DUOTRACE filters multi-agent traces before attribution and lifts step-level blame accuracy 7% across six baselines

arXiv·medium signalXBlueskyLinkedInCopy link

LLM-based failure attribution degrades on long trajectories while topology and spectrum methods miss fine-grained semantics. DUOTRACE inserts a detect-before-attribute stage: a VAE anomaly detector over dual-view semantic-structural node representations with a Tree-LSTM trajectory encoder, plus prefix-chain and LLM data augmentation to cope with scarce failure data, then hands only the focused evidence to the downstream attributor. Tested against six attribution baselines it improves agent-level accuracy 8.7% and step-level 7.0%, and it is plug-and-play rather than a replacement attributor.