Original title: EDGE: Error Dependency Graph-Guided Multi-Error Attribution in Multi-Agent LLM Systems
Jun Hou, Priya Pitre, Yi Fang, Xuan Wang · September 2026
When groups of AI programs, called large language model agents, work together, they often fail because of several linked mistakes rather than just one slip-up. Older tools usually point out one responsible agentAn autonomous AI program that can perform tasks and make decisions., step, or root cause, but they miss how mistakes trigger each other. To fix this, researchers built EDGE. EDGE works by drawing a map, called an error dependency graphA diagram showing how different errors in a system are connected to each other., that links observed mistakes together. It then tests these links using "counterfactual rolloutA testing method that simulates what would happen if certain errors were removed or altered.," checking what happens if a step changes. Next, an AI judge uses this map in two stages to diagnose what went wrong. The researchers found that EDGE improved multi-error attribution across most evaluated models, settings, and prompting styles, proving that tracking error connections diagnoses failures better.
Imagine preparing a feast with a team of chefs. If one chef uses too much salt in a sauce, it might ruin multiple dishes that rely on that sauce. Finding just the salty sauce isn't enough; you need to trace how that mistake affects other ingredients and dishes. In this paper, EDGE works like a system that maps how cooking errors propagate across chefs and courses, then tests fixes by simulating adjustments. This helps pinpoint the exact chain of mistakes so the entire meal can be corrected.
This matters for anyone building AI systems where multiple agents work together. When failures happen, understanding the chain of errors allows for quicker fixes and prevents the same issues from recurring.
The abstract doesn't specify limitations, but the method is evaluated only on the TRAIL and MAST benchmarks, so its performance on other multi-agent systems isn't confirmed. It also relies on LLM-as-judge, which could introduce biases.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
EDGE: Error Dependency Graph-Guided Multi-Error Attribution in Multi-Agent LLM Systems — Jun Hou, Priya Pitre, Yi Fang et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.