LLM‑driven multi‑agent systems (MAS) solve complex tasks by coordinated reasoning, tool usage, and interaction with external resources. Yet attributing failures remains difficult because the observed outcome often does not reveal the exact error that caused the execution to fail. This work defines the attribution target as the decisive error—the agent‑step pair whose correction would recover the entire execution. Existing approaches mainly flag suspicious steps without explicitly modeling how errors propagate across interactions or persist in unresolved loops, making it hard to separate decisive errors from downstream symptoms. To address this, we propose Error‑Propagation Modeling for Failure Attribution (EMFA). EMFA first builds a structured representation of the failed trajectory, then models two propagation mechanisms simultaneously:
- Cascading propagation: the sequential path through which an error spreads across agents;
- Persistent loops: scenarios where an error repeatedly appears in unresolved interaction cycles. Leveraging this representation, EMFA performs propagation‑aware candidate screening to select potential decisive error pairs, followed by counterfactual verification (re‑executing the trace assuming the error is fixed) to confirm the true decisive error. Experiments on the Who&When benchmark show that EMFA achieves state‑of‑the‑art step‑level attribution accuracy while remaining competitive at the agent level. Compared with the previous best, EMFA improves step‑level accuracy by 3.45 % on the Hand‑Crafted subset and 4.40 % on the Algorithm‑Generated subset.
Review: By explicitly capturing both error propagation and looping behaviors, EMFA markedly improves failure localization in LLM‑based multi‑agent systems, offering a more reliable debugging tool for real‑world deployments.