English

Generative Verifiers: Reward Modeling as Next-Token Prediction

Machine Learning 2025-02-25 v3

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 large performance gains with Best-of-N, namely 5% \rightarrow 45.3% on algorithmic tasks and 73% \rightarrow 93.4% on GSM8K. In easy-to-hard generalization settings, we observe improvements of 28% \rightarrow 44.6% on MATH, and 37.9% \rightarrow 53.5% on MMLU abstract algebra. Furthermore, we find that training GenRM with synthetic verification rationales is sufficient to pick out subtle errors on math problems. Finally, we demonstrate that GenRM scales favorably with model size and test-time compute.

Keywords

Cite

@article{arxiv.2408.15240,
  title  = {Generative Verifiers: Reward Modeling as Next-Token Prediction},
  author = {Lunjun Zhang and Arian Hosseini and Hritik Bansal and Mehran Kazemi and Aviral Kumar and Rishabh Agarwal},
  journal= {arXiv preprint arXiv:2408.15240},
  year   = {2025}
}

Comments

ICLR 2025

R2 v1 2026-06-28T18:25:43.897Z