Beyond Drift: Stabilizing Subjective LLM Evaluation with Information-Theoretic Rubrics

Wang Xu, Shan Zhao, Haolan Chen, Tianwei Yan, Chengyu Wang, Shizhao Chen, Qian Wan
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:142864-142881, 2026.

Abstract

Despite the growing use of large language models (LLMs) in subjective tasks such as role-playing, humor, emotional intelligence, and dialogue quality, their evaluation faces a pressing reproducibility crisis: even the same evaluator may contradict itself when re-judging the exact same sample. We attribute this instability to dimension drift, where free-form evaluation protocols (e.g., Chain-of-Thought reasoning) unpredictably shift their implicit criteria, undermining reliability. To address this fundamental challenge, we reformulate subjective evaluation as an information-theoretic optimization problem. Specifically, we propose an Expected Information Gain (EIG)-based framework that constructs a stable yet adaptive personalized rubric to eliminate dimension drift. Our two-stage “generate–then–score” design first produces a diverse pool of candidate evaluation questions and then selects the most informative subset via EIG, yielding explicit and repeatable criteria. Experiments on six benchmarks, including CharacterEval, The rJokes, and MT_bench, demonstrate that our approach substantially improves both evaluation consistency and alignment with human judgments, outperforming CoT-based and fixed-questionnaire baselines. These results highlight that information-theoretic questionnaire construction offers a principled and reliable path toward reproducible evaluation of subjective tasks.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-xu26cm, title = {Beyond Drift: Stabilizing Subjective {LLM} Evaluation with Information-Theoretic Rubrics}, author = {Xu, Wang and Zhao, Shan and Chen, Haolan and Yan, Tianwei and Wang, Chengyu and Chen, Shizhao and Wan, Qian}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {142864--142881}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/xu26cm/xu26cm.pdf}, url = {https://proceedings.mlr.press/v306/xu26cm.html}, abstract = {Despite the growing use of large language models (LLMs) in subjective tasks such as role-playing, humor, emotional intelligence, and dialogue quality, their evaluation faces a pressing reproducibility crisis: even the same evaluator may contradict itself when re-judging the exact same sample. We attribute this instability to dimension drift, where free-form evaluation protocols (e.g., Chain-of-Thought reasoning) unpredictably shift their implicit criteria, undermining reliability. To address this fundamental challenge, we reformulate subjective evaluation as an information-theoretic optimization problem. Specifically, we propose an Expected Information Gain (EIG)-based framework that constructs a stable yet adaptive personalized rubric to eliminate dimension drift. Our two-stage “generate–then–score” design first produces a diverse pool of candidate evaluation questions and then selects the most informative subset via EIG, yielding explicit and repeatable criteria. Experiments on six benchmarks, including CharacterEval, The rJokes, and MT_bench, demonstrate that our approach substantially improves both evaluation consistency and alignment with human judgments, outperforming CoT-based and fixed-questionnaire baselines. These results highlight that information-theoretic questionnaire construction offers a principled and reliable path toward reproducible evaluation of subjective tasks.} }
Endnote
%0 Conference Paper %T Beyond Drift: Stabilizing Subjective LLM Evaluation with Information-Theoretic Rubrics %A Wang Xu %A Shan Zhao %A Haolan Chen %A Tianwei Yan %A Chengyu Wang %A Shizhao Chen %A Qian Wan %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-xu26cm %I PMLR %P 142864--142881 %U https://proceedings.mlr.press/v306/xu26cm.html %V 306 %X Despite the growing use of large language models (LLMs) in subjective tasks such as role-playing, humor, emotional intelligence, and dialogue quality, their evaluation faces a pressing reproducibility crisis: even the same evaluator may contradict itself when re-judging the exact same sample. We attribute this instability to dimension drift, where free-form evaluation protocols (e.g., Chain-of-Thought reasoning) unpredictably shift their implicit criteria, undermining reliability. To address this fundamental challenge, we reformulate subjective evaluation as an information-theoretic optimization problem. Specifically, we propose an Expected Information Gain (EIG)-based framework that constructs a stable yet adaptive personalized rubric to eliminate dimension drift. Our two-stage “generate–then–score” design first produces a diverse pool of candidate evaluation questions and then selects the most informative subset via EIG, yielding explicit and repeatable criteria. Experiments on six benchmarks, including CharacterEval, The rJokes, and MT_bench, demonstrate that our approach substantially improves both evaluation consistency and alignment with human judgments, outperforming CoT-based and fixed-questionnaire baselines. These results highlight that information-theoretic questionnaire construction offers a principled and reliable path toward reproducible evaluation of subjective tasks.
APA
Xu, W., Zhao, S., Chen, H., Yan, T., Wang, C., Chen, S. & Wan, Q.. (2026). Beyond Drift: Stabilizing Subjective LLM Evaluation with Information-Theoretic Rubrics. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:142864-142881 Available from https://proceedings.mlr.press/v306/xu26cm.html.

Related Material