Generative Verifiers: Reward Modeling as Next-Token Prediction

Lunjun Zhang, Arian Hosseini, Hritik Bansal, Mehran Kazemi, Aviral Kumar, Rishabh Agarwal

Keywords: LLM reasoning, reward models, verifiers

TL;DR: We pose the problem of reward modeling as next-token prediction, allowing for Chain-of-Thought and Majority Voting in reward models

Abstract: Verifiers or reward models are often used to enhance the reasoning performance of large language models (LLMs). A common approach is the Best-of-N method, where N candidate solutions generated by the LLM are ranked by a verifier, and the best one is selected. While LLM-based verifiers are typically trained as discriminative classifiers to score solutions, they do not utilize the text generation capabilities of pretrained LLMs. To overcome this limitation, we instead propose training verifiers using the ubiquitous next-token prediction objective, jointly on verification and solution generation. Compared to standard verifiers, such generative verifiers (GenRM) can benefit from several advantages of LLMs: they integrate seamlessly with instruction tuning, enable chain-of-thought reasoning, and can utilize additional test-time compute via majority voting for better verification. We demonstrate that GenRM outperforms discriminative, DPO verifiers, and LLM-as-a-Judge, resulting in a 16-40% improvement in the number of problems solved with Best-of-N on algorithmic and math reasoning tasks. Furthermore, we find that training GenRM with synthetic verification rationales is sufficient to pick out subtle errors on math problems. Finally, we demonstrate that generative verifiers scale favorably with model size and inference-time compute.

Concurrent Submissions: Workshop Submission (System-2 Reasoning at Scale) at NeurIPS'24

Submission Number: 85