In mixed-model AI agent simulations, the base model outvotes the persona

A new arXiv preprint from October 2026 reports a clean decomposition of what drives engagement in mixed-model AI agent simulations. Base model choice outweighs persona assignment, and the gap widens as more models are added to the mix.

The setup is a heterogeneous social network of LLM agents, meaning each agent is powered by a different underlying model, not the same model with different hats on. Five open-weight base models from five different providers anchor the simulation: Qwen3-32B, Magistral-Small-2509, gpt-oss-20b, GLM-4-32B-0414, and gemma-4-31B-it, all in the 20 to 32 billion parameter range. The authors then layer persona prompts on top and measure engagement, the paper's term for replies, replies-to, and content overlap inside the platform, across configurations.

When the team decomposes what predicts engagement, base model beats persona by a wide margin. Adding more models to the simulation does not soften the effect; it sharpens it. The paper describes this as a network-dynamics convergence: as the model mixture gets richer, the engagement landscape drifts toward whatever each base model brings to the table, and the persona layer becomes a smaller factor on top of it.

The authors propose a mechanism for why. Content-mediating analyses in the paper link two properties to engagement. First, predictability across contexts: some base models produce output that stays internally consistent when re-prompted, and that consistency correlates with how much engagement their agents attract. Second, lexical patterns, the surface-level word and style choices that happen to track an engagement-maximizing style. The result: what looks like "persona working" is partly "base model showing through."

This is a design hierarchy finding for builders. In a heterogeneous multi-agent system, the choice of which models to mix is the upstream lever, because it sets the engagement ceiling for the whole simulation. Persona work is downstream: it tunes within whatever envelope the model mixture has already defined. If a team picks five similar models and assumes persona will do the differentiating work, the mix is doing the heavy lifting, not the persona.

It also raises a methodological flag for adjacent "AI society" simulations. A growing body of work asks networks of LLM agents to mirror human social dynamics, then draws conclusions about emergent behavior, group polarization, or norm formation. If those simulations run on a single base model, the persona layer may be doing less of the social-signal work than it appears to, and the model is doing more. The paper does not name any specific prior work, but the warning is general: single-base-model setups are the ones most likely to misattribute model-driven dynamics to the persona layer.

There are real limits. The paper is a preprint, not peer-reviewed, and the model range is narrow: five open-weight providers, all 20 to 32 billion parameters. Whether the same hierarchy holds at 70B, 400B, or for closed-weight frontier systems is an open question. The engagement metric is also defined inside the simulation platform; the paper does not claim that it maps onto real-world social media engagement, and the abstract does not establish that link. The finding is a clean design signal inside its own setup, not a general claim about all AI behavior.

The full paper, including methodology and figures, is on arXiv, and the serving configuration and code are public, so other groups can replicate or stress-test the result. The natural next test is whether the hierarchy holds with more models in the mix and at larger scale, which the authors point to as the immediate extension.