arxivcs.LG2026-06-26
The Weakest Link Tells It All: Outcome-Supervised Process Reward Modeling via Learnable Credit Assignment
Tianyu Jia, Yue Fang, Hongxin Ding, Rihong Qiu, Zhibang Yang, Zhijing Wu, et al.
Process reward models (PRMs) enhance the reasoning capabilities of large language models (LLMs) by providing fine-grained feedback, yet training PRMs typically requires expensive stepwise annotations. Outcome-supervised PRMs offer a scalable alternative by learning from final-ans…