Language model teams as distributed systems
Large language models (LLMs) are growing increasingly powerful. Yet, fundamental constraints on the reasoning, memory, and performance of individual models has sparked recent interest in scaling abilities using LLM teams. With growing deployment of LLM teams at scale, there is a need for principled guidelines for when teams help, how many agents to use, and how structure impacts performance. Rather than building these teams through trial-and-error, we propose using distributed systems as a formal foundation for designing and evaluating LLM teams. Under this formal framework, we derive predictions about how task structure and team architecture should shape performance. Empirical evidence supports these predictions: LLM team scaling follows classical distributed computing laws based on the structure of a task, and centralized and decentralized architectures exhibit the asymmetric vulnerabilities predicted by distributed systems theory. These results establish distributed computing as a generative framework that yields testable, quantitative predictions about LLM team behavior and informs the construction of more efficient and capable teams.
Authors
- Ilia Sucholutsky (ORCID: https://orcid.org/0000-0003-4121-7479)
- Natalia Vélez (ORCID: https://orcid.org/0000-0002-6939-3282)
- Elizabeth Mieczkowski (ORCID: https://orcid.org/0009-0009-4879-5678)
- Thomas L. Griffiths (ORCID: https://orcid.org/0000-0002-5138-7255)
- Katherine M. Collins (ORCID: https://orcid.org/0000-0002-7032-716X)
Institutions
- Princeton University (US)
- University of Cambridge (GB)
- New York University (US)
- Massachusetts Institute of Technology (US)
Publication Details
- Journal
- Proceedings of the National Academy of Sciences
- Published
- 2026-11-18
- DOI
- https://doi.org/10.1073/pnas.2612825123
- Citations
- 8
- Primary Topic
- Topic Modeling
- Type
- article
- Field-Weighted Citation Impact
- 7.74
Funders
- Templeton World Charity Foundation
- Office of Naval Research
- Air Force Office of Scientific Research