Reproduce “Emergence of Preferential Attachment and Glass-Ceiling Effects in Autonomous Networks of LLMs” 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
Yiming Zhang, Vikram Krishnamurthy
The paper studies how LLM agents form communication hierarchies by choosing collaborators under a protocol that samples candidate sources according to prominence. In tests with identical Gemini base models, some roles typically gained more influence than counterparts, and this pattern persisted when disadvantaged roles received stronger models. Thus, base-model strength alone did not explain the observed hierarchy in these settings.
Key insights
- Autonomous LLM agent networks can spontaneously develop complex sociological phenomena such as preferential attachment and the glass-ceiling effect (GCE).
- The GCE can be induced by system prompt design alone, creating communication hierarchies where less capable agents gain disproportionate prominence, even overriding the inherent strength of base models.
- Network structures significantly impact information quality, with capability-aligned dominance enhancing truthfulness and GCE accelerating the spread of hallucinations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.