Pretrained Persona Mixture Models and Tandem Models for Human Simulation

Minwoo Kang; Téa Wright; Seun Eisape; Ayush Raj; Suhong Moon; Joseph Suh; Alane Suhr; David M. Chan

cs.CLSep 18, 2026

Pretrained Persona Mixture Models and Tandem Models for Human Simulation

Authors: Minwoo Kang, Téa Wright, Seun Eisape, Ayush Raj, Suhong Moon, Joseph Suh, Alane Suhr, David M. Chan, +1 more

Organizations: University of California, Berkeley · Google

Abstract

We argue here that the current dominant practice in LLM human simulation: prompting instruction-tuned assistant language models to role-play personas, is inaccurate and produces stereotyped predictions (lacking natural diversity). It has previously been shown that LLMs can be bound to personas using naturalistic, freetext dialog avoiding stereotyping. Here we show that binding can also be achieved using short, individual samples of dialog from specific people. Demographics can be added later without negative effects by simply querying the model. We use the term Persona Mixture Models (PMMs) for well-calibrated human models, currently realized as pretrained base models. We show that PMMs produce more accurate predictions than instruction-tuned models and retain more of the lexical, semantic, and pragmatic diversity found in human dialog. We measure realism and diversity of LLMs simulating human interlocutors across a diverse set of corpora spanning open-domain text, human-AI chat, and task-oriented dialogue between human speakers. However, base pretrained models can produce out-of-domain dialog and may lose some of the human's internal state over long contexts. We propose and explore tandem models which combine a pre-trained model with an instruction-tuned supervisor. Tandem models achieve the best overall accuracy and diversity in our experiments.

Figures & tables

Corpus	Dialogue type	Sampled Contexts	Total
Reddit (ConvoKit) [ 67 ]	Open-domain; Human–Human	1,024	8,192
WildChat-1M [ 68 ]	Human-LLM Chat; Human–AI	1,024	8,192
LMSYS-Chat-1M [ 69 ]	Human-LLM Chat; Human–AI	1,024	8,192
MultiWOZ 2.1 [ 70 ]	Task-Oriented; Human–Human	807	6,456
DialOp [ 8 ]	Task-Oriented; Human–Human	306	2,448

Table 1 : Conversational corpora used in our case study. For all corpora, we use similar next-utterance prediction setups with LMs and sample 8 generations per context. For Multiwoz and DialOp, we take 128 and 117 dialogues, respectively, and generate utterances of all human-user turns; for all other datasets, we take the final user turn and generate model utterances in place of the human interlocutor.

Figure 1 : Per-token self-entropy and negative log-likelihood. For each model, we report model self-entropy (shown as solid bars) and negative log-likelihood of human utterances given dialogue context. See Appendix E for full results.

Figure 2 : MAUVE results.

\mathrm{MAUVE}

is a sample-based score in

[0,1]

that quantifies how close two text distributions are. The score is high when the set of model samples and the set of human references cover the same regions of the distribution under a text embedding model; conversely, score is low when either distribution places mass where the other does not. A value near

1.0

indicates that the two samples are statistically indistinguishable, while lower values mean the model is sampling from a different region of utterance space than humans.

	JSD ( $\downarrow$ )			V Overlap ( $\uparrow$ )
Model	$N=1$	$2$	$3$	$N=1$	$2$	$3$
Tandem	.088	.219	.280	0.08	.75	.84
M24B-B	.181	.277	.330	0.08	.75	.74
M24B-I	.195	.299	.346	0.08	.68	.67
Q2.5-14B-B	.175	.272	.325	0.08	.68	.76
Q2.5-14B-I	.119	.273	.321	0.08	.68	.75

Table 2: Full JSD and Vocabulary Overlap results (

K=100

). B, I, and R respectively indicate Base, Instruct, and Reasoning model variants.

