ASCon does failure attribution in multi-agent systems as one model instead of three, lifting failure-mode detection 14.73 points

arXiv·medium signalXBlueskyLinkedInCopy link

Submitted August 11, ASCon tackles the debugging problem every multi-agent operator hits: when a run fails, which agent, which step, and what kind of failure? Rather than training a separate model per attribution target, it builds contextualized representations from system trajectories using direction-aware graph attention, masked step-to-agent attention and agent-conditioned step contextualization, then attaches specialized heads. Reported gains over prior work are +5.83% micro-accuracy on faulty-agent detection, +10.63% on faulty-step detection, and +14.73 Macro-F1 on failure-mode detection, with improved out-of-domain performance when paired with LLM-based attribution. Directly relevant to anyone running orchestration in production without a way to localize blame in a trace.