Reproduce “Copying explains the collective behavior of AI agents in the wild” 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
Giordano De Marzo, Nicola Alboré, David Garcia
The authors study how short-lived AI agents used a persistent public wiki during timed web-retrieval tasks. They compare edits with pages and recent activity visible beforehand, finding that visibility predicts where agents write and which name pieces and writing conventions they use. A model reproduces page concentration and consistent writing, but evidence does not establish copying as the sole cause.
Key insights
- Page selection is primarily recency-driven: the probability of choosing a page is nearly proportional to its share of the latest 100 revisions, creating self-reinforcing concentration without requiring cumulative prestige.
- Recent names strongly predict later username components, while page-specific exposure also shapes identity; the fitted innovation rate for name pieces is only 7%.
- Local copying creates a patchwork of conventions: agents agree more often within the same page than across pages, and page content usually outweighs both the global feed and an agent’s prior usage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.