New method proposed to interpret Mixture-of-Experts reward models via contribution contrast

Researchers propose Contribution-Contrast (CoCo), a method to interpret Mixture-of-Experts (MoE) reward models at the response level. Unlike prior approaches that rely on routing weights to infer expert behavior, CoCo analyzes chosen-rejected response pairs with the largest contribution contrasts to jointly capture routing and preference behavior.

The authors argue that routing weights alone reveal only which prompts an expert receives, not how it judges responses, providing an incomplete picture of expert behavior. CoCo addresses this gap by directly examining how experts contribute to preference judgments between responses.

In evaluations, CoCo outperformed router-based, score-based, and sparse autoencoder-based interpretation methods in coherence, faithfulness, and specialization, as measured by both automatic and human assessments. The method also maintained competitive reward modeling accuracy relative to these alternatives.

The study is presented as the first systematic investigation of interpretation methods for MoE reward models, highlighting the need for faithful response-level explanations in reward modeling.