arXiv cs.MA· Tuo Zhang, Alin-Ionut Popa, Yan Xu, Rui Song, Dimitrios Dimitriadis·· 2 天前AI 评分34
PIVOT：通过轨迹精修打通 LLM 智能体的规划与执行
PIVOT: Bridging Planning and Execution in LLM Agents via Trajectory Refinement
AI 导读
PIVOT 提出一种自监督框架，将智能体轨迹视为可优化对象，通过 PLAN、INSPECT、EVOLVE、VERIFY 四阶段迭代精修，缓解 LLM 智能体规划与执行脱节的问题。在 DeepPlanning 和 GAIA 上，带人类反馈的版本约束满足率相对提升最高达 94%，全自主版本也保持显著增益。该方法所需 token 比同类精修方法少 3x 至 5x。
正文 · 原文
Abstract:Large language model (LLM)-based agents frequently generate seemingly coherent plans that fail upon execution due to infeasible actions, constraint violations, and compounding errors over extended horizons. PIVOT (Plan-Inspect-eVOlve Trajectories) addresses this plan-execution misalignment through a self-supervised framework that treats trajectories as optimizable objects iteratively refined via environment interaction. The framework comprises four stages: PLAN generates candidate trajectories; INSPECT executes them and computes structured losses with textual gradients encoding plan-execution discrepancies; EVOLVE applies these signals to produce improved trajectories; and VERIFY performs a final global check against task constraints. A monotonic acceptance process ensures a non-decreasing solution quality. Empirical evaluations on DeepPlanning and GAIA demonstrate state-of-the-art performance: with human-in-the-loop (HITL) feedback, PIVOT establishes a strong upper bound up to 94% relative improvement in constraint satisfaction, while its fully autonomous variant retains substantial gains, showing that the core trajectory-refinement mechanism remains effective without external supervision. At the same time, PIVOT remains computationally efficient, requiring up to 3x to 5x fewer tokens than competing refinement methods. These findings establish that (self- or human-supervised) feedback-based trajectory optimization is a principled methodology for mitigating plan-execution gaps in autonomous agent systems.
| Subjects: | Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Multiagent Systems (cs.MA) |
| Cite as: | arXiv:2605.11225 [cs.AI] |
| (or arXiv:2605.11225v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2605.11225 arXiv-issued DOI via DataCite |
Submission history
From: Dimitrios Dimitriadis [view email]
[v1]
Mon, 11 May 2026 20:43:20 UTC (4,250 KB)
[v2]
Tue, 27 Oct 2026 04:50:15 UTC (4,180 KB)
来源：arXiv cs.MA · arxiv.org