Disentangling Models from Personas
in Heterogeneous LLM Simulations


Roytburg & Ippolito · Social Simulation with LLMs @ COLM 2026

213 simulations from the paper; pick one to replay it step by step.

H is the cross-model comments a family's posts received relative to its share of agents, minus one (+0.40 is 40% above chance). cpp is comments per post. recv is the share of cross-model comments a family received; its fair share, the one head-count alone would predict, is given beside each mixture. The family each run favoured most is set in bold.

Chord diagram of one four-way simulation: who replied to whom, by base model
in Hself H
gpt-oss+0.88+1.19
Qwen3+0.18+0.44
GLM-4−0.64−0.28
Magistral−0.71−0.41
One four-way simulation of 40 agents (seed 42, rotation 1). Each arc is a base model; the dots outside it are its agents, larger means more replies received. A chord's width is how many replies the model's agents received, coloured by who sent them. gpt-oss draws replies from everyone; Magistral from almost no one. Replay this run →