Beyond Post-Hoc Temperature Scaling: Bilevel Optimization for LLM Calibration

TL;DR - An arXiv preprint proposing to fix LLM overconfidence from preference alignment during training rather than with post-hoc temperature scaling, by using bilevel optimization to maximize predictive entropy. It matters because calibration fitted post-hoc on one domain doesn't transfer, while this approach targets out-of-domain generalization.

• Frames calibration as maximizing the entropy of predictive distributions, directly penalizing overly concentrated (overconfident) predictions.
• Uses a bilevel formulation inspired by temperature scaling: the lower level trains the model under a parametric loss, the upper level selects loss hyperparameters to maximize entropy.
• Applies an efficient first-order approximation to avoid explicit second-order computation, making it tractable at LLM scale.
• Evaluated on multiple-choice and open-ended generative QA, reporting better-calibrated models with particular gains out-of-domain (no numeric results given in the abstract).