EDGE: Error Dependency Graph-Guided Multi-Error Attribution in Multi-Agent LLM Systems
EDGE constructs an error dependency graph from observed error events and validates a reliable causal subset through counterfactual rollout. The inference graph guides a two-stage LLM-as-judge detector for error attribution. Experiments on TRAIL and MAST show that EDGE improves category-level multi-error attribution across most evaluated models and settings. Experiments with adapted Who&When-style prompts show that the graph helps across prompting strategies.
Large language model (LLM) agent failures often contain multiple related errors rather than a single mistake. Existing attribution methods usually identify a responsible agent, step, or root cause, but do not explicitly model dependency between errors. EDGE introduces an Error Dependency Graph-guided multi-Error attribution framework. It constructs an error dependency graph from observed error events and validates a reliable causal subset through counterfactual rollout. The inference graph guides a two-stage LLM-as-judge detector for error attribution, and the intervention-validated subgraph provides a more reliable basis for explanation and repair analysis. Experiments on TRAIL and MAST show that EDGE improves category-level multi-error attribution across most evaluated models and settings. Experiments with adapted Who&When-style prompts show that the graph helps across prompting strategies. These results suggest that dependency structure is a useful diagnostic prior for agent failures beyond isolated root-cause prediction.
EDGE models error dependencies as a graph and uses counterfactual rollouts to validate causal links, then employs a two-stage LLM-as-judge detector guided by the validated subgraph. This shifts attribution from isolated root-cause prediction to structured multi-error causal reasoning, improving category-level attribution across TRAIL and MAST benchmarks.
The framework addresses a practical gap in multi-agent LLM system debugging: failures are rarely single-point. By providing dependency-aware attribution, EDGE could reduce manual investigation time and improve reliability engineering for agentic systems, though it remains a research artifact without reported production deployment.
For organizations operating multi-agent LLM systems, EDGE offers a potential method to improve error diagnosis and reduce downtime or debugging costs. However, no commercial product, pricing, or adoption data is provided, so business value is prospective and tied to future engineering adoption.
Next observable signals include open-source release of EDGE code or integration into agent evaluation harnesses, follow-up work applying dependency graphs to repair or intervention planning, and independent replication on additional multi-agent benchmarks beyond TRAIL and MAST.