← News

Models & toolsSources: arXiv · 19 Nov 2026

CoCo method reads MoE reward model experts at the response level

An arXiv paper proposes Contribution-Contrast (CoCo), an interpretation method that characterizes sparse Mixture-of-Experts reward model experts by how they judge responses, not just which prompts they receive.

In plain words

Reward models score how well AI answers match human preferences, and sparse Mixture-of-Experts versions split that job among specialized experts. The paper's authors argue that existing interpretations only show which prompts each expert receives, not how it judges answers. Their CoCo method instead describes each expert using response pairs where its contribution differs most, aiming to reveal its actual preference behavior.

What was announced

The authors' starting point is that routing weights, the signal usually used to characterize experts in sparse MoE reward models, give only a partial account: they show which prompts an expert receives but not how it judges responses. CoCo addresses this by identifying chosen-rejected response pairs with the largest contribution contrasts, so each expert is described through both its routing and its preference behavior. The paper frames this as a response-level interpretation of expert roles.

According to the abstract, automatic and human evaluations found CoCo produced more coherent, faithful, and specialized interpretations than router-based, score-based, and sparse autoencoder-based alternatives, while reward modeling accuracy stayed competitive. The authors also position the work as a controlled study of how different example-selection signals affect the interpretability of MoE reward models. The evidence available is the paper's abstract only, so these evaluation results have not been independently verified and the full paper's details are not covered here.

What it does not show

  • The evidence is limited to the paper's abstract; evaluation details, method specifics, and scope of experiments are not available in the supplied passages.
  • The reported evaluation results are the authors' own claims and have not been independently confirmed.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.