SupportCal improves confidence calibration of post-trained language models
SupportCal: Label-Free Calibration of Post-Trained LLMs via Reference Support and Corroboration
Machine LearningComputation and Language
Summary
Language models trained for specific tasks can become overconfident in their answers, which makes their predictions less trustworthy. The researchers found that using the original, pretrained model as a reference helps adjust this overconfidence without needing extra labeled data. Their new method, SupportCal, cleverly balances examples where the task-trained model agrees or disagrees with the original model to better gauge confidence. Tests on medical and math question-answering tasks show SupportCal makes the model’s confidence more accurate than previous methods.
What this means in practice
- •For machine learning engineers: Calibrate large language models after task-specific training without using labeled calibration data to improve reliability of confidence scores.
- •For healthcare data teams: Improve confidence estimates of AI models answering medical questions by referencing pretrained language models, enhancing trustworthiness without costly labels.
Authors
Linhan Luo, Lequan Lin, Dai Shi, Feng Chen, José Miguel Hernández-Lobato, Junbin Gao
Abstract
Post-training often improves task performance but can degrade confidence calibration, leaving post-trained language models (PoLMs) more overconfident than their corresponding pretrained language models (PLMs). Because task-specific labeled calibration data can be costly or unavailable, the corresponding pretrained PLM provides a natural label-free reference for post-hoc calibration. Prior agreement-gated PLM-referenced calibration fits a scalar temperature using only examples on which the PoLM and its PLM reference agree, excluding disagreement examples because direct alignment can drive the fitted temperature excessively high and induce under-confidence. We revisit this binary treatment. A controlled reintroduction diagnostic reveals a non-monotonic aggregate effect: admitting a moderate fraction of disagreement examples can improve calibration, whereas the benefit diminishes as unit-weight inclusion approaches the full disagreement set. We introduce SupportCal, a label-free post-hoc method that retains agreement examples at unit weight and assigns disagreement examples continuous weights based on the own-base PLM's relative support and corroboration from pretrained references selected from a size-compatible candidate pool. We further characterize when the resulting weighted objective admits a finite optimal temperature. Across MedMCQA and MathQA, SupportCal yields lower ECE than the agreement-only baseline for nearly all evaluated target-model configurations; supplementary TweetEval Sentiment results show the same pattern on a fixed-label classification task.