DCFA: Dual-view Causal-inspired Attribution for Failure Reasoning in LLM-based Multi-agent Systems

To address the issue of failure in reasoning and coordination in multi-agent systems for large language models (LLMs), researchers proposed a training-free fault attribution framework called DCFA. This framework constructs a causal heuristic dependency graph using global modules to identify decisive errors and uses local modules for anti-facto heuristic reasoning optimization. In the Who&When benchmark, DCFA increased the step-level accuracy of six LLMs by 8.27% compared to state-of-the-art baselines. Existing attribution methods face two major challenges: first, shallow attribution only captures minor deviations and misses decisive causes; second, the growth of system traces leads to context degradation, resulting in a decline in model reasoning capabilities.