In recent years, multi‑agent systems built on large language models (LLMs) have expanded rapidly, yet they remain fragile, often producing reasoning and coordination errors that cause system‑level failures. Existing failure attribution relies on tracing natural‑language interactions among agents to locate the decisive error—the earliest action whose correction would reverse the failure.
Two major challenges arise:
- Shallow attribution – current methods tend to capture minor deviations such as incomplete retrievals or formatting glitches, which verification mechanisms can fix, while missing the true cause of the collapse.
- Contextual degradation – as the length of system traces grows, the model’s reasoning ability deteriorates sharply.
To address these issues, we introduce DCFA, a training‑free framework for failure attribution. DCFA consists of:
- Global module: builds a structured causal‑inspired dependency graph from system traces to identify the initial decisive error.
- Local module: applies counterfactual‑inspired reasoning locally to refine the global causal attribution.
Experiments on the Who&When benchmark across six mainstream LLMs show that DCFA improves step‑level accuracy by up to 8.27% over state‑of‑the‑art baselines.
Review