Mind the Gaps: Mixture-of-Minds for Human Simulation

Pranav Dahiya

REVIEW 3 major objections 5 minor 44 references

Mind the Gaps: Mixture-of-Minds for Human Simulation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Anacreon claims that a mixture of per-cluster LLM adapters trained on public text reaches state-of-the-art individual-level survey accuracy, with ordinal alignment 0.775, on a narrow audience-simulation domain.

desk verdict A well-written paper with a genuinely new architecture, but its central SOTA claim rests on a circular evaluation: test questions are mined from the training corpus, and the same held-out split is used for both checkpoint selection and final scoring. read the letter →

arxiv 2608.06115 v1 pith:NWUV7RUH submitted 2026-08-06 cs.AI

Pranav Dahiya This is my paper · ORCID

classification cs.AI

keywords humansimulationLLMsimulatorsordinalalignmentdigitaltwinsmixtureofmindschain-of-emotioncontrastiveauthorshipembeddingsurveyprediction

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that LLM-based human simulators capture a population's average opinions while flattening its variation, and that the way out is to specialise: model a narrow, well-defined audience with a separate trained model per cluster of similar people. Anacreon does this by learning a contrastive authorship embedding that separates individuals in public text, clustering the corpus around seed people, then fine-tuning a dedicated LoRA adapter—a 'mixture of minds'—on a Gemma 4 12B base for every cluster. Records are augmented with harvested demographics, psychological traits, and mined survey answers, and with a chain-of-emotion that traces emotional state from stimulus to response; training also shuffles response options and balances positive and negative framings to counter prompt brittleness and sycophancy. On a large externally sourced survey of SME merchants, Anacreon reports ordinal alignment 0.775, which the paper presents as the best individual-level accuracy among systems that report the measure, ahead of 0.717, 0.740, and 0.748. A sympathetic reading: if this holds, individual-level prediction from public text is achievable for bounded populations, and the remaining gap to aggregate inference is a calibration problem rather than a modelling impossibility.

What carries the argument

The load-bearing mechanism is a 'mixture of minds': one specialised adapter per cluster of similar individuals, rather than a single dense model for everyone. The pipeline is: (1) a contrastive authorship-representation objective embeds individuals so samples from the same person are pulled together and samples from different people are pushed apart; (2) every record is assigned to its nearest seed individual, partitioning the corpus into clusters; (3) harvested demographics, psychological traits, and mined survey responses are appended to each record, together with a chain-of-emotion, a short trace of emotional states and appraisals from stimulus to response; (4) a separate high-rank LoRA adapter is fine-tuned per cluster on a Gemma 4 12B base with QLoRA, while response-option order is shuffled and non-neutral questions are included in both positive and negative framing; (5) the best checkpoint is selected by maximising $\alpha A_{\mathrm{mc}}(\theta) + \beta A_{\mathrm{ord}}(\theta) - \gamma \mathrm{PPL}(\theta)$ on held-out data. The chain-of-emotion is what conditions the final response on an inferred internal state rather than on the prompt's wording alone.

What would settle it

Ask Anacreon to answer a newly written survey on the same SME-merchant domain whose questions were not mined from the training corpus, and compare its ordinal alignment on those fresh items with the reported 0.775; a large drop would show the score reflects question-format familiarity rather than generalizable individual prediction. A quicker check is to recompute the score after excluding the newest-20% held-out records from the checkpoint-selection objective in eq. (5).

Watch

Extended reading notes

Core claim

The central claim is that individual-level survey responses can be predicted more accurately by a mixture of specialist models than by a single prompted or fine-tuned LLM, and Anacreon is the demonstration on a bounded population: SME merchants who take physical, in-person card payments. In the paper's own terms, the discovery is an extension: clustering a real qualitative corpus around seed people, training one high-rank LoRA adapter per cluster on Gemma 4 12B, augmenting each record with a chain-of-emotion, and selecting checkpoints by a combined score of multiple-choice accuracy, ordinal alignment, and chain-of-emotion perplexity yields a median ordinal alignment of 0.775 and a median top-1 accuracy of 0.679 on a large, externally sourced survey. The paper states that 0.775 is the state of the art on the ordinal-alignment measure, exceeding the reported 0.717, 0.740, and 0.748 of prior digital-twin systems, and that the residual positive bias is small though not eliminated. It frames the result as narrowing, not dissolving, the individual-aggregate gap, with population-level fidelity still to be shown.

