PRISM: Probabilistic Reward Model with Inherent Structural Modeling

摘要

Standard evaluators, such as reward models, compress diverse human judgments into a single scalar, conflating valid Subjective Preference with Cognitive Uncertainty. This structural mismatch often leads to brittle alignment and reward hacking. To address this, we propose PRISM which reinterprets reward evaluation as a conditional distribution parameterized by a Mixture of Gaussians. PRISM structurally disentangles these factors: distinct Gaussian experts emerge to capture conflicting preference dimensions, while their variance estimates quantify uncertainty, acting as a dynamic reliability gate during optimization. We introduce a two-stage training strategy to learn these disentangled representations from scalable pairwise comparisons without requiring massive fine-grained annotations. Empirical results show that PRISM significantly outperforms scalar baselines in both accuracy and generalization. Furthermore, in downstream Reinforcement Learning, PRISM effectively mitigates reward hacking, yielding policies that are more robust and resilient to distribution shifts.