Disentangling Models from Personas in Heterogeneous LLM Simulations

Dani Roytburg; Daphne Ippolito

cs.MAOct 6, 2026

Disentangling Models from Personas in Heterogeneous LLM Simulations

Authors: Dani Roytburg, Daphne Ippolito

Organizations: Carnegie Mellon University

Abstract

Multi-agent simulations with large language models (LLMs) often operate networks of agents with a single base model. This overlooks the inter-model effects which may dominate engagement dynamics in real-world deployments. To show this, we simulate a heterogeneous social network powered by several different base models and show that the amount of engagement an agent receives depends more on its base model than on its assigned persona. The attraction or repulsion effects of a base model strengthen dramatically when more models are added in the mix, suggesting that networks dynamics may converge to base model effects at scale. To help explain this effect, we conduct a series of content-mediating analyses, showing the predictability of base models across contexts as well as the relationship between a model's lexical patterns and an engagement-maximizing style. In light of recent developments in mass multi-agent interaction, this work underscores the relevance of heterogeneous compositions in driving the outcomes of those networks

Figures & tables

Figure 1: Left: an interaction graph from one randomly-chosen social network simulation of 40 agents (dots, larger means higher in-degree) colored by the base model running the agent. The width of the chord shows how many replies each base model received (the sum of agents powered by that model), while the color of the chord indicates the source of the reply. self $n$ indicates how much traffic came from agents powered by the same model. $H$ is defined in Section 3.2 . Right: one such interaction from a four-way run. A gpt-oss agent’s post and the first reply from an agent run by another model.

Panel of Figure 2 — Figure 2: Excess cross-model engagement $H_{X\to Y}$ for two-way (left) and three-way (right) mixtures. Rows are source (commenter) families; columns are target (post-author) families; color encodes mean $H$ , and the $\pm$ value in each cell is the run-to-run standard deviation (Section 4 ). Per-cell numeric tables are in Appendix C.8 (Tables 8 , 9 ).

Figure 3: (a) Proportional variance of comments-per-post (CPP) explained by an agent’s own base model vs. its persona, by number of unique models in play (in-sample $R^{2}$ ; the parameter-count–robust held-out CV and $\Delta R^{2}$ are in Appendix D ). Base model overtakes persona at every size and the gap widens. (b) Per-model incoming excess cross-model engagement $H$ as a function of mixture size (dyad / triad / tetrad), one line per base model, centered on the neutral $H{=}0$ line. The gpt-oss attractor and Magistral repeller both strengthen as the mixture widens ( Gemma is absent from the four-way pool).

Excluded	Regime	$n$	Persona	Own	Selection
baseline	Two-way	1586	8.6%	10.2%	27.8%
	Three-way	2093	7.0%	15.6%	26.7%
Qwen	Two-way	923	12.4%	6.2%	20.5%
	Three-way	781	14.2%	19.2%	23.0%
gpt-oss	Two-way	893	10.4%	11.7%	27.2%
	Three-way	825	9.8%	9.3%	14.0%

Table 1: Leave-one-model-out variance decomposition (comments per post). Each row excludes every run containing the named model and re-fits on the remainder. Numbers are

R^{2}

for persona-only, own-model-only, and model-selection (own + partner-pair) feature blocks.

model	CV acc.	LPO acc.	characteristic phrases
gpt-oss-20b	0.98	0.97	merkle root , audit trail , zk snark , tamper evident
Qwen3-32B	0.88	0.87	let’s build , let’s make , here’s twist , i’ll draft
GLM-4-32B	0.85	0.84	beautifully captures , resonates deeply , aligns perfectly
Magistral-Small	0.88	0.85	ah user , alright listen , strikes chord , neon lights
Gemma-4-31B	0.98	0.94	you’re just , stop trying , just fancy , fancy way

Table 2: Predicting the writing model from a text’s lexical features per model, under 5-fold cross-validation (CV) and a leave-personas-out (LPO) split (train/test on disjoint personas). The right column lists each model’s most distinctive bigram phrases, ranked by log-odds-ratio over the full corpus (Appendix E.6 ).

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Metric	Value
Avg. cross-comments / agent	$25.2\pm 2.3$
Gini of in-degree	$0.49\pm 0.07$
Post:comment ratio	$0.13\pm 0.07$
Gini of posts / agent	$0.30\pm 0.04$
Max / mean in-degree	$3.65\pm 1.05$

Appendix

Table 3: Network sanity statistics (mean

\pm

std over 190 valid runs).