Load-bearing premise

The load-bearing premise is that the survey questions mined from the same public corpus, and the newest 20% of each cluster's records, are genuinely held out, so the 0.775 score measures prediction of new questions rather than recall of training data—even though those same held-out records are also used to select checkpoints.

Editorial extensions

If this is right

If the reported 0.775 survives clean held-out evaluation, per-cluster fine-tuning on mined public text becomes the strongest reported approach for individual-level survey prediction in a bounded domain.
The recipe—seed clustering, contrastive authorship embedding, attribute mining, per-cluster adapters, chain-of-emotion—can be transferred to other narrow populations that leave enough public text, without needing a single dense retrain.
Because response-option shuffling and mirrored positive/negative framings are part of training, models built this way should be less sensitive to reworded questions than prompted simulators.
The paper's own next step is to post-stratify the per-cluster models to a target population; that is what would convert individual-level fidelity into aggregate prediction.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper: the 0.775 is the median over the 400 retained clusters after a long tail of non-converging clusters was pruned, so an uncurated population would likely score lower unless cluster quality is consistent.
Beyond the paper: because the test questions are mined from the same corpus that supplies training data, some of the gain may reflect familiarity with question format; a genuinely new instrument on the same merchant population is the clean test.
Beyond the paper: the small residual positive bias (+0.437 scale positions) suggests that a simple per-cluster bias-correction layer added at inference could raise ordinal alignment further without changing the architecture.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know two things about this paper. First, the architecture is real: a mixture of per-cluster LoRA adapters on a 12B base, with authorship embeddings used to cluster a corpus, is a sensible way to attack individual-level heterogeneity. Second, the headline number, 0.775 ordinal alignment, is not an unbiased estimate. The paper itself gives you the evidence for that. Section 3.2 says "Survey questions are also mined from the text," which directly contradicts the abstract's "externally sourced survey." Section 3.3 selects the best checkpoint by maximizing Eq. (5) on the same held-out temporal split that Section 4.2 then uses to report the final score. That is selection on the test set. Table 1 concedes the comparisons are "indicative rather than a controlled comparison," so the SOTA claim is unsupported.

What the paper does well: it is clearly written, the motivation is grounded in a real literature (Graunt to Simulmatics to digital twins), and it tackles known failure modes—prompt brittleness and positive bias—with concrete techniques. The chain-of-emotion is a reasonable extension of prior appraisal-based work, and the discussion of variance versus human test-retest is honest. The engineering is substantial: 420 clusters, 110B parameters, and a deployed B2B system. Credit is due there.

Where it falls short is the evaluation. Beyond the circularity, the comparison set is not controlled. Park et al. is interview-grounded and tested on different questions; Twin-2K-500 and the mega-study use different instruments. So the 0.775 number is not comparable to 0.748 without a shared benchmark. The paper needs a disjoint validation split and an external benchmark, ideally on a standard dataset like a published survey, to make the SOTA claim credible.

Who is this for? Researchers working on LLM-based human simulation, particularly digital twins and audience modeling. The architecture idea is worth exploring, but the current paper does not provide a valid test of it. I would send it to peer review because the idea is meaningful and the flaws are fixable—but it needs major revision. The claims about state of the art should be withdrawn until the evaluation is clean.

Recommendation: send to a serious venue, require the authors to fix the evaluation, and do not accept the current version.

Referee Report

3 major / 5 minor

Summary. The paper introduces Anacreon, an LLM-based audience simulator that clusters a public text corpus around seed individuals, trains a per-cluster LoRA adapter on top of a Gemma 4 12B base, and augments records with mined attributes and a chain-of-emotion. The central empirical claim is a state-of-the-art ordinal alignment of 0.775 on what the abstract calls a large, externally sourced survey, together with a top-1 accuracy of 0.679 on mined multiple-choice questions. The method is presented with attention to known failure modes such as prompt brittleness and positivity bias. However, the evaluation has two load-bearing problems: the survey questions are not externally sourced but mined from the same corpus used for training, and the checkpoint-selection procedure in Eq. (5) uses the same held-out split on which the final scores are then reported. As a result, the headline 0.775 figure is not an unbiased estimate of generalization to new questions, and the state-of-the-art claim is not supported by the evidence presented.

