Rethinking Confidence in Reasoning Models
Cohere, Tuesday, October 20th, 2026
Cohere researchers introduce CALIBER, a method that trains reasoning models to estimate and update their chance of being right.
Cohere notes that producing a detailed reasoning trace does not guarantee a model can reliably assess its own uncertainty.
A useful confidence estimate should change as reasoning reveals the answer is likely right or wrong.
Cohere's CALIBER method trains a reasoning model to estimate its chance of success before it begins and then assess its specific answer afterward.
On Cohere's math benchmark, a calibrated 7B model substantially reduced calibration error.