The authors developed TrajEvolve for tasks in which a language-model agent interacts with an environment over many steps and receives reward mainly at completion.
Trajectory refinement improves reinforcement learning for long-horizon language-model agents
TrajEvolve revises unnecessary steps in historical trajectories and uses the resulting paths alongside fresh policy rollouts during training.
Chinese Tech
Yi Wen · Hao Chen · Wanyu Wang · Maolin Wang · Pengyue Jia · Derong Xu · +6 more
City University of Hong Kong · Tencent · Nanjing University
Research Digest··3 min read
Wen et al.
§