Significance. If the claimed result were established by a valid, externally grounded evaluation, it would be a meaningful advance for individual-level human simulation: prior digital-twin systems report ordinal alignment values around 0.717-0.748, so a validated 0.775 would be a clear improvement. The paper also deserves credit for addressing two documented failure modes in LLM simulation (option-order prompt brittleness and positive/sycophantic bias) and for explicitly acknowledging in Table 1 that cross-dataset comparisons are indicative rather than controlled. These strengths do not compensate for the evaluation-circularity problem, because the paper's main contribution is the empirical claim itself. The method may be interesting, but the manuscript as written does not establish that it outperforms prior systems.

major comments (3)

[Abstract; §3.2] The abstract and §4.2 describe the evaluation as a 'large, externally sourced survey,' but §3.2 states that 'Survey questions are also mined from the text.' The test questions therefore come from the same public corpus that supplies the training records. Even though survey questions are mined separately on train and test splits, the ground truth is not an external survey instrument; it is derived from the same population and the same text genre used for training. This directly contradicts the abstract and means the 0.775 figure should not be presented as performance on a new, externally administered survey.
[§3.3, Eq. (5); §4.2] The best checkpoint is chosen by maximizing α·A_mc(θ) + β·A_ord(θ) - γ·PPL(θ) on 'held-out data,' and §4.2 then reports the median per-cluster A_ord on what appears to be the same temporal 20% split described in §3.1. No disjoint validation split is described anywhere in the paper. Selecting the checkpoint with a criterion that includes A_ord and then reporting A_ord on the same data makes the reported 0.775 a selected maximum rather than an unbiased estimate of performance; the magnitude of the optimistic bias is unknown and could be substantial, especially with 420 clusters and 400 retained models.
[Table 1; §4.2] The 'state of the art' claim rests on comparisons across different datasets, different survey instruments, and somewhat different experimental protocols. The table caption itself concedes that the scores are 'indicative rather than a controlled comparison.' A higher number on a non-identical evaluation does not by itself establish state-of-the-art status. A controlled comparison on a common benchmark or a matched external survey is required before the headline claim can be made.

minor comments (5)

[§3.2] The sentence 'Due to the positivity bias in LLM responses [10]. most of the ordinal questions...' contains a misplaced period after the citation; it should read 'responses [10], most of...'.
[§3.3] The paper does not report the number of seed individuals, the minimum data threshold for seed inclusion, the clustering algorithm, or the selection weights α, β, and γ in Eq. (5). These details are needed for reproducibility.
[§4.1] The pooled signed-error analysis reports that each model was prompted five times with shuffled options, but the aggregation of those five responses into a single prediction is not described; the reader cannot tell whether the reported error uses the mean, median, or a single sample.
[Figure 7; §5] Figure 7 is labeled 'illustrative' and appears to be a schematic representation rather than measured data; the text should state explicitly that the comparison is not an empirical measurement of Anacreon's skew against human test-retest.
[§5] The discussion refers to 'a belief network [27]' as an analogy for the cluster training data, but the connection is not developed; as written it is a metaphor rather than a formal or quantitative claim.

Circularity Check

1 steps flagged · score 6.0 of 10

The SOTA ordinal-alignment claim is inflated by selecting checkpoints on the test set: Eq. (5) maximizes Aord on the same held-out split whose Aord is later reported as 0.775.

fitted input called prediction [Section 3.3 (Eq. 5), Section 3.1 (temporal split), Section 4.2 (Figure 5a), Figure 1 caption]
"The best checkpoint is chosen by balancing three quantities on held-out data. ... The second is the ordinal alignment Aord(θ) of eq. (1) on the ordinal questions. ... θ⋆ = arg maxθ α Amc(θ) + β Aord(θ) − γ PPL(θ) ... Figure 5a shows ordinal alignment across the evaluation clusters. The distribution is centred at 0.775 ... This is the state of the art on this measure."

The 'held-out data' used in Eq. (5) is the same newest-20% test split described in Section 3.1; no separate validation split is introduced. Eq. (5) selects the checkpoint θ⋆ by maximizing β Aord(θ) on that split, and Section 4.2 then reports Aord(θ⋆) on the same split as the final, state-of-the-art score. The reported 0.775 is therefore a maximum over candidate checkpoints evaluated against the test set itself, not an unbiased estimate of generalization. The 'prediction' is selected against the exact quantity later presented as the measured result, so the score is optimistically biased by construction.

full rationale

The paper contains no self-citation chain or imported uniqueness theorem: the authorship embedding, chain-of-emotion, and survey mining all cite independent prior work, and the single-author paper does not lean on its own prior results. The central derivation is also not circular in the sense of re-deriving its own assumptions. However, the headline claim rests on one formally circular evaluation step: Eq. (5) uses Aord on held-out data to choose the best checkpoint, and Section 4.2 reports Aord on that same split as the SOTA figure. Because there is no described separate validation set, the final 0.775 is a selected maximum against the test data rather than an independent measurement. This is a genuine fitted-input-called-prediction step, and it alone justifies a score of 6. A related concern, that the 'externally sourced survey' is actually mined from the same public text used for training (Section 3.2), further weakens external validity, but it is a data-independence issue rather than a formal reduction, so it does not raise the score further.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

The central claims rest on several assumptions about public text, mined survey questions, and checkpoint selection. The most important is that survey questions mined from the training corpus are independent ground truth, which the paper does not establish. The free parameters are mostly undisclosed, and the chain-of-emotion is not shown to help.

free parameters (6)

Contrastive temperature tau (eq. 3) = not reported
Controls how strongly the authorship embedding separates individuals; the value is undisclosed and directly affects clustering quality.
Number of seed individuals / clusters = 420 trained, 400 retained
The granularity of the mixture of minds is a design choice; the paper does not report how seeds were selected.
LoRA rank = 64
Larger rank increases trainable parameters per cluster; the choice is not justified against alternatives.
Checkpoint selection weights alpha, beta, gamma (eq. 5) = not reported
These weights balance top-1 accuracy, ordinal alignment, and chain-of-emotion perplexity. Their values are not given, and they may have been tuned on the same held-out data used for evaluation.
Minimum data threshold for seed inclusion = not reported
Inclusion in the seed population requires a minimum quantity of data per individual, but the cutoff is not specified.
Cluster pruning threshold = not reported
A long tail of clusters failed to converge and was pruned; the criterion is not defined.

assumptions (6)

domain assumption Public behavior predicts private disposition.
Section 1 argues public text reveals demographics, traits, and survey responses, citing Kosinski et al. The entire attribute-harvesting pipeline relies on this.
domain assumption Survey questions mined from free text yield valid ground-truth responses.
Section 3.2 converts user text into multiple-choice questions with known answers. If this mining is noisy or constructed, all reported accuracy numbers are affected.
domain assumption The temporal 80/20 split prevents leakage.
Section 3.1 holds out the newest 20% of records per cluster, but the paper does not show that topics or phrasings do not reappear across the split.
ad hoc to paper The held-out set can be used for checkpoint selection and final evaluation without overfitting.
Section 3.3 selects the best checkpoint using a score that includes Aord on held-out data, and Section 4.2 then reports Aord on that same held-out data. No separate validation split is mentioned.
ad hoc to paper Chain-of-emotion can be reliably generated and improves response fidelity.
Section 3.2.1 introduces the chain-of-emotion, but the paper provides no ablation showing it improves accuracy. Section 4.3 only reports perplexity, which is a convergence signal, not a validity check.
domain assumption Ordinal alignment Aord = 1 - MAD/R is the appropriate individual-level fidelity measure.
Section 2.3 asserts the field has converged on this measure. The paper acknowledges it is necessary but not sufficient, and adds no distribution-level check for the headline result.

invented entities (1)

Chain-of-emotion (internal emotional-state trace)
purpose: Conditions response generation on a sequence of appraisals and emotional states before producing the answer.
No ablation or external validation shows that the chain improves accuracy; the paper only reports its perplexity in Section 4.3.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Mind the Gaps: Mixture-of-Minds for Human Simulation." pith.science (2026). https://pith.science/paper/NWUV7RUH

@misc{pith2026260806115,
  author       = {Pith},
  title        = {Pith review of: Mind the Gaps: Mixture-of-Minds for Human Simulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NWUV7RUH}},
  note         = {Machine review of arXiv:2608.06115}
}

read the original abstract

Predicting how a population will answer a new question is a long-standing goal. Statistical methods succeed at the level of the mass but falter at the level of the individual. Large language model simulators inherit this gap. They recover a population's central tendencies while flattening its heterogeneity, and they carry social biases and prompt brittleness that distort individual predictions. This paper introduces Anacreon, an audience simulation model that targets the individual level within a narrow, well-specified domain. Anacreon learns an authorship embedding that separates individuals, clusters a real qualitative corpus around seed people, and trains a dedicated adapter for each cluster, a mixture of minds, on a Gemma~4 12B base. It harvests demographics, psychological traits, and survey responses from public text, and augments each record with a chain-of-emotion. It reduces prompt brittleness by shuffling response options and reduces positive bias by balancing the training distribution. On a large, externally sourced survey, Anacreon reaches a state-of-the-art ordinal alignment of 0.775, the individual-level accuracy measure on which the field has converged, with a small residual bias. The work is a step toward drawing aggregate insight from faithfully simulated individuals.

