When Calibration Rankings Reverse: Accuracy-Controlled Evaluation for Fair Comparison of LLMs

• URL: http://arxiv.org/abs/2606.30814v1
• Date: Mon, 29 Jun 2026 18:37:48 GMT
• Title: When Calibration Rankings Reverse: Accuracy-Controlled Evaluation for Fair Comparison of LLMs
• Abstract summary: We show that comparisons are confounded by differences in model accuracy.<n>For fairer cross-model comparison, we propose ACE, an accuracy-controlled evaluation framework.
• Score: 44.00626308821981
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Calibration evaluates whether a model confidence aligns with its empirical accuracy. Existing studies often compare the calibration of different large language models using global calibration metrics such as Expected Calibration Error and Brier Score. We begin by showing, both theoretically and empirically, that such comparisons are confounded by differences in model accuracy. For fairer cross-model comparison, we then propose ACE, an accuracy-controlled evaluation framework with three complementary views: Instance-Aligned, Distribution-Aligned, and Candidate-Aligned calibration. Across multiple benchmarks, model families, and confidence elicitation methods, we use ACE to study two practically important comparison axes, small versus large models and thinking versus non-thinking models. We find that many previously reported calibration advantages under raw global metrics weaken substantially after accuracy control. We also find that ranking reversal is frequent: models favored by raw metrics often cease to be favored once accuracy is controlled. Our results show that raw global calibration metrics are not robust for cross-model comparison, and that fair calibration comparison requires accuracy-aware evaluation.

Related papers

• CalArena: A Large-Scale Post-Hoc Calibration Benchmark [48.0798861811642]
  We introduce a large-scale, standardized benchmark for post-hoc calibration.<n>Our benchmark aggregates predictions from a diverse set of classical models, modern deep learning architectures, and foundation models.<n>We argue that Post-Hoc Improvement (PHI) in proper scoring rules offers a principled alternative to traditional calibration error estimators.
  arXiv  Detail & Related papers  (2026-05-28T16:31:36Z)
• All Models Are Miscalibrated, But Some Less So: Comparing Calibration with Conditional Mean Operators [12.103487148356747]
  We propose a kernel calibration error based on the Hilbert-Schmidt norm of the difference between conditional mean operators.<n>Our experiments show that CKCE provides a more consistent ranking of models by their calibration error and is more robust against distribution shift.
  arXiv  Detail & Related papers  (2025-02-17T05:52:09Z)
• Reassessing How to Compare and Improve the Calibration of Machine Learning Models [7.183341902583164]
  A machine learning model is calibrated if its predicted probability for an outcome matches the observed frequency for that outcome conditional on the model prediction.<n>We show that there exist trivial recalibration approaches that can appear seemingly state-of-the-art unless calibration and prediction metrics are accompanied by additional generalization metrics.
  arXiv  Detail & Related papers  (2024-06-06T13:33:45Z)
• Calibrating Large Language Models with Sample Consistency [76.23956851098598]
  We explore the potential of deriving confidence from the distribution of multiple randomly sampled model generations, via three measures of consistency. Results show that consistency-based calibration methods outperform existing post-hoc approaches. We offer practical guidance on choosing suitable consistency metrics for calibration, tailored to the characteristics of various LMs.
  arXiv  Detail & Related papers  (2024-02-21T16:15:20Z)
• From Uncertainty to Precision: Enhancing Binary Classifier Performance through Calibration [0.3495246564946556]
  Given that model-predicted scores are commonly seen as event probabilities, calibration is crucial for accurate interpretation. We analyze the sensitivity of various calibration measures to score distortions and introduce a refined metric, the Local Score. We apply these findings in a real-world scenario using Random Forest classifier and regressor to predict credit default while simultaneously measuring calibration.
  arXiv  Detail & Related papers  (2024-02-12T16:55:19Z)
• Calibration by Distribution Matching: Trainable Kernel Calibration Metrics [56.629245030893685]
  We introduce kernel-based calibration metrics that unify and generalize popular forms of calibration for both classification and regression. These metrics admit differentiable sample estimates, making it easy to incorporate a calibration objective into empirical risk minimization. We provide intuitive mechanisms to tailor calibration metrics to a decision task, and enforce accurate loss estimation and no regret decisions.
  arXiv  Detail & Related papers  (2023-10-31T06:19:40Z)
• On Calibrating Semantic Segmentation Models: Analyses and An Algorithm [51.85289816613351]
  We study the problem of semantic segmentation calibration. Model capacity, crop size, multi-scale testing, and prediction correctness have impact on calibration. We propose a simple, unifying, and effective approach, namely selective scaling.
  arXiv  Detail & Related papers  (2022-12-22T22:05:16Z)
• Better Uncertainty Calibration via Proper Scores for Classification and Beyond [15.981380319863527]
  We introduce the framework of proper calibration errors, which relates every calibration error to a proper score. This relationship can be used to reliably quantify the model calibration improvement.
  arXiv  Detail & Related papers  (2022-03-15T12:46:08Z)
• Localized Calibration: Metrics and Recalibration [133.07044916594361]
  We propose a fine-grained calibration metric that spans the gap between fully global and fully individualized calibration. We then introduce a localized recalibration method, LoRe, that improves the LCE better than existing recalibration methods.
  arXiv  Detail & Related papers  (2021-02-22T07:22:12Z)
• Mitigating Bias in Calibration Error Estimation [28.46667300490605]
  We introduce a simulation framework that allows us to empirically show that ECE_bin can systematically underestimate or overestimate the true calibration error. We propose a simple alternative calibration error metric, ECE_sweep, in which the number of bins is chosen to be as large as possible.
  arXiv  Detail & Related papers  (2020-12-15T23:28:06Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.