CoPE adds one forecasting loss to GRPO, doubling long-horizon agent sample efficiency
Lianhuiq · x · 2026-10-17
Researchers introduce CoPE, a simple fix for sample-inefficient RL in long-horizon language agents: one extra loss on top of GRPO that forecasts the agent's next K actions on successful rollouts, encouraging action chunking. It achieves up to 2× sample efficiency across 5 long-horizon agent benchmarks with zero inference-time cost. Paper, code and project page are available.
Related event: UIUC's CoPE Boosts GRPO Sample Efficiency with Action Forecasting Loss(2 posts)→
More from coding & agent
- All technical work is becoming evals: engineers' job is building agent verification loops — rachittshah · 2026-10-17
- Cursor Cloud Agents can now build interactive canvases you can share as links — belce_dogru · 2026-10-17
- 20 free single-file HTML landing pages that let Codex reskin your whole app — Azra_Nysus · 2026-10-17
- You Can Generate 3D HTML Animations Directly in Runway's Agent Mode — CurieuxExplorer · 2026-10-17
- OpenAI Agents API weekly: 6-Astra Ultrafast, sandbox snapshots, idle-environment billing fix — OpenAIDevs · 2026-10-17
- Engineer uses Databricks Genie One for org discovery, not just SQL/BI — durlabha · 2026-10-17