Coming Soon: CoPE — Co-Evolving Acting and Forecasting for Language Agents
Published:
A blog post on CoPE is on the way. In the meantime, check out the paper and the code.
CoPE adds multi-step action targets to agentic RL: on successful trajectories, the policy also learns to forecast the actions that follow each history, so acting and forecasting co-evolve within a single backbone, with no extra models and no inference overhead.
The upcoming post will walk through the motivation, the method, and results on ALFWorld, ScienceWorld, WebShop, τ²-bench, and AppWorld. Stay tuned!