Figure 3 : Jensen–Shannon divergence (JSD) between human and model dialogue-act

N

-gram distributions on MultiWOZ-2.1. For each

N\in\{1,\dots,3\}

, we form the empirical distribution of speaker-marked dialogue-act

N

-grams from the aligned human reference (

P_{N}

) and from the model under evaluation (

Q_{N}

), then plot

\mathrm{JSD}(P_{N},Q_{N})

. JSD is bounded in

[0,1]

; lower is better , with

0

corresponding to identical distributions. Higher

N

probes longer pragmatic context—a model that matches marginal act frequencies (low JSD at

N{=}1

) can still fail to reproduce the ordering of acts across turns (high JSD at

N\geq 3

Figure 4 : Utterance-level diversity diagnostics on Reddit. For each model we report three complementary measures of how varied the model-generated continuations are within a given dialogue context, averaged across contexts. Self-BLEU (left) computes BLEU between every pair of continuations from the same context: lower is better , since a low value means the continuations are lexically distinct from one another. Self-BERT-F1 (center) is the analogous quantity in contextual-embedding space, capturing semantic rather than surface variation: again, lower indicates greater diversity . Vendi score (right) reports the effective number of semantically distinct samples, computed from the eigenspectrum of the embedding similarity kernel [ 78 ] : higher is better , with a value near $1$ corresponding to total mode collapse.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

System	PPL-H	PPL-G rand	Dist-1	Dist-2	Self-BLEU	Self-ROUGE-L	Self-BERT-F1	Vendi	MAUVE
Llama-3.1-8B	13.2424	19.2998	0.6743	0.9602	0.0122	0.0776	0.4796	6.3915	0.9650
Llama-3.1-8B-Instruct	23.1513	6.9341	0.5984	0.9339	0.0545	0.1304	0.5337	4.7601	0.6946
Mistral-24B-Base	10.7243	17.2362	0.6562	0.9581	0.0152	0.0845	0.4884	6.2010	0.9658
Mistral-24B-Instruct	13.4477	10.4384	0.6147	0.9263	0.0790	0.1473	0.5560	5.0822	0.8999
Qwen2.5-14B	12.8851	16.4860	0.6770	0.9593	0.0109	0.0799	0.4831	6.4020	0.9666
Qwen2.5-14B-Instruct	32.1536	2.7724	0.5726	0.8520	0.2017	0.2176	0.6285	4.1150	0.7728

Appendix

Table 3 : Full Reddit point-estimate results. PPL-H scores aligned human continuations; PPL-G rand scores one randomly sampled generated continuation per context.

System	PPL-H	PPL-G rand	Dist-1	Dist-2	Self-BLEU	Self-ROUGE-L	Self-BERT-F1	Vendi	MAUVE
Llama-3.1-8B	4.3322	8.1718	0.6483	0.8977	0.0785	0.1122	0.5141	6.1210	0.8608
Llama-3.1-8B-Instruct	5.2963	9.5169	0.6068	0.9044	0.0865	0.1510	0.5840	4.7968	0.7813
Mistral-24B-Base	3.5866	8.9617	0.6321	0.8875	0.0902	0.1170	0.5166	5.9821	0.9659
Mistral-24B-Instruct	4.0625	7.0132	0.6758	0.9275	0.0598	0.1315	0.5812	5.5360	0.6233
Qwen2.5-14B	3.8005	7.4436	0.6463	0.8871	0.0922	0.1114	0.5102	6.0735	0.9691
Qwen2.5-14B-Instruct	5.8571	2.4996	0.5625	0.7978	0.2501	0.2386	0.6641	4.0393	0.5007

Appendix

Table 4 : Full WildChat-1M point-estimate results. PPL-H scores aligned human continuations; PPL-G rand scores one randomly sampled generated continuation per context.

