Claude Haiku 4.5 replication

Homophily under model substitution

A scaled-down topic simulation tests whether LLM agents preferentially engage with semantically similar others when the model family changes from GPT-3.5 to Claude Haiku 4.5.

Agents

100

Rounds

12

p-value

< 0.001

Canonical result

Reproduces, with attenuated magnitude.

Final modularity sits above the degree-preserving bootstrap null, while assortativity and content-engagement correlation remain weakly positive.

Mechanism

Agents post, read a mixed feed, and choose 1-3 posts to like, follow, or ignore.

Measurement

Weighted engagement graphs, Louvain communities, MiniLM embeddings, and a degree-preserving null.

Constraint

No prompt tells agents to seek similar others. The homophily signal has to emerge.