Co-Evolving Acting and Forecasting for Language Agents

One-line takeaway

CoPE enables co-evolution of acting and forecasting via multi-step action targets, improving performance and sample efficiency without extra models or overhead.

提出CoPE方法，通过多步动作目标协同训练语言代理的决策与预测能力，在多个环境中实现性能提升和样本效率显著提高，且无需额外模型或推理开销。

Key points

CoPE introduces multi-step action targets to enable co-evolution between acting and forecasting in language agents. By training the policy to generate future actions from history and simultaneously predicting those actions, it leverages intermediate supervision effectively. This approach improves performance across multiple environments while achieving high sample efficiency without additional models or inference overhead.

• CoPE introduces multi-step action targets for joint acting and forecasting training.
• It enables effective use of intermediate trajectory supervision through temporal ensembles.
• Outperforms GRPO and RL baselines in performance across five benchmarks.
• Achieves up to 2x sample efficiency gains on ScienceWorld with a single policy backbone.

Written only from the official arXiv abstract by an open-source model, then checked automatically (every number and name must appear in the abstract). The paper is authoritative.