DynaCF: Mitigating Shortcut Learning in Reward Models via Dynamic Counterfactual Sensitivity

This paper studies what discriminatively trained reward models (RMs) memorize by measuring counterfactual memorization on two human preference datasets. We show that RMs 1) misallocate memorization to easy, high margin preference pairs, 2) memorize dataset-specific shortcuts (e.g., model identity, user sampling strategy), and 3) overgeneralize simple heuristic correlates of human preference (e.g., length, compliance) when confronted with unseen preference pairs. Overall, our findings indicate that discriminative training of RMs from human preference data results in biased RMs not yet capable of judging response quality in context-dependent scenarios.

Despite their widespread use, the role of reward models in shaping reinforcement learning is poorly understood. Reward models offer a tempting promise: they automatically estimate response quality in the absence of verifiers or human judges. Unlike "verifiable rewards" which typically produce binary scores, reward models typically produce continuous scores, allowing them to be sensitive to fine-grained differences in responses. However, we show this apparent strength is a serious weakness: many popular reward models are oversensitive, assigning different scores to equally good responses. Theoretically, we show that seemingly perfect reward models can be highly oversensitive; empirically, this oversensitivity can lead to bad policies. In place of existing notions of "reward model accuracy," we propose evaluating reward models using distinct measures of "discriminative ability" and "specificity" (the complement of oversensitivity). As a solution, we describe a training-free algorithm that uses Monte Carlo dropout on any neural reward model to produce discrete reward clusters. Theoretically, we prove there exist discretizations that reduce oversensitivity at minimal expense of discriminative ability; empirically we show, in both controlled and natural RL settings, that discretizing rewards leads to less reward hacking and better policies than training on the original rewards.

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.