DMoERM: Recipes of Mixture-of-Experts for Effective Reward Modeling

TL;DR

This paper introduces DMoERM, a novel Mixture-of-Experts approach for reward modeling in large language models, addressing generalization and noise issues, and demonstrating superior performance and consistency with human preferences.

Contribution

We propose the Double-Layer MoE RM (DMoERM), combining sparse and dense models with task-specific experts to improve reward modeling effectiveness and robustness.

Findings

Outperforms advanced generative approaches in human preference alignment

Reduces overoptimization in reward modeling

Achieves superior consistency with human annotations

Abstract

The performance of the reward model (RM) is a critical factor in improving the effectiveness of the large language model (LLM) during alignment fine-tuning. There remain two challenges in RM training: 1) training the same RM using various categories of data may cause its generalization performance to suffer from multi-task disturbance, and 2) the human annotation consistency rate is generally only $60\%$ to $75\%$, causing training data to contain a lot of noise. To tackle these two challenges, we introduced the idea of Mixture-of-Experts (MoE) into the field of RM for the first time. We propose the Double-Layer MoE RM (DMoERM). The outer layer MoE is a sparse model. After classifying an input into task categories, we route it to the corresponding inner layer task-specific model. The inner layer MoE is a dense model. We decompose the specific task into multiple capability dimensions and individually fine-tune a LoRA expert on each one. Their outputs are then synthesized by an MLP to compute the final rewards. To minimize costs, we call a public LLM API to obtain the capability preference labels. The validation on manually labeled datasets confirms that our model attains superior consistency with human preference and outstrips advanced generative approaches. Meanwhile, through BoN sampling and RL experiments, we demonstrate that our model outperforms state-of-the-art ensemble methods of RM and mitigates the overoptimization problem. Our code and dataset are available at: https://github.com/quanshr/DMoERM-v1.