leiqf/UtilityBench

UtilityBench: Measuring Language Model Utility When Errors Have Different Costs

Introduction

As language models are increasingly deployed in real-world scenarios with important outcomes, their mistakes can have consequences that correctness alone does not capture. A minor error in a low-stakes question and a dangerous error in medical advice should not carry the same practical significance. We therefore propose a utility-based evaluation paradigm that explicitly incorporates the varying consequences of errors, encouraging more responsible model behavior. As an early trial of this paradigm, we construct UtilityBench with continuous, sample-specific consequence annotations and introduce UtilityAUC.

Usage

Import to your project using huggingface dataset, or refer to the github repository for an end-to-end demo.

Dataset Fields

The dataset contains 15,000 records in a single test split.

Field	Type	Description
id	int64	Unique record identifier
input	string	Full prompt sent to the model
answerability	int8	1 = answerable; 0 = refusal required
reference_answer	string / null	Reference answer; null for refusal-required prompts
source_id	string	Source dataset and item identifier
eta	float64	Relative error-consequence score; higher means greater consequences
L	float64	Positive cost of an incorrect answer
evaluator_id	string	Evaluation rule or code-test identifier

Acknowledgement

UtilityBench builds on 18 source datasets. We thank their authors for making these resources publicly available. Record-level provenance is preserved in source_id.

Citations

If you find this work useful, please consider citing our paper:

@misc{lei2026utilitybench,
  author        = {Qifeng Lei and Zhiyong Yang and Zitai Wang and Shilong Bao and Linchao Zhang and Qianqian Xu and Qingming Huang},
  title         = {{UtilityBench}: Measuring Language Model Utility When Errors Have Different Costs},
  year          = {2026},
  eprint        = {2610.25924},
  archiveprefix = {arXiv},
}

Contact

If you encounter any problems using the dataset or evaluating your models, please open an issue in the github repository.