System	PPL-H	PPL-G rand	Dist-1	Dist-2	Self-BLEU	Self-ROUGE-L	Self-BERT-F1	Vendi	MAUVE
Llama-3.1-8B	4.0134	7.1166	0.6723	0.8993	0.0660	0.1117	0.5177	6.3393	0.8206
Llama-3.1-8B-Instruct	5.8418	3.5852	0.5955	0.8857	0.1070	0.1662	0.5911	4.8261	0.7639
Mistral-24B-Base	3.3318	7.3487	0.6714	0.8959	0.0746	0.1162	0.5185	6.2141	0.9503
Mistral-24B-Instruct	4.3267	4.9720	0.6673	0.9010	0.0918	0.1625	0.5977	5.3387	0.7026
Qwen2.5-14B	3.4565	6.1836	0.6676	0.8830	0.0837	0.1194	0.5196	6.1912	0.9499
Qwen2.5-14B-Instruct	12.5903	2.1503	0.5200	0.7306	0.3556	0.3049	0.6864	3.6589	0.6036

Appendix

Table 5 : Full LMSYS-Chat-1M point-estimate results. PPL-H scores aligned human continuations; PPL-G rand scores one randomly sampled generated continuation per context.

System	PPL-H	PPL-G rand	Dist-1	Dist-2	Self-BLEU	Self-ROUGE-L	Self-BERT-F1	Vendi	MAUVE
Llama-3.1-8B	5.5176	7.9276	0.7019	0.9066	0.0809	0.1862	0.5428	5.1383	0.6172
Llama-3.1-8B-Instruct	10.7274	2.8377	0.5258	0.8028	0.2477	0.2610	0.6700	3.6958	0.3526
Mistral-24B-Base	4.8951	5.8217	0.6910	0.8948	0.0956	0.1845	0.5804	5.0722	0.6743
Mistral-24B-Instruct	5.0876	3.6643	0.5731	0.8003	0.2424	0.2851	0.6932	3.8587	0.6687
Qwen2.5-14B	5.1235	6.3018	0.6654	0.8824	0.1153	0.2005	0.6367	4.9447	0.7132
Qwen2.5-14B-Instruct	20.0783	1.7114	0.4251	0.6084	0.5541	0.4712	0.6949	2.5738	0.7499

Appendix

Table 6 : Full MultiWOZ continuation point-estimate results. PPL-H scores aligned human continuations; PPL-G rand scores one randomly sampled generated continuation per context.

System	PPL-H	PPL-G rand	Dist-1	Dist-2	Self-BLEU	Self-ROUGE-L	Self-BERT-F1	Vendi	MAUVE
Llama-3.1-8B	12.4540	16.5786	0.7327	0.9570	0.0125	0.0829	0.5614	6.5064	0.4638
Llama-3.1-8B-Instruct	23.6024	4.4736	0.5588	0.8957	0.0868	0.1711	0.6723	4.1654	0.3901
Mistral-24B-Base	10.9471	12.0142	0.7194	0.9522	0.0180	0.0821	0.5996	6.4174	0.5942
Mistral-24B-Instruct	12.3744	9.9723	0.5982	0.9307	0.0434	0.1367	0.6078	4.8471	0.4900
Qwen2.5-14B	12.3213	8.0706	0.7019	0.9454	0.0245	0.0911	0.5743	6.2673	0.5640
Qwen2.5-14B-Instruct	57.3444	2.4584	0.5177	0.7909	0.2729	0.2582	0.6922	3.4711	0.3958

Appendix

Table 7 : Full DialOp point-estimate results. PPL-H scores aligned human continuations; PPL-G rand scores one randomly sampled generated continuation per context.

Figure 5 : Individual (N=1) dialogue act distributions for each model compared against the human distribution.

	JSD ( $\downarrow$ )			Overlap K=10 ( $\uparrow$ )			Overlap K=100 ( $\uparrow$ )			Overlap K=200 ( $\uparrow$ )
System	$N=1$	$2$	$3$	$N=1$	$2$	$3$	$N=1$	$2$	$3$	$N=1$	$2$	$3$
Llama-3.1-8B-Base	0.198	0.293	0.346	0.700	0.600	0.600	0.080	0.700	0.770	0.040	0.600	0.645
Llama-3.1-8B-Inst	0.265	0.366	0.407	0.400	0.700	0.700	0.070	0.640	0.640	0.035	0.495	0.570
Mistral-24B-Base	0.181	0.277	0.330	0.700	0.800	0.500	0.080	0.750	0.740	0.040	0.620	0.660
Mistral-24B-Inst	0.195	0.299	0.346	0.700	0.700	0.800	0.080	0.680	0.670	0.040	0.575	0.635
Qwen2.5-7B-Base	0.168	0.271	0.325	0.700	0.800	0.600	0.080	0.760	0.780	0.040	0.635	0.660

Appendix

Table 8 : Comprehensive results for N-gram Dialogue Act sequences (

N=1

3

). Results include Jensen-Shannon Divergence (JSD

\downarrow

) and Vocabulary Overlap (

\uparrow

) at multiple

K

thresholds.

