RESEARCH · RESEARCH · #2253
CALIBER trains reasoning models to report initial and final confidence
CALIBER is a training method that teaches reasoning models to emit two confidence estimates — an initial confidence (probability of solving the problem before reasoning) and a final confidence (probability the produced answer is correct). On a math benchmark, a calibrated 7B model using CALIBER cut calibration error (ECE) by more than half versus the best single-confidence baseline while also optimizing AUROC to help rank likely-correct answers.
KEY POINTS
- CALIBER is a training method that teaches reasoning models to emit two confidence estimates — an initial confidence (probability of solving the problem before reasoning) and a final confidence (probability the produced answer is correct).
- On a math benchmark, a calibrated 7B model using CALIBER cut calibration error (ECE) by more than half versus the best single-confidence baseline while also optimizing AUROC to help rank likely-correct answers.
- More accurate, instance-specific confidence estimates make model outputs more actionable by indicating when to trust, verify, or seek help for a given answer.
WHY IT MATTERS
More accurate, instance-specific confidence estimates make model outputs more actionable by indicating when to trust, verify, or seek help for a given answer.