Inference-Time Scaling for Generalist Reward Modeling

2025-04-03

AI Summary

To address the poor scalability and weak cross-domain generalization of generic reward models (RMs) during inference, this paper proposes Self-Principled Critique Tuning (SPCT), an inference-time computation-augmented reward modeling paradigm. SPCT integrates generative pairwise reward modeling (GRM), adaptive principle-driven critical feedback generation, parallel multi-path sampling, and a meta-reward model ensemble voting mechanism—achieving substantial gains in inference quality and robustness without increasing training overhead. Experiments demonstrate that SPCT consistently outperforms existing training-time scaling methods across multiple benchmarks, offering superior inference efficiency, reduced bias, and strong domain adaptability. The open-sourced DeepSeek-GRM model series empirically validates SPCT’s effectiveness and practical utility.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search
• Natural Language Processing: Safety and Robustness

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

Reinforcement learning (RL) has been widely adopted in post-training for large language models (LLMs) at scale. Recently, the incentivization of reasoning capabilities in LLMs from RL indicates that $ extit{proper learning methods could enable effective inference-time scalability}$. A key challenge of RL is to obtain accurate reward signals for LLMs in various domains beyond verifiable questions or artificial rules. In this work, we investigate how to improve reward modeling (RM) with more inference compute for general queries, i.e. the $ extbf{inference-time scalability of generalist RM}$, and further, how to improve the effectiveness of performance-compute scaling with proper learning methods. For the RM approach, we adopt pointwise generative reward modeling (GRM) to enable flexibility for different input types and potential for inference-time scaling. For the learning method, we propose Self-Principled Critique Tuning (SPCT) to foster scalable reward generation behaviors in GRMs through online RL, to generate principles adaptively and critiques accurately, resulting in $ extbf{DeepSeek-GRM}$ models. Furthermore, for effective inference-time scaling, we use parallel sampling to expand compute usage, and introduce a meta RM to guide voting process for better scaling performance. Empirically, we show that SPCT significantly improves the quality and scalability of GRMs, outperforming existing methods and models in various RM benchmarks without severe biases, and could achieve better performance compared to training-time scaling. DeepSeek-GRM still meets challenges in some tasks, which we believe can be addressed by future efforts in generalist reward systems. The models will be released and open-sourced.

Problem

• Improving reward modeling for general queries using inference compute
• Enhancing performance-compute scaling with effective learning methods
• Addressing challenges in accurate reward signals for diverse domains

Innovation

• Pointwise generative reward modeling for flexibility
• Self-Principled Critique Tuning for scalable rewards
• Parallel sampling and meta RM for better scaling