Figure 6 : Transition probabilities for N=2 (i.e., probability of user intent given previous assistant intent).

Panel of Figure 7 — Figure 7 : Comparison of distributional divergence and vocabulary overlap between model outputs and human-human (H-H) bootstrapped baselines. H-H results are generated via 100 $\times$ random corpus splits. We observe that for $N\leq 3$ , H-H performance consistently exceeds model-human similarity, establishing $N=3$ as our threshold complexity limit.

Explore similar work

How Well Do Large Language Models Capture Human Personality?
May 12, 2026Aanisha Bhattacharyya, Yaman Kumar Singla, Rajiv Ratn Shah +2LLM Evaluation Representation Collapse
Adobe Media and Data Science Research (MDSR) · IIIT-Delhi · SUNY at Buffalo
Beyond Cooperative Simulators: Generating Realistic User Personas for Robust Evaluation of LLM Agents
May 13, 2026Harshita Chopra, Kshitish Ghate, Aylin Caliskan +3Human Behavior Simulation LLM Agent Evaluation
University of Washington, Seattle, WA · Georgetown University, Washington, DC
The Chameleon's Limit: Investigating Persona Collapse and Homogenization in Large Language Models
Apr 27, 2026Yunze Xiao, Vivienne J. Zhang, Chenghao Yang +3LLM Evaluation Personality Modeling in Language Models
CMU · UChicago · MIT +3

May 12, 2026, cs.HCJ/K move · Enter open · S save

How Well Do Large Language Models Capture Human Personality?

Aanisha Bhattacharyya, Yaman Kumar Singla, Rajiv Ratn Shah, Changyou Chen+1

Adobe Media and Data Science Research (MDSR) · IIIT-Delhi · SUNY at Buffalo

LLM Evaluation Representation Collapse Large Language Model-Based Role-Play Simulation Personality Modeling in Language Models Human Behavior Simulation Large Language Model-Based Social Simulation+1

Open paper AskPDF arXiv

Large language models (LLMs) are increasingly used to simulate human populations via persona prompting, often under the assumptions that richer persona descriptions improve behavioral fidelity, similarly sized attribute combinations are equally simulatable, and persona definitions generalize across tasks. In this work, we formalize these assumptions and systematically evaluate them across multiple architectures, scales, and simulation settings. We identify a fundamental limitation we term persona manifold collapse, where increasingly expressive persona specifications lead to systematic contraction of representational and behavioral diversity. Across models, increasing persona complexity consistently reduces inter-persona separation in latent space and weakens behavioral differentiation in downstream simulation tasks. These effects persist across multiple analyses as richer personas fail to preserve human subgroup disagreement, performance varies across attribute combinations of similar size, and adding descriptive detail often degrades rather than improves simulation fidelity. Surprisingly, simple Age-Gender personas consistently outperform richly specified Ideal Customer Profiles (ICPs) across industries, achieving substantially higher downstream prediction accuracy. We find that collapse is not uniform across attributes. Certain combinations remain behaviorally stable and preserve stronger alignment with human responses, forming localized regions we term alignment bridges. Together, our results provide empirical and conceptual foundations for understanding the limits of persona-conditioned simulation, highlighting the need for representation-aware persona construction rather than increasing persona expressivity alone.

Abstract

Explore similar work

How Well Do Large Language Models Capture Human Personality?

Beyond Cooperative Simulators: Generating Realistic User Personas for Robust Evaluation of LLM Agents

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