Regime	N runs	Max-dev mean $\pm$ std
Dyadic (target 50%/50%)	100	$4.79\%\pm 1.69\%$
Triadic (target 33.3%/33.3%/33.3%)	90	$8.72\%\pm 2.23\%$

Appendix

Table 4: Empirical post-share deviation from target across all valid 5-model runs. “Max-dev” is the per-run maximum across families of

|\hat{\sigma}_{X}-\sigma_{X}^{*}|

. Tolerance

\delta=0.05

is enforced as a soft cap.

gpt-oss comment share $\geq$	fraction of runs
$50\%$	$69\%$
$55\%$	$52\%$
$60\%$	$35\%$
$65\%$	$21\%$
$70\%$	$13\%$
$75\%$	$5\%$

Appendix

Table 5: Fraction of the

94

gpt-oss -containing runs in which gpt-oss ’s comment share exceeds each threshold (pooled dyadic and triadic). Because gpt-oss is a minority of agents, its share rarely clears

70\%

despite its consistent over-representation.

Model	Q0	Q1	Q2	Q3
GLM-4	$-0.11$	$-0.08$	$-0.12$	$-0.05$
Mag	$-0.03$	$-0.21$	$-0.19$	$-0.20$
Qwen	$-0.12$	$+0.17$	$+0.17$	$+0.21$
gpt-oss	$+0.59$	$+0.52$	$+0.42$	$+0.39$
Gemma	$+0.05$	$-0.15$	$-0.05$	$-0.03$

Appendix

Table 6: In-group self-preference

H_{X\to X}

by simulation quartile (combined dyadic + triadic runs).

Source	Q0	Q1	Q2	Q3
GLM-4	$+0.30$	$+0.54$	$+0.49$	$+0.32$
Mag	$+0.72$	$+0.62$	$+0.58$	$+0.50$
Qwen	$+0.51$	$+0.39$	$+0.26$	$+0.13$
gpt-oss	$+0.59$	$+0.52$	$+0.42$	$+0.39$
Gemma	$+0.50$	$+0.40$	$+0.28$	$+0.21$

Appendix

Table 7: Incoming

H

to gpt-oss by simulation quartile:

H_{X\to\text{gpt-oss}}

(combined dyadic + triadic runs).

	Qwen	GLM-4	Mag	oss	Gem
Qwen	$+.04\pm.25$	$-.17\pm.24$	$-.15\pm.08$	$+.16\pm.16$	$-.00\pm.07$
GLM-4	$-.01\pm.20$	$-.08\pm.22$	$-.13\pm.11$	$+.27\pm.11$	$-.03\pm.08$
Mag	$+.08\pm.24$	$+.04\pm.14$	$-.13\pm.31$	$+.42\pm.28$	$+.03\pm.16$
oss	$-.24\pm.13$	$-.37\pm.17$	$-.36\pm.18$	$+.34\pm.18$	$-.27\pm.17$
Gem	$-.05\pm.13$	$-.04\pm.09$	$-.14\pm.16$	$+.22\pm.16$	$+.04\pm.18$

Appendix

Table 8:

H_{X\to Y}

(mean

\pm

std) — Dyadic runs (100 valid). Rows: commenter; cols: post author. oss = gpt-oss, Gem = Gemma.

	Qwen	GLM-4	Mag	oss	Gem
Qwen	$+.16\pm.25$	$-.23\pm.24$	$-.32\pm.22$	$+.38\pm.18$	$-.22\pm.20$
GLM-4	$+.05\pm.25$	$-.09\pm.20$	$-.27\pm.27$	$+.46\pm.22$	$-.23\pm.24$
Mag	$+.10\pm.26$	$-.13\pm.34$	$-.17\pm.33$	$+.67\pm.32$	$-.24\pm.35$
oss	$-.06\pm.19$	$-.31\pm.21$	$-.34\pm.26$	$+.59\pm.28$	$-.37\pm.15$
Gem	$+.18\pm.24$	$-.10\pm.22$	$-.08\pm.26$	$+.40\pm.17$	$-.11\pm.26$

Appendix

Table 9:

H_{X\to Y}

(mean

\pm

std) — Triadic runs (90 valid). Same layout as Table 8 .

