Research map / Multi-agent systems and games / LLM agents
CoPE: A Framework for Optimizing Coordination between Planning and Execution in LLM-based Agents
Huanxi Liu, Kun Hu, Qiang Wang, Yuanzhao Zhai, Feng Dawei, Bo Ding, Huaimin Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CoPE, a framework for fine-tuning LLM agents that coordinates long-term planning with multi-step execution to avoid impractical plans and plan-deviated trajectories.
Problem
Fine-tuning LLMs as autonomous agents on domain-specific data overlooks the coordination between long-term planning and multi-step execution, causing impractical plans and plan-deviated trajectories to propagate.
Approach
A framework that explicitly optimizes the coordination between planning and execution during agent fine-tuning, rather than treating them independently; details beyond the abstract are not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add coordination weighting to planning DPO (ablation) | CoPE planning agent · AlfWorld Seen / ScienceWorld Seen | average reward: 84.47 -> 88.00 and 61.22 -> 64.67 | Improved |
| use coordination label instead of reward-based weighting (ablation) | CoPE planning agent · AlfWorld Seen / ScienceWorld Seen | average reward: 82.60 -> 88.00 and 58.14 -> 64.67 | Improved |
| use coordination-weighted DPO instead of IPO-style loss (ablation) | CoPE planning agent · AlfWorld Seen / ScienceWorld Seen | average reward: 80.93 -> 88.00 and 58.28 -> 64.67 | Improved |
| add SFT loss to execution KTO (ablation) | CoPE execution agent · AlfWorld Seen / ScienceWorld Seen | average reward: 81.81 -> 88.00 and 51.63 -> 64.67 | Improved |
| add execution action pruning (ablation) | CoPE execution agent · AlfWorld Seen / ScienceWorld Seen | average reward: 76.65 -> 88.00 and 55.20 -> 64.67 | Improved |
| use coordination-label SFT weight instead of constant weight (ablation) | CoPE execution agent · AlfWorld Seen / ScienceWorld Seen | average reward: 72.49 -> 88.00 and 45.11 -> 64.67 | Improved |
| add replanning in MCTS node expansion (ablation) | CoPE data collection · AlfWorld | retention rate: 53.69% -> 57.43% | Improved |
| add replanning in MCTS node expansion (ablation) | CoPE data collection · AlfWorld | Self-BLEU of plans (lower more diverse): 0.48 -> 0.25 | Improved |
| add coordination modeling to preference optimization over iterations (ablation) | CoPE · AlfWorld/ScienceWorld Seen | average reward | Improved |
| use coordination-aware data instead of task-reward filtering | CoPE vs RFT · ScienceWorld | invalid action rate: 10.78% -> 2.67% | Improved |
| raise trajectory filter threshold beyond optimum (ablation) | CoPE · AlfWorld | average reward: peak at 0.8 then declines | Hurt |
| train baselines on same data as CoPE | RFT/ETO/MPO/AGENTEVOL · AlfWorld Seen | average reward: CoPE 88.00 vs AGENTEVOL 82.74 | No clear effect |
Closest papers by meaning
- SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LLM-Based Agents (NeurIPS 2025 · similarity 0.90)
- AgentSquare: Automatic LLM Agent Search in Modular Design Space (ICLR 2025 · similarity 0.90)
- Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks (ICML 2025 · similarity 0.90)
- AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.89)
- $\texttt{Multi}^2$: Hierarchical Multi-Agent Decision-Making with LLM-Based Agents in Interactive Environments (ICML 2026 · similarity 0.89)
- FT-Dojo: Towards Autonomous LLM Fine-Tuning with Language Agents (ICML 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