Figures

Figures reproduced from arXiv: 2608.06115 by the authors.

**Figure 2.** Generation of a mined survey question. Due to the positivity bias in LLM responses [10]. most of the ordinal questions generated by the mining agent had positive responses, skewing towards "agree"/"strongly agree" in 90% of cases. To fix this problem, every question with a non-neutral response was included twice in the training and test sets, once framed positively and once negatively. This not only aims to address … view at source ↗

**Figure 3.** Signed prediction error, in semantic scale positions. (a) pooled over all survey rows, peaked [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Demographic composition over age, gender, income, location, marital status, and merchant [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Frequency distribution of per-cluster survey performance. (a) ordinal alignment ( [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Chain-of-emotion perplexity frequency distribution per cluster. The vertical line marks the [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 7.** Figure 7: Illustrative response distributions on a five-point Likert item, each against the same ground [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 21 canonical work pages

[1]

Christopher H. Achen. Mass political attitudes and the survey response.American Political Science Review, 69(4):1218–1231, 1975. doi:10.2307/1955282

work page doi:10.2307/1955282 1975
[2]

Alwin and Jon A

Duane F. Alwin and Jon A. Krosnick. The reliability of survey attitude measurement: The influence of question and respondent attributes.Sociological Methods & Research, 20(1): 139–181, 1991. doi:10.1177/0049124191020001005

work page doi:10.1177/0049124191020001005 1991
[3]

Argyle, Ethan C

Lisa P. Argyle, Ethan C. Busby, Nancy Fulda, Joshua R. Gubler, Christopher Rytting, and David Wingate. Out of one, many: Using language models to simulate human samples.Political Analysis, 31(3):337–351, 2023. doi:10.1017/pan.2023.2

doi:10.1017/pan.2023.2 2023
[4]

Bettman, Mary Frances Luce, and John W

James R. Bettman, Mary Frances Luce, and John W. Payne. Constructive consumer choice processes.Journal of Consumer Research, 25(3):187–217, 1998. doi:10.1086/209535

doi:10.1086/209535 1998
[5]

Eckstein, Noémi Élteto, Thomas L

Marcel Binz, Elif Akata, Matthias Bethge, Franziska Brändle, Fred Callaway, Julian Coda-Forno, Peter Dayan, Can Demircan, Maria K. Eckstein, Noémi Élteto, Thomas L. Griffiths, Susanne Haridi, Akshay K. Jagadish, Li Ji-An, Alexander Kipnis, Sreejan Kumar, Tobias Ludwig, Marcel Mathony, Marcelo Mattar, and Eric Schulz. A foundation model to predict and capt...

doi:10.1038/s41586-025-09215-4 2025
[6]

Clinton, Cassy Dorff, Brenton Kenkel, and Jennifer M

James Bisbee, Joshua D. Clinton, Cassy Dorff, Brenton Kenkel, and Jennifer M. Larson. Synthetic replacements for human survey data? the perils of large language models.Political Analysis, 32(4):401–416, 2024. doi:10.1017/pan.2024.5

doi:10.1017/pan.2024.5 2024
[7]

An appraisal-based chain-of-emotion architecture for affective language model game agents.PLoS ONE, 19(5): e0301033, 2024

Maximilian Croissant, Madeleine Frister, Guy Schofield, and Cade McCall. An appraisal-based chain-of-emotion architecture for affective language model game agents.PLoS ONE, 19(5): e0301033, 2024. doi:10.1371/journal.pone.0301033

doi:10.1371/journal.pone.0301033 2024
[8]

QLoRA: Efficient fine- tuning of quantized LLMs

Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. QLoRA: Efficient fine- tuning of quantized LLMs. InAdvances in Neural Information Processing Systems (NeurIPS), volume 36, 2023. URL https://proceedings.neurips.cc/paper_files/paper/2023/ hash/1feb87871436031bdc0f2beaa62a049b-Abstract-Conference.html

work page 2023

Show all 44 references

[9]

Epstein and Robert Axtell.Growing Artificial Societies: Social Science from the Bottom Up

