Co-Evolving Acting and Forecasting for Language Agents
Forecasting future actions during training improved agent performance without adding inference overhead, the authors report.
The method trains a single policy backbone on successful trajectories, teaching it to predict action sequences from earlier interaction histories. This gives intermediate steps additional supervision beyond the final reward. Across five benchmarks, the authors report stronger aggregate results in simulated environments and improved overall performance on two realistic applications. On ScienceWorld, sample efficiency reached up to 1.9 times that of ECHO and 2.0 times that of GRPO.
ArXiv · AI/CL/LG's note
The method trains a single policy backbone on successful trajectories, teaching it to predict action sequences from earlier interaction histories. This gives intermediate steps additional supervision beyond the final reward. Across five benchmarks, the authors report stronger aggregate results in simulated environments and improved overall performance on two realistic applications. On ScienceWorld, sample efficiency reached up to 1.9 times that of ECHO and 2.0 times that of GRPO.
ArXiv · AI/CL/LG's note
score 5