Automatic Blame Attribution in LLM-Based Multi-Agent Systems: New Research

Researchers from Pennsylvania State University and Duke University, in collaboration with Google DeepMind, University of Washington, Meta, Nanyang Technological University, and Oregon State University, have for the first time formalized the task of automatic fault attribution in multi-agent systems based on large language models (LLMs). They created the first benchmark, Who&When, which includes 127 fault logs collected from LLM-agent systems, both algorithmically generated and manually composed by experts. Each log contains annotations: the responsible agent (Who), the error step (When), and a description of the cause (Why). Three methods are proposed for the task: All-at-Once (simultaneous analysis of the entire log), Step-by-Step (stepwise review), and Binary Search (binary search over the log). Experiments with GPT-4o showed that the best single method achieves 53.5% accuracy in identifying the responsible agent and only 14.2% in exactly pinpointing the error step. No method outperforms random guessing on many metrics. Even advanced models like o1 and DeepSeek R1 struggle. Hybrid approaches improve results but significantly increase computational costs. As context length increases, performance of all methods decreases, especially in determining the error step. The work was accepted as a Spotlight at ICML 2025; code and data are open-sourced.