Reliability-Aware LLM Alignment from Inconsistent Human Feedback

Jingyi Huang, Ruohan Zong, Yujun Feng, Liran Ma, Lanyu Shang, Yang Zhang
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:48601-48620, 2026.

Abstract

Reinforcement Learning from Human Feedback (RLHF) is critical for aligning Large Language Models (LLMs) with human preferences. However, its efficacy is often compromised by the inherent inconsistency and subjectivity of human annotations. Existing preference optimization frameworks, such as Direct Preference Optimization (DPO), typically treat ambiguous pairs with high annotator disagreement identically to those with unanimous consensus, forcing models to overfit to inconsistent supervision signals and leading to suboptimal alignment. In this work, we propose $\textit{Reliability-Guided Preference Optimization}$ (RGPO), a robust framework designed to mitigate the impact of inconsistent human feedback. RGPO estimates annotator reliability and infers latent ground truth labels from noisy human feedback to identify robust preferences. Furthermore, we introduce a reliability-aware consistency optimization that dynamically modulates the training objective based on the consensus level of annotations, ensuring the model prioritizes high-consensus supervision signals. Extensive experiments on LLM alignment benchmarks demonstrate that RGPO effectively reduces inconsistency and noise in training data and achieves superior performance compared to widely adopted RLHF baselines. Our code and configurations are available at https://github.com/GenieHuang/RGPO.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-huang26co, title = {Reliability-Aware {LLM} Alignment from Inconsistent Human Feedback}, author = {Huang, Jingyi and Zong, Ruohan and Feng, Yujun and Ma, Liran and Shang, Lanyu and Zhang, Yang}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {48601--48620}, 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/huang26co/huang26co.pdf}, url = {https://proceedings.mlr.press/v306/huang26co.html}, abstract = {Reinforcement Learning from Human Feedback (RLHF) is critical for aligning Large Language Models (LLMs) with human preferences. However, its efficacy is often compromised by the inherent inconsistency and subjectivity of human annotations. Existing preference optimization frameworks, such as Direct Preference Optimization (DPO), typically treat ambiguous pairs with high annotator disagreement identically to those with unanimous consensus, forcing models to overfit to inconsistent supervision signals and leading to suboptimal alignment. In this work, we propose $\textit{Reliability-Guided Preference Optimization}$ (RGPO), a robust framework designed to mitigate the impact of inconsistent human feedback. RGPO estimates annotator reliability and infers latent ground truth labels from noisy human feedback to identify robust preferences. Furthermore, we introduce a reliability-aware consistency optimization that dynamically modulates the training objective based on the consensus level of annotations, ensuring the model prioritizes high-consensus supervision signals. Extensive experiments on LLM alignment benchmarks demonstrate that RGPO effectively reduces inconsistency and noise in training data and achieves superior performance compared to widely adopted RLHF baselines. Our code and configurations are available at https://github.com/GenieHuang/RGPO.} }
Endnote
%0 Conference Paper %T Reliability-Aware LLM Alignment from Inconsistent Human Feedback %A Jingyi Huang %A Ruohan Zong %A Yujun Feng %A Liran Ma %A Lanyu Shang %A Yang Zhang %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-huang26co %I PMLR %P 48601--48620 %U https://proceedings.mlr.press/v306/huang26co.html %V 306 %X Reinforcement Learning from Human Feedback (RLHF) is critical for aligning Large Language Models (LLMs) with human preferences. However, its efficacy is often compromised by the inherent inconsistency and subjectivity of human annotations. Existing preference optimization frameworks, such as Direct Preference Optimization (DPO), typically treat ambiguous pairs with high annotator disagreement identically to those with unanimous consensus, forcing models to overfit to inconsistent supervision signals and leading to suboptimal alignment. In this work, we propose $\textit{Reliability-Guided Preference Optimization}$ (RGPO), a robust framework designed to mitigate the impact of inconsistent human feedback. RGPO estimates annotator reliability and infers latent ground truth labels from noisy human feedback to identify robust preferences. Furthermore, we introduce a reliability-aware consistency optimization that dynamically modulates the training objective based on the consensus level of annotations, ensuring the model prioritizes high-consensus supervision signals. Extensive experiments on LLM alignment benchmarks demonstrate that RGPO effectively reduces inconsistency and noise in training data and achieves superior performance compared to widely adopted RLHF baselines. Our code and configurations are available at https://github.com/GenieHuang/RGPO.
APA
Huang, J., Zong, R., Feng, Y., Ma, L., Shang, L. & Zhang, Y.. (2026). Reliability-Aware LLM Alignment from Inconsistent Human Feedback. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:48601-48620 Available from https://proceedings.mlr.press/v306/huang26co.html.

Related Material