Verifying failure hypotheses against whole interaction trajectories beat predicting which agent erred, and improved both open-weight and API-based backbones
AspireAI community · arXiv
LLM-driven multi-agent systems excel at complex tasks, but unreliable agents remain a key bottleneck to system-level reliability, which makes automatic failure attribution critical. Existing approaches, such as directly predicting agent-error pairs or agent-first attribution, rely on local agent logs and miss global failures that only manifest over full interaction trajectories, such as cross-step inconsistencies and inter-agent coordination errors, while directly predicting failures induces a large combinatorial search space that hinders fine-grained attribution. VerifyMAS formulates and verifies failure hypotheses against full trajectories instead of predicting faulty agents and error types directly, decomposing attribution into trajectory-level error validation and fine-grained agent localization. The authors add a hypothesis-based data construction strategy grounded in a structured error taxonomy and fine-tune a specialized LLM verifier for trajectory-level failure verification and agent attribution. On Aegis-Bench and Who&When, VerifyMAS consistently improves diverse backbone models including open-source Qwen and API-based GPT models, outperforming prior methods without sacrificing inference efficiency for long multi-agent trajectories. The limitation is that attribution accuracy is only as good as the error taxonomy behind it, so a failure mode the taxonomy does not name will not be found by the verifier.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.