← Back to all news
arXiv Computation and Language

Calibration as a First-Class Criterion in LLM Evaluation

The paper argues that calibration—how well a language model’s confidence aligns with its actual correctness—should be a standard evaluation metric for large language models (LLMs). It notes that while calibration metrics exist, they are rarely applied outside specialized NLP subfields, leading to unverified confidence scores in new models, datasets, and benchmarks. The authors highlight the risks of miscalibration both at deployment (overconfident errors causing harm) and in research workflows (affecting LLM-as-a-judge, synthetic data generation, and active learning). They call for every NLP subfield to pair its primary performance metric with a calibration score, treating calibration as an essential property of every model.

Related stories

arXiv Computation and Language
Sep 1

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

The paper argues that traditional global calibration metrics, such as Expected Calibration Error and Brier Score, are confounded by differences in model accuracy when comparing large language models. It introduces ACE, an accuracy‑controlled evaluation framework that offers Instance‑Aligned, Distribution‑Aligned, and Candidate‑Aligned views to provide fairer cross‑model comparisons. Experiments across various benchmarks reveal that many reported calibration advantages disappear after accuracy control and that model rankings often reverse, indicating that raw global metrics are unreliable for cross‑model calibration assessment.

By Zhichao Yang, Caiqi Zhang, Ruihan Yang, Chengzu Li, Nigel Collier, Deqing Yang
llmsbenchmarks
More like this →
arXiv Computation and Language
Oct 22

S-EDL: Eliciting Self-Evidence from Sequence Likelihoods for Semantic Calibration of LLMs

S-EDL is a new evidential method that calibrates large language models for open‑ended generation by using the model’s own sequence likelihoods to build a differentiable Dirichlet prior over prompt‑specific semantic classes. This approach optimizes an evidential loss that provides a calibration‑aware training signal without relying on fixed multiple‑choice options. Experiments on four open‑ended benchmarks and three models show that S-EDL reduces semantic calibration error and improves accuracy, offering a framework for trustworthy LLM deployment in complex applications.

By Yawei Li, Jiazheng Li, David R\"ugamer, Bernd Bischl, Yulan He, Mina Rezaei
llmsbenchmarks
More like this →
arXiv Computation and Language
Sep 1

Beyond the Final Layer: Intermediate Representations for Better Multilingual Calibration in Large Language Models

The paper investigates multilingual confidence calibration in large language models, revealing that non‑English languages are systematically less well calibrated than English. By analyzing internal representations, the authors find that late‑intermediate layers provide a more reliable confidence signal than the final layer, which is biased by English‑centric training. They propose training‑free methods such as Language‑Aware Confidence Ensemble (LACE) to adaptively select optimal layers per language, aiming to improve global equity and trustworthiness of LLMs.

By Ej Zhou, Caiqi Zhang, Tiancheng Hu, Chengzu Li, Nigel Collier, Ivan Vuli\'c, Anna Korhonen
llmssafety
More like this →