Breaking: Researchers Unveil Automated Failure Attribution for LLM Multi-Agent Systems
A team of researchers from Penn State University and Duke University, in collaboration with Google DeepMind, the University of Washington, Meta, Nanyang Technological University, and Oregon State University, has announced a breakthrough in debugging complex LLM-based multi-agent systems. Their work introduces a new research problem they call "Automated Failure Attribution" and presents the first benchmark dataset, named Who&When, to identify which agent caused a task failure and at what point the failure occurred. The paper has been accepted as a Spotlight presentation at the top-tier machine learning conference ICML 2025, and the code and dataset are fully open-source.
Multi-agent systems powered by large language models (LLMs) have shown remarkable promise in tackling complex tasks through collaboration. However, when these systems fail—and they often do—developers face a daunting challenge: sifting through massive interaction logs to pinpoint the root cause. This manual process is time-consuming, expertise-dependent, and inefficient. The new research aims to automate this diagnosis.
Background
LLM-based multi-agent systems are increasingly used in areas like software development, scientific reasoning, and customer support. In these systems, multiple agents communicate and coordinate to solve problems. Failures can occur due to a single agent's error, miscommunication between agents, or mistakes in information transmission. Identifying the precise source of failure is crucial for system iteration and optimization, yet it has remained a manual, labor-intensive process—often described by developers as "finding a needle in a haystack."
The need for automated attribution is urgent, as the complexity and autonomy of these systems grow. Without a systematic method, debugging stalls development and limits reliability. The Who&When dataset fills this gap by providing a standardized benchmark to evaluate automated attribution methods.
What This Means
This research paves the way for faster, more reliable debugging of multi-agent systems. Developers can now use automated attribution methods to quickly identify failing agents and the exact timing of failures, reducing downtime and accelerating system improvements. The open-source availability of Who&When and the associated code enables the broader AI community to build on this work.
Ming Yin, co-first author and researcher at Duke University, said: "Our work transforms failure diagnosis from a manual detective hunt into an automated, measurable process. This is a critical step toward building more trustworthy multi-agent systems." Shaokun Zhang, co-first author from Penn State, added: "Without knowing 'who' caused a failure and 'when', developers are stuck in a cycle of guesswork. Who&When gives them a clear starting point."
The implications extend beyond debugging. Automated failure attribution can inform system design, improve agent training, and enhance overall system robustness. As multi-agent systems scale, this capability becomes essential for deploying them in high-stakes environments like autonomous driving, healthcare, and finance.
Technical Details and Availability
The researchers developed multiple automated attribution methods and evaluated them on the Who&When benchmark. They compared performance across different failure types and system configurations. Preliminary results show that their methods can accurately identify the responsible agent and failure point in a majority of cases, though they acknowledge the task remains challenging.
For those interested in exploring the work further:
- Paper: https://arxiv.org/pdf/2505.00212
- Code: https://github.com/mingyin1/Agents_Failure_Attribution
- Dataset: https://huggingface.co/datasets/Kevin355/Who_and_When
The dataset includes diverse multi-agent scenarios with annotated failure attribution labels, enabling reproducible research. The team encourages the community to contribute new attribution methods and extend the benchmark to more complex systems.