English

Adversarial Training for Process Reward Models

Machine Learning 2025-12-01 v1 Artificial Intelligence

Abstract

Process Reward Models (PRMs) enhance reasoning ability of LLMs by providing step-level supervision. However, their widespread adoption is limited due to expensive manual step-level annotation and poor generalization of static training data to novel errors. We introduce Adversarially Trained PRMs (\texttt{APRM}), where a Generator (GG) learns to produce reasoning errors to deceive a PRM (RR), while RR concurrently learns to detect them. This interaction yields progressively harder negatives for RR, improving its robustness and generalization to novel errors without requiring manual step-level labels. Averaged across diverse mathematical reasoning benchmarks, \texttt{APRM} improves solver accuracy by +3.4+3.4 percentage points (pp) over the strongest PRM baseline. \texttt{APRM} achieves gains of +5.3+5.3 pp on out-of-distribution tasks.

Keywords

Cite

@article{arxiv.2511.22888,
  title  = {Adversarial Training for Process Reward Models},
  author = {Gurusha Juneja and Deepak Nathani and William Yang Wang},
  journal= {arXiv preprint arXiv:2511.22888},
  year   = {2025}
}
R2 v1 2026-07-01T07:58:48.836Z