Skip to content
arXiv· arXiv:2602.10418· Weichen Yu, Ravi Mangal, Yinyi Luo, Kai Hu, Jingxuan He, Corina S. Pasareanu, Matt Fredrikson·Indexed by this site · Originally published

SecCodePRM: Process Reward Model for Code Security

SecCodePRM: A Process Reward Model for Code Security

AI summary

SecCodePRM is a process reward model for code security that provides context-aware step-level security scores along code trajectories, outperforming existing methods on full-code and partial-code vulnerability detection and secure code generation. The model is trained with step-level supervision from static analyzers and expert annotations; vulnerability detection uses risk-sensitive aggregation to highlight high-risk steps, and code generation achieves inference-time scaling by ranking candidate continuations. In the experiments, code functional correctness is maintained throughout, with no security-utility trade-off observed.

Read originalICML 2026