Reproduce “Microscopic dynamics of consensus formation in multi-agent LLM Naming Games” 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
Cristiano De Nobili, Vijayasri Iyer, Alessandro Codello, Raffaella Burioni
The authors model how an LLM listener’s yes-or-no decisions affect convention formation in a minimal Naming Game. They measure recognition and false-acceptance rates across temperatures for three open-weight models, finding that temperature’s association with consensus time differs by architecture; phi3 converges faster in hotter tested runs despite lower recognition. The rates are a first-order diagnostic, not a general deployment guarantee.
Key insights
- LLM architectures exhibit three distinct 'listener archetypes' (permissive, near-deterministic, conservative), characterized by unique $\pi(T)$ and $\phi(T)$ profiles that dictate their macroscopic consensus dynamics.
- Decoding temperature is not a universal 'noise knob'; its effect on consensus time is architecture-dependent, capable of slowing, having no impact, or even accelerating consensus, often mediated by effects like agent inventory accumulation.
- The derived analytical ordering condition $3\pi - 2\phi - 1 > 0$ provides a quantitative statistical-physics framework to predict and interpret the collective behavior of LLM populations based on their internal stochasticity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.