CoPE adds one loss to GRPO, boosting language agent sample efficiency up to 2×

Lianhuiq · x · 2026-10-17

An arXiv paper proposes CoPE (Co-Evolving Acting and Forecasting for Language Agents), adding a single auxiliary loss to GRPO: on successful rollouts, the agent forecasts its own next K actions.

Why actions, not observations

Value targets derive from one scalar reward per trajectory; observation targets (web pages, terminal logs) are long, noisy, and only one step ahead
Action sequences are short and informative, and successful rollouts already record them — no simulator, critic, or annotation needed
Each action is supervised both when executed and when forecast from earlier histories, an implicit temporal ensemble (like action chunking in robotics) absorbed into the shared backbone during training

Results

Across ALFWorld, ScienceWorld, WebShop, τ²-bench, and AppWorld, CoPE beats GRPO and RL baselines with up to 2× sample efficiency
WebShop success: 64.5 → 77.0; ALFWorld: 90.5 → 95.5
At inference the agent still acts one step at a time, with zero overhead

Original post →

CoPE adds one loss to GRPO, boosting language agent sample efficiency up to 2×

More from coding & agent