Joshua M. Epstein and Robert Axtell.Growing Artificial Societies: Social Science from the Bottom Up. Complex Adaptive Systems. Brookings Institution Press and MIT Press, Washington, DC, 1996. doi:10.7551/mitpress/3374.001.0001. 11

1996 doi
[10]

SycEval: Evaluating LLM sycophancy

Aaron Fanous, Jacob Goldberg, Ank Agarwal, Joanna Lin, Anson Zhou, Sonnet Xu, Vasiliki Bikia, Roxana Daneshjou, and Sanmi Koyejo. SycEval: Evaluating LLM sycophancy. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES), volume 8, pages 893–900. AAAI Pres...

2025 doi
[11]

Gallegos, Ryan A

Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K. Ahmed. Bias and fairness in large language models: A survey.Computational Linguistics, 50(3):1097–1179, 2024. doi:10.1162/coli_a_00524

2024 doi
[12]

Gallup and Saul Forbes Rae.The Pulse of Democracy: The Public-Opinion Poll and How It Works

George H. Gallup and Saul Forbes Rae.The Pulse of Democracy: The Public-Opinion Poll and How It Works. Simon & Schuster, New York, 1940

1940
[13]

Gemma 4 technical report

Gemma Team. Gemma 4 technical report. Technical report, Google DeepMind, 2026. URL https://arxiv.org/abs/2607.02770

2026 arXiv
[14]

Printed by Tho

John Graunt.Natural and Political Observations Mentioned in a Following Index, and Made upon the Bills of Mortality. Printed by Tho. Roycroft, for John Martin, James Allestry, and Tho. Dicas, London, 1662
[15]

Population-aligned persona generation for LLM-based social simulation.arXiv preprint arXiv:2509.10127, 2025

Zhengyu Hu, Jianxun Lian, Zheyuan Xiao, Max Xiong, Yuxuan Lei, Tianfu Wang, Kaize Ding, Ziang Xiao, Nicholas Jing Yuan, and Xing Xie. Population-aligned persona generation for LLM-based social simulation.arXiv preprint arXiv:2509.10127, 2025. URL https: //arxiv.org/abs/2509.10127

2025
[16]

Mercer, Lalit R

Frederick Jelinek, Robert L. Mercer, Lalit R. Bahl, and James K. Baker. Perplexity—a measure of the difficulty of speech recognition tasks.Journal of the Acoustical Society of America, 62 (S1):S63, 1977. URLhttps://pubs.aip.org/asa/jasa/article/62/S1/S63/642598

1977
[17]

Understanding the effects of RLHF on LLM generalisation and diversity

Robert Kirk, Ishita Mediratta, Christoforos Nalmpantis, Jelena Luketina, Eric Hambro, Edward Grefenstette, and Roberta Raileanu. Understanding the effects of RLHF on LLM generalisation and diversity. InThe Twelfth International Conference on Learning Representations (ICLR),
[18]

Bernstein

Akaash Kolluri, Shengguang Wu, Joon Sung Park, and Michael S. Bernstein. Finetuning LLMs for human behavior prediction in social science experiments. InProceedings of the 2025 Con- ference on Empirical Methods in Natural Language Processing (EMNLP), pages 30096–30111, Suzhou, ...

2025 doi
[19]

Private traits and attributes are predictable from digital records of human behavior.Proceedings of the National Academy of Sciences, 110 (15):5802–5805, 2013

Michal Kosinski, David Stillwell, and Thore Graepel. Private traits and attributes are predictable from digital records of human behavior.Proceedings of the National Academy of Sciences, 110 (15):5802–5805, 2013. doi:10.1073/pnas.1218772110

2013 doi
[20]

Liveright, New York, 2020

Jill Lepore.If Then: How the Simulmatics Corporation Invented the Future. Liveright, New York, 2020

2020
[21]

Lord and Melvin R

Frederic M. Lord and Melvin R. Novick.Statistical Theories of Mental Test Scores. Addison- Wesley, Reading, MA, 1968

1968
[22]

George Allen & Unwin, London, 1962

Maurice Edward Ogborn.Equitable Assurances: The Story of Life Assurance in the Experience of the Equitable Life Assurance Society, 1762–1962. George Allen & Unwin, London, 1962

1962
[23]

Park, Andrew Gelman, and Joseph Bafumi

David K. Park, Andrew Gelman, and Joseph Bafumi. Bayesian multilevel estimation with poststratification: State-