Process-reward optimization advances computer use agents

Arxiv CS.LG (machine learning)

Two papers propose process-reward optimization for training computer use agents (CUAs), addressing limitations of sparse reward and costly live environment interaction. Methods like filtered behavior cloning and multi-granularity reward models improve agent performance on complex digital workflows.

process-reward

See 1 more source

PRO-CUA: Process-Reward Optimization for Computer Use Agents20 days agoYifei He, Rui Yang, Hao Bai, Tong Zhang, Han Zhao

Arxiv CS.LG (machine learning)·Jaehun JungXiming LuBrandon CuiMuhammad KhalifaShaokun ZhangHao ZhangJin XuAmala Sanjay DeshmukhKaran SapraAndrew TaoYejin ChoiJan KautzMingjie LiuYi Dong

1 day ago