A team of researchers from Penn State University and Duke University, in collaboration with Google DeepMind, University of Washington, Meta, Nanyang Technological University, and Oregon State University, has introduced a groundbreaking approach to automatically identify the specific agent and moment responsible for a failure in large language model (LLM)-driven multi-agent systems. The work, accepted as a Spotlight presentation at ICML 2025, defines the novel problem of Automated Failure Attribution and provides the first benchmark dataset, Who&When, along with multiple automated attribution methods. The code and dataset are now fully open-source and available on Hugging Face. Read about the research background below and what this means for AI development.
“Debugging multi-agent systems has been a manual, time-consuming process. Our method turns that into an automated, systematic task—like moving from searching for a needle in a haystack to using a metal detector,” said Shaokun Zhang, co-first author and researcher at Penn State University. “This is a critical step toward building more reliable and transparent AI systems.”
The research addresses a pressing challenge: as LLM multi-agent systems grow in complexity and adoption, failures become both more frequent and harder to diagnose. Current debugging relies on manual log reviews and deep expert knowledge—a “needle in a haystack” problem that slows iteration and deployment. The team’s automated attribution methods can pinpoint the failing agent and the exact interaction step, dramatically cutting debugging time.
Background
LLM-driven multi-agent systems coordinate multiple AI agents to solve complex tasks collaboratively. Despite their promise, these systems are fragile: a single agent’s error, a misunderstanding between agents, or a broken information chain can cause the entire task to fail. Developers often fall back on manual “log archaeology” and deep system expertise to locate the root cause.
To formalize and tackle this issue, the researchers introduced Automated Failure Attribution as a new research problem. They built the Who&When dataset—a collection of multi-agent failure scenarios with ground-truth labels indicating which agent failed and at which step. Several automated methods were then developed and evaluated, achieving promising accuracy on the benchmark.
“The dataset and methods we release allow the community to measure progress on this crucial problem,” said Ming Yin, co-first author from Duke University. “We hope this opens a new line of research toward self-diagnosing and self-healing multi-agent systems.”
What This Means
Automated failure attribution could accelerate the development cycle for multi-agent AI applications—from research prototypes to production systems in robotics, software engineering, and autonomous decision-making. By quickly identifying failure sources, developers can iterate faster and improve system reliability.
This work also lays the foundation for more transparent AI. Understanding exactly where and why a failure occurs is a step toward explainability and trust, especially in safety-critical domains. The open-source release invites further innovation from academia and industry alike.
The Spotlight acceptance at ICML 2025 underscores the significance of the contribution. With the code and dataset freely available, the researchers hope to spark a new field of study centered on automated failure detection and correction in multi-agent systems.
Institutions involved: Penn State University, Duke University, Google DeepMind, University of Washington, Meta, Nanyang Technological University, Oregon State University.