Reproduce “Social Networks of LLM Agents” 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
Kaixuan Liu, Guojun Xiong, Weinan Zhang, Shengpu Tang
Researchers from Emory University and Shanghai Jiao Tong University developed SNLA, a framework for Large Language Model (LLM) agent societies that models the distinction between visible exposure and realized influence. This framework identifies an "attention width" parameter that controls the transition between collective knowledge aggregation and herding, and introduces a decentralized equalization protocol that improved collective accuracy by +0.64 on a hidden-profile reasoning task.
Key insights
- The "attention width" parameter (beta) serves as a critical control mechanism, governing the transition between optimal "wisdom of crowds" behavior and a collapse into "herding" within LLM agent populations.
- Narrow attention causes herding by keeping the effective sample size bounded (e.g., <= 8), irrespective of population size, particularly when a unique dominant source exists in the network structure.
- The equalization protocol drives the realized-influence matrix towards a doubly stochastic state, theoretically ensuring optimal collective information aggregation and achieving the highest possible effective sample size.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.