Problem
Research questions and friction points this paper is trying to address.
confidence calibration
reasoning language models
state-dependent confidence
expected calibration error
out-of-distribution generalization
CALIBER: Calibrating Confidence Before and After Reasoning in Language Models
🤖 AI Summary
Current language models lack a confidence calibration mechanism that aligns with the evolving information state during reasoning, often relying on single-point estimates either before or after inference, which inadequately captures problem-solving capability or answer correctness. This work proposes CALIBER, a framework that explicitly distinguishes between pre-reasoning confidence (predicting problem-solving success) and post-reasoning confidence (predicting answer correctness). By designing position-target alignment to generate appropriate supervision signals, CALIBER achieves unified and precise two-stage calibration. Experiments demonstrate that the method substantially reduces Expected Calibration Error (ECE)—by 52.5% on BigBench-MathDigits for a 7B model—and consistently outperforms baselines on out-of-distribution benchmarks such as GPQA and TriviaQA, exhibiting strong scalability and robustness.
📝 Abstract
Reasoning language models are increasingly asked not only to answer difficult questions, but also to estimate their likelihood of success. Existing methods typically elicit confidence only once: either before thinking or after answering. We argue that confidence in reasoning models is state-dependent: before thinking, confidence should estimate the chance of the model correctly solving the prompt, while after thinking it should predict whether the realized answer is likely to be correct. This distinction determines the appropriate supervision target: prompt-level success should supervise confidence estimates made after seeing the prompt, while individual answer-level correctness should supervise confidence estimates made after answering. We introduce CALIBER (Calibration Before and After Reasoning), which elicits both estimates and supervises each with the target matched to its information state. Under this unified protocol, CALIBER reduces Expected Calibration Error (ECE) by 52.5% over the strongest single-confidence baseline on BigMathDigits for the 7B model, while achieving the best Brier score and AUROC, and remains within 2.1 points of the best accuracy. Further, on a larger 30B model, CALIBER achieves the best ECE on BigMathDigits while remaining competitive in Brier score and AUROC. Out of distribution, it achieves the best ECE and Brier score on GPQA and TriviaQA, and remains competitive on SimpleQA. Ablations further show that this position-target alignment is most beneficial under distribution shift where it consistently reduces calibration error across all out-of-distribution benchmarks.
Innovation
Methods, ideas, or system contributions that make the work stand out.
confidence calibration
reasoning language models
state-dependent confidence
Expected Calibration Error
CALIBER
🔎 Similar Papers
No similar papers found.
