arXiv cs.AI· Yifan Wang, Jinyi Mu, Mayank Jobanputra, Yu Wang, Soyoung Oh, Isabel Valera, Vera Demberg·· 10 小时前AI 评分18
CoCo:用贡献对比实现 MoE 奖励模型的响应级忠实解释
Beyond Routing Weights: Faithful Response-Level Interpretation of Mixture-of-Experts Reward Models via Contribution Contrast
AI 导读
针对稀疏 MoE 奖励模型仅靠路由权重只能说明专家"接收哪些提示词"、无法反映其"如何评判响应"的问题,研究者提出 Contribution-Contrast(CoCo)响应级解释方法,用贡献对比最大的 chosen-rejected 响应对刻画专家角色。在自动与人工评估中,CoCo 的解释比基于路由器、基于分数和基于稀疏自编码器的方案更连贯、忠实且更具专精性,同时保持有竞争力的奖励建模准确率。
正文 · 原文
Abstract:Reward models are central to learning from human preferences, yet identifying what drives their predictions remains challenging. Recent sparse Mixture-of-Experts (MoE) reward models seek to improve interpretability by routing prompts to specialized experts and characterizing experts through examples with high routing weights. However, routing weights only reveal which prompts an expert $\textit{receives}$, not how it $\textit{judges}$ responses, providing only a partial account of expert behavior. We therefore propose $\textbf{Co}$ntribution-$\textbf{Co}$ntrast ($\textbf{CoCo}$) response-level interpretation, which faithfully characterizes experts' roles using chosen-rejected response pairs with the largest contribution contrasts, jointly capturing routing and preference behavior. Across automatic and human evaluations, CoCo yields more coherent, faithful, and specialized interpretations than router-based, score-based, and sparse autoencoder-based alternatives while maintaining competitive reward modeling accuracy. Overall, we provide a controlled study of how different example-selection signals affect the interpretability of MoE reward models.
| Comments: | EACL 2027 (Findings) |
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2608.06400 [cs.AI] |
| (or arXiv:2608.06400v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2608.06400 arXiv-issued DOI via DataCite |
Submission history
From: Yifan Wang [view email]
[v1]
Fri, 31 Jul 2026 17:43:30 UTC (66 KB)
[v2]
Tue, 17 Nov 2026 23:12:42 UTC (72 KB)
来源:arXiv cs.AI · arxiv.org