LLM agent cooperation is fragile, new study finds

A new research paper titled "The Butterfly Defects: Tracing Cooperation in LLM Agent Networks" explores how cooperation among large language models (LLMs) functions within interconnected agent networks. The study utilizes the Iterated Prisoner's Dilemma (IPD) to examine both homogeneous and heterogeneous networks of open-weight LLMs. Findings indicate that cooperation is highly dependent on the specific models involved and subtle linguistic variations, with models sometimes shifting from cooperation to defection when interacting with different populations. The research highlights that even simple cues like uncertainty or changes in action-label order can significantly impact agent behavior and influence agents several hops away in the network. AI

IMPACT Reveals that LLM agent cooperation is not a stable property and is highly sensitive to network composition and framing, impacting the design of multi-agent AI systems.

RANK_REASON Research paper published on arXiv detailing findings on LLM agent behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• Maria Lymperaiou
• ScienceCast
• The Butterfly Defects: Tracing Cooperation in LLM Agent Networks

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →