Enriching Scarce Root-cause Supervision for Failure Attribution via Source-conditioned Supervision Refinement

Failure attribution in LLM-based multi-agent systems aims to identify the decisive root-cause step within a long execution trajectory, yet each failed trajectory typically contains only one such step among many non-root-cause ones, resulting in highly imbalanced supervision. Latent-space synthesis provides a natural way to expand supervision around scarce root-cause examples by constructing source-conditioned latent neighborhoods, but naive expansion introduces two challenges: source-level training influence can grow with synthesis multiplicity, while individual synthetic samples may differ substantially in positive-support compatibility and redundancy. We introduce Source-conditioned Supervision Refinement (SSR), which fixes the total synthetic supervision weight contributed by each source and softly adjusts sample-level supervision according to positive-support compatibility and source-specific non-redundancy. This preserves broad local coverage while emphasizing better-supported and less redundant synthetic samples. SSR operates only during training; inference uses deterministic latent encoding and a lightweight attribution detector without additional generation, sampling, or trajectory replay. Experiments on three failure-attribution benchmarks show that SSR consistently outperforms matched representation and synthesis baselines.