New Method Repairs AI Planning Graphs by Targeting Root Errors.

As AI agents tackle increasingly complex, multi-step workflows, failures within their extensive planning graphs become inevitable. Traditional methods of replanning the entire graph after each error are computationally inefficient and can overwhelm the LLM with irrelevant information. This research proposes WM-SAR, a new approach to correct these planning failures in place. Instead of merely scanning for visible symptoms, WM-SAR works backward from error amplification, pinpointing the specific nodes and edges that repeatedly cause issues. By focusing the LLM's attention on this compact, causal subgraph, the system can achieve near-whole-graph stabilization with significantly fewer tokens. Experiments demonstrate that WM-SAR substantially outperforms common engineering correctors, especially when operating within realistic token constraints, by providing the LLM with a cleaner, more targeted repair objective.