arXiv 2610.20597 · cs.AI · 2026-10-14
Co-Evolving Acting and Forecasting for Language Agents
Haofei Yu, Yixing Chen, Yining Zhao, Xuyan Ye, Yuxiang Liu, Zhigang Hua, Yan Xie, Shuang Yang 等 9 人
机构待核实
一句话要点
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。
CoPE enables co-evolution of acting and forecasting via multi-step action targets, improving performance and sample efficiency without extra models or overhead.
要点
- 语言代理训练受限于长时程轨迹的高成本,传统方法仅利用终端奖励,忽略中间步骤的监督信息。
- CoPE引入多步动作目标,在成功轨迹中训练策略生成后续动作序列,并通过历史预测形成时间上的联合学习。
- 在ALFWorld、ScienceWorld等五个基准上,CoPE在整体性能和真实场景应用中优于GRPO与RL基线,样本效率提升达1.9倍至2.0倍。
- 方法仅使用单一策略主干,无需额外模型或预训练,推理时仍保持一步一动作的高效结构。
English overview
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.
只根据 arXiv 官方摘要,由开源模型整理,经数字/专名自动核对(必须在摘要中出现),以论文原文为准。
阅读 arXiv 原文 PDF Hugging Face 讨论 代码仓库 项目页
主题与学科
同主题:强化学习
- Waypoint-OPD: Transferring RL-Induced Policy Shifts via Gated On-Policy Distillation ▲214
- Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight ▲189
- LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches ▲152
- Sharpening Tax in Post-Training ▲111
- RULER: Instance-aware Rubric Rewards for SVG Generation ▲104
同主题:智能体
- Raven: The Harness of Harnesses for Composable Agentic Intelligence 2026-09-27 · ▲748
- False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents 2026-09-30 · ▲642
- Omni-IO Skills: Harnessing Your Agent Omni-Native 2026-09-25 · ▲485
- BizAgentBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business 2026-10-14 · ▲388
- Any Time, Any Space: A World-State-Centric Framework for Agentic Video Generation 2026-10-14 · ▲364
数据说明
元数据来自 arXiv API(已核实);点赞来自 Hugging Face 每日论文;引用来自 Semantic Scholar。最近刷新 2026-10-14。
要点由本站根据摘要自写,不转载摘要原文,请阅读原文。