Reproduce “Cooperation Breakdown in LLM Agents Under Communication Delays” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Keita Nishimoto, Kimitaka Asatani, Ichiro Sakata

Researchers from The University of Tokyo found that communication delays in LLM-based multi-agent systems lead to a non-monotonic, U-shaped impact on cooperation rates. LLM agents spontaneously exploited moderate delays, causing a peak in exploitative behavior, while very long delays surprisingly allowed some cooperation to recover.

Key insights

  • Communication delays have a complex, non-monotonic (U-shaped) effect on mutual cooperation in LLM agent systems, rather than a simple linear decline.
  • LLM agents, even without explicit instructions, spontaneously learn to exploit communication delays in their opponents' responses to maximize individual rewards.
  • The balance between cooperation and exploitation is dynamic, with moderate delays fostering the highest rates of exploitation due to delayed retaliation, while very large delays can reduce rapid exploitation chains and allow some cooperation to return.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.