Closed-Loop Multi-Round Planning for Large Language Model Agents via Self-Reflection and Error Correction

Abstract

To address the instability and error cascading problems in multi-round interaction planning for large language model agents, this paper proposes a multi-round planning method with self-reflection and error correction mechanisms. This method integrates task objective and constraint modeling, state updating, candidate plan generation, plan selection and scoring, risk assessment, reflection triggering, and local patching into a closed-loop process. After each round, the method performs consistency verification on intermediate states and constraint satisfaction based on observable feedback. When potential deviations are detected, a reflection module is triggered to attribute errors and locate patching points. An error correction module then minimizes key steps to restore feasible trajectories. Simultaneously, a unified scoring function is used to select from candidate plans, balancing task completion, tool availability, and constraint compliance, thereby reducing invalid rounds and violation risks. To facilitate process interpretation and diagnosis, this paper further provides visualization analysis methods, such as trajectory alignment and failure type distribution, to characterize the structural patterns of error triggering, reflection, and patching at different stages. Comparative experiments show that the proposed method performs better in terms of task success, tool invocation stability, and constraint violation control. It is also more efficient and stable in terms of multi-round interaction costs, verifying the effective role of closed-loop error correction in improving the reliability and controllability of planning.