A team of researchers from Penn State University and Duke University, in collaboration with Google DeepMind and other institutions, has introduced a groundbreaking framework to automatically identify which agent caused a failure in large language model (LLM) multi-agent systems. The work, accepted as a Spotlight presentation at the top-tier machine learning conference ICML 2025, tackles what they call 'Automated Failure Attribution.'
'Developers often spend hours combing through logs to find the source of a failure,' said Shaokun Zhang, co-first author and researcher at Penn State University. 'Our approach provides a systematic way to answer the critical question: which agent, at what point, was responsible?' The team also released the first benchmark dataset for this task, named Who&When, along with open-source code and evaluation methods.
Background
LLM-powered multi-agent systems have shown promise in solving complex problems collaboratively, but they are notoriously fragile. A single agent's error, a misunderstanding between agents, or a mistake in information transmission can derail the entire task.
'Currently, debugging is like finding a needle in a haystack,' explained Ming Yin, co-first author from Duke University. 'Developers manually review long interaction logs and rely heavily on intuition. This becomes impractical as systems grow more complex.' The researchers note that without automated attribution, system iteration and optimization grind to a halt.
Key Challenges Addressed
- Manual Log Archaeology: Developers must manually sift through thousands of lines of interaction logs to pinpoint issues.
- Reliance on Expertise: Effective debugging demands deep understanding of both the system and the task, making it inaccessible to many developers.
What This Means
The introduction of Automated Failure Attribution marks a paradigm shift. Instead of ad-hoc debugging, developers can now use data-driven methods to diagnose failures rapidly. 'This could significantly accelerate the development cycle of multi-agent systems,' said a spokesperson for Google DeepMind, an institution involved in the research.
The Who&When dataset provides a standardized benchmark for evaluating failure attribution methods, enabling fair comparisons and fostering innovation. The open-source release of code and data means that the broader AI community can build on this work immediately.
'We expect this to become a standard tool for anyone building reliable multi-agent systems,' Zhang added. 'The ability to automatically attribute failures will enhance trust and robustness in these systems, from customer service bots to complex simulation environments.'
The paper is available on arXiv, and the code and dataset are hosted on GitHub and Hugging Face respectively.