Cambridge's ACE framework: LLM calibration rankings often reverse when accuracy is controlled

A new paper on arXiv from Cambridge researchers (Zhichao Yang, Caiqi Zhang, Ruihan Yang, Chengzu Li, Nigel Collier, Deqing Yang) takes a critical look at how we compare calibration across LLMs. Calibration measures whether a model's confidence matches its actual accuracy. The team shows both theoretically and empirically that standard global metrics like Expected Calibration Error (ECE) and Brier Score are confounded by differences in model accuracy—meaning a model may appear better calibrated simply because it is more or less accurate, not because it gives better confidence estimates.

To solve this, they introduce ACE (Accuracy-Controlled Evaluation), a framework with three views: Instance-Aligned, Distribution-Aligned, and Candidate-Aligned calibration. Applying ACE across multiple benchmarks, model families, and confidence elicitation methods, they demonstrate that many previously reported calibration advantages weaken after controlling for accuracy. Crucially, ranking reversals are frequent: models favored by raw global metrics often cease to be favored once accuracy is equalized. Their results call into question the robustness of existing calibration comparisons and advocate for accuracy-aware evaluation going forward.