Measuring how one agent's mistakes become another agent's mistakes, conditioned on whether the sender was right
AspireAI community · arXiv
LLM-based multi-agent systems introduce a reliability problem with a specific shape: actions taken by upstream agents affect the correctness of downstream agents. Prior work has investigated error propagation and methods for estimating inherited uncertainty, but has not explicitly modelled how realised sender-receiver interactions contribute to downstream error risk under different sender correctness states. This paper studies agent-level error-risk estimation from multi-agent execution traces and proposes Interaction-Effect Propagation, an interaction-conditioned model that learns separate effects for correct and incorrect upstream agents. For each sender-receiver interaction the model combines those conditional effects using the sender's estimated error risk, and recursively propagates the resulting estimates through the execution graph. Evaluated across three multi-agent architectures and three benchmarks, the approach raises final-node AUROC over the established baseline by approximately 0.05 points on average, and tracking risk across multiple agent steps contributes a distinct additional 0.03 AUROC gain in sequential systems. The results show that realised inter-agent interactions carry predictive information beyond receiver-local evidence, while the value of multi-step risk tracking depends on the system's information-flow structure. The limitation is that the method needs execution traces with identifiable sender-receiver structure, which constrains where it can be deployed and how much instrumentation an operator must add first.
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.