2025년 7월 8일

Pre-Trained Policy Discriminators are General Reward Models

(Shihan Dou, Shichun Liu, Yuming Yang, Yicheng Zou, Yunhua Zhou, Shuhao Xing, Chenhao Huang, Qiming Ge, Demin Song, Haijun Lv, Songyang Gao, Chengqi Lv, Enyu Zhou, Honglin Guo, Zhiheng Xi, Wenwei Zhang, Qipeng Guo, Qi Zhang, Xipeng Qiu, Xuanjing Huang, Tao Gui, Kai Chen)

We offer a novel perspective on reward modeling by formulating it as a policy discriminator, which quantifies the difference between two policies to generate a reward signal, guiding the training policy towards a target policy with desired behaviors. Based on this conceptual insight, we propose a scalable pre-training method named Policy Discriminative Learning (POLAR), which trains a reward model (RM) to discern identical policies and discriminate different ones. Unlike traditional reward modeling methods relying on absolute preferences, POLAR captures the relative difference between one policy and an arbitrary target policy, which is a scalable, high-level optimization objective suitable for modeling generic ranking relationships. Leveraging the POLAR pre-training paradigm, we present a series of RMs with parameter scales from 1.8B to 7B. Empirical results show that POLAR substantially outperforms traditional non-pre-trained methods, significantly enhancing RM performance. For instance, POLAR-7B could improve preference accuracy from 54.8% to 81.0% on STEM tasks and from 57.9% to 85.5% on creative writing tasks compared to SOTA baselines. POLAR also shows robust generalization capabilities in RLHF using Reinforcement Fine-tuning (RFT), providing reliable reward signals and markedly enhancing policy performance--improving LLaMa3.1-8B from an average of 47.36% to 56.33% and Qwen2.5-32B from 64.49% to 70.47% on 20 benchmarks. Moreover, scaling experiments reveal a clear power-law relationship between computation and performance, supported by linear correlation coefficients approaching 0.99. The impressive performance, strong generalization, and scaling properties suggest that POLAR is a promising direction for developing general and strong reward models.

Reward Model Pretraining. 다양한 Policy에서 샘플링한 궤적을 구분하는 형태로 학습했습니다. 이걸 Trillion 단위로 했네요. 이미 정렬된 모델도 Policy Pool에 포함되어 있다는 것을 고려했을 때 이 Pool의 구성이 어떤 영향을 미치는지 궁금하네요.

Reward model pretraining. They trained the model to distinguish between trajectories sampled from various policies. This was done on a trillion-token scale. Considering that already-aligned models are included in the policy pool, I'm curious about how the composition of this pool affects the results.

#reward-model #scaling-law