PrefMoE: Robust Preference Modeling with Mixture-of-Experts Reward Learning

TL;DR

PrefMoE introduces a mixture-of-experts framework for preference modeling in reinforcement learning, enhancing robustness to heterogeneous and conflicting preference data.

Contribution

It proposes a novel mixture-of-experts reward learning method with adaptive trajectory-level routing and load-balancing, improving robustness over single reward models.

Findings

PrefMoE outperforms single-model baselines in preference prediction accuracy.

PrefMoE leads to more reliable policy learning in locomotion and manipulation tasks.

The framework effectively captures diverse latent preferences under noisy supervision.

Abstract

Preference-based reinforcement learning offers a scalable alternative to manual reward engineering by learning reward structures from comparative feedback. However, large-scale preference datasets, whether collected from crowdsourced annotators or generated by synthetic teachers, often contain heterogeneous and partially conflicting supervision, including disagreement across annotators and inconsistency within annotators. Existing reward learning methods typically fit a single reward model to such data, forcing it to average incompatible signals and thereby limiting robustness. To solve this, we propose PrefMoE, a mixture-of-experts reward learning framework for robust preference modeling. PrefMoE learns multiple specialized reward experts and uses trajectory-level soft routing to combine them adaptively, enabling the model to capture diverse latent preference patterns under noisy and heterogeneous preference supervision. A load-balancing regularizer further stabilizes training by preventing expert collapse. Across locomotion benchmarks from D4RL and manipulation tasks from MetaWorld, PrefMoE improves preference prediction robustness and leads to more reliable downstream policy learning than strong single-model baselines.