Warn before the action, not after the failure: a GNN over historical trajectory graphs lifted pass ratio 14.69%

arXiv 2607.27443·medium signalXBlueskyLinkedInCopy link

The framework converts past agent trajectories into a probabilistic graph, trains a GNN to recognize action sequences that historically precede failure, and fires an early warning *before* the agent executes a flagged step so it can self-correct — a diagnostic pass rather than post-hoc repair. Average improvement was 14.69% pass ratio across four benchmarks and three LLM agents, with no fine-tuning of the underlying model. If you already log agent trajectories, this is a pattern you can build on data you're keeping anyway; it also complements the growing evidence that feeding failures back for self-repair is the weaker move.