composition	runs	gpt-oss	Qwen	GLM-4	Magistral	Gemma
without Gemma	24	${+}0.93$	${+}0.12$	${-}0.45$	${-}0.70$	—
without Qwen	21	${+}0.84$	—	${-}0.24$	${-}0.19$	${-}0.51$
without Magistral	12	${+}0.77$	${+}0.18$	${-}0.43$	—	${-}0.63$
without GLM-4	6	${+}0.68$	${+}0.15$	—	${-}0.29$	${-}0.65$
pooled	63	${+}0.84$	${+}0.14$	${-}0.37$	${-}0.44$	${-}0.57$

Appendix

Table 10: Mean incoming

H

per model in each four-way composition. gpt-oss leads in

58/63

runs; the last place goes to Magistral (

23/24

) when Gemma is absent and to Gemma (

30/39

) when present.

Figure 4: Pooled four-way $H_{X\to Y}$ over the $24$ complete runs of the Qwen + gpt-oss + Magistral + GLM-4 composition. Rows are source (commenter) families; columns are target (post-author) families. gpt-oss is a uniform attractor column, Magistral a uniform repeller.

Block	Dy $R^{2}$	Tri $R^{2}$	Tet $R^{2}$
Persona only	0.086	0.070	0.056
Model only	0.102	0.156	0.317
Model selection (own + partner)	0.278	0.267	0.317
gpt-oss indicator only	0.008	0.093	0.288
Run only	0.350	0.219	0.049
Persona + model	0.190	0.224	0.364

Appendix

Table 11: Hierarchical OLS

R^{2}

for comments per post by feature block (in-sample).

\Delta R^{2}_{\text{model}}

is the unique variance attributable to base model after controlling for persona and run. Dy/Tri are the five-model pool (dyadic / triadic); Tet is the four-way pool (

24

runs,

2

seeds of Qwen , gpt-oss , Magistral , GLM-4 ), where all four models appear in every run, so the partner composition is constant and model selection reduces to own-model identity.

Block	Dy	Tri	Tet
Persona only	0.068	0.025	$-0.070$
Model only	0.067	0.149	0.316
Persona + model	0.147	0.183	0.297

Appendix

Table 12: Held-out 5-fold CV

R^{2}

for within-run-

z

CPP (Dy / Tri / Tet). Base model’s held-out

R^{2}

grows with mixture size; persona’s stays near zero.

	Dyad		Triad		Tetrad
Model	raw	adj	raw	adj	raw	adj
gpt-oss	$+0.41$	$+0.20$	$+0.63$	$+0.29$	$+0.90$	$+0.46$
Gemma	$+0.16$	$+0.17$	$+0.19$	$+0.27$	–	–
Qwen	$-0.13$	$+0.07$	$-0.12$	$+0.06$	$-0.08$	$+0.18$
GLM-4	$-0.31$	$-0.24$	$-0.44$	$-0.24$	$-0.54$	$-0.25$
Magistral	$-0.12$	$-0.23$	$-0.18$	$-0.32$	$-0.39$	$-0.55$

Appendix

Table 13: Per-model engagement (within-run-

z

CPP), raw vs. after partialling out mean post length. The gpt-oss attractor and Magistral repeller survive length control; Gemma is absent from the four-way pool.

representation	model CV	model LPO	persona CV
TF-IDF (topic + style)	0.80	0.70	0.73
StyleDistance (style)	0.76	0.74	0.22
Gemma (semantic)	0.79	0.75	0.83
MPNet (semantic)	0.72	0.66	0.76

Appendix

Table 14: Comparing different representation methods: recovering base model and persona from agent text (4,509 agents, posts only). Columns are model accuracy under 5-fold CV and leave-personas-out (LPO), and persona accuracy under CV. TF-IDF is the representation used throughout the paper.

axis	top $+$ terms	top $-$ terms	model spread (mean score)
1	let’s, just, digital, i’m	ici je, citer, je viens	all ${\approx}{+}0.2$ (language)
2	let’s, i’m, systems, you’re	proof, token, audit, zk	gpt-oss ${-}0.13$ vs rest ${\sim}{+}0.04$
3	actually, stop, noise, finally	explore, journey, greetings	Gemma ${+}0.11$ vs rest ${\leq}{+}0.01$
4	strategic, governance, architecture	tech, hey, share, got	Gemma ${+}0.03$ vs Mag ${-}0.05$

Appendix

Table 15: Selected SVD style axes: highest-loading terms at each pole (from

V^{\top}

) and the spread of mean axis scores across base models. Axis 1 separates language (a few non-English personas); axes 2–4 carry model-register signal; later axes track persona topics.

model	persona accuracy	$\times$ chance
Gemma	$50.8\%$	$22\times$
Magistral	$42.7\%$	$18\times$
gpt-oss	$33.6\%$	$14\times$
GLM-4	$30.7\%$	$13\times$
Qwen	$22.3\%$	$10\times$

Appendix

Table 16: Within-model persona recovery: accuracy of a persona classifier trained and tested on a single model’s own output (

3

-fold CV), and the multiple over chance. Every model expresses its assigned persona far above chance.

feature set	single-set $R^{2}$	unique $R^{2}$
length	$0.175$	$0.051$
style	$0.192$	$0.016$
model	$0.096$	$0.001$
persona	$0.057$	$0.008$
all four		$0.259$

Appendix

Table 17: Content vs. identity as predictors of (within-run

z

-scored) engagement. Cross-validated

R^{2}

: each feature set alone, and its unique contribution (full-model

R^{2}

minus the model omitting that set).

model	direction stability	cos to engagement
gpt-oss-20b	$0.96$	$+0.67$
Gemma-4-31B	$0.98$	$+0.13$
Qwen3-32B	$0.88$	$-0.19$
Magistral-Small	$0.95$	$-0.36$
GLM-4-32B	$0.89$	$-0.55$

Appendix

Table 18: Style geometry in the SVD space. “Direction stability” is the cosine between a model’s centroid recomputed on disjoint run-halves (mean over

60

splits;

1=

perfectly stable). “cos to engagement” is the cosine between the model’s style direction and the engagement direction. The latter recovers the attractor–repeller ordering of Finding 1.

Explore similar work

The Chameleon's Limit: Investigating Persona Collapse and Homogenization in Large Language Models
Apr 27, 2026Yunze Xiao, Vivienne J. Zhang, Chenghao Yang +3Artificial Intelligence Personas Personality
CMU · UChicago · MIT +3
The $\textit{Silicon Society}$ Cookbook: Design Space of LLM-based Social Simulations
Apr 30, 2026Aurélien Bück-Kaeffer, Sneheel Sarangi, Maximilian Puelma Touzel +3Scaling Improve Social Simulation Chip Design
McGill University · Mila - Quebec Artificial Intelligence Institute · Ubisoft La Forge +2
Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems
Sep 28, 2026Xavier Del Giudice, Alessio Palma, Matteo Migliarini +2Multi-Agent Large Language Model Systems Coalition Formation
Sapienza University of Rome, Italy

Apr 27, 2026, cs.CLJ/K move · Enter open · S save

The Chameleon's Limit: Investigating Persona Collapse and Homogenization in Large Language Models

Yunze Xiao, Vivienne J. Zhang, Chenghao Yang, Ningshan Ma+2

CMU · UChicago · MIT+3

Artificial Intelligence Personas Personality Collapse

Open paper AskPDF arXiv

Applications based on large language models (LLMs), such as multi-agent simulations, require population diversity among agents. We identify a pervasive failure mode we term \emph{Persona Collapse}: agents each assigned a distinct profile nonetheless converge into a narrow behavioral mode, producing a homogeneous simulated population. To quantify persona collapse, we propose a framework that measures how much of the persona space a population occupies (Coverage), how evenly agents spread across it (Uniformity), and how rich the resulting behavioral patterns are (Complexity). Evaluating ten LLMs on personality simulation (BFI-44), moral reasoning, and self-introduction, we observe persona collapse along two axes: (1) Dimensions: a model can appear diverse on one axis yet structurally degenerate on another, and (2) Domains: the same model may collapse the most in personality yet be the most diverse in moral reasoning. Furthermore, item-level diagnostics reveal that behavioral variation tracks coarse demographic stereotypes rather than the fine-grained individual differences specified in each persona. Counter-intuitively, \textbf{the models achieving the highest per-persona fidelity consistently produce the most stereotyped populations}. We release our toolkit and data to support population-level evaluation of LLMs.

Abstract

Explore similar work

The Chameleon's Limit: Investigating Persona Collapse and Homogenization in Large Language Models

The Silicon Society\textit{Silicon Society}Silicon Society Cookbook: Design Space of LLM-based Social Simulations

Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems

The $\textit{Silicon Society}$ Cookbook: Design Space of LLM-based Social Simulations