Co-Evolving Acting and Forecasting for Language Agents
arXiv cs.AI · 15 Oct 2026, 03:59 AEST
A new method called CoPE enhances the training of language agents by introducing multi-step action targets. This approach allows for improved performance and sample efficiency in various simulated environments.
Key facts
- Training language agents is bottlenecked by the high cost of long-horizon rollouts, yet standard agentic reinforcement learning (RL) learns only from a terminal scalar reward, leaving the supervision in intermediate steps largely unused.
- We propose CoPE, which adds a third kind of multi-step action targets.
- Across five benchmarks (ALFWorld, ScienceWorld, WebShop, τ²-bench, and AppWorld), CoPE achieves stronger aggregate performance than GRPO and RL baselines with predictive auxiliary targets across simulated environments.
Read aloud
A new method called CoPE enhances the training of language agents by introducing multi-step action targets. This approach allows for improved performance and sample efficiency in various simulated environments. Training language agents is bottlenecked by the high cost of long-horizon rollouts, yet standard agentic reinforcement learning RL learns only from a terminal scalar reward, leaving the supervision in intermediate steps largely unused. We propose CoPE, which adds a third kind of multi-step action targets. Across five benchmarks ALFWorld, ScienceWorld, WebShop, τ²-bench, and AppWorld, CoPE achieves stronger aggregate performance than GRPO and RL baselines with predictive auxiliary targets across simulated environments.
Sources in this cluster
- Co-Evolving Acting and Forecasting for Language Agents
arXiv cs.AI · 15 Oct 2026, 03:59 AEST