMathDigits evaluation split, used Cohere’s Command A as the judge for TriviaQA and SimpleQA and as a fallback judge for BigMathDigits, and removed malformed or repetitive-loop generations before calculating metrics.

The results come from a Cohere Labs-authored preprint and Cohere’s own blog; the cited sources provide no independent replication or external audit. Cohere also says it did not test whether the confidence estimates improve decisions to verify an answer, retrieve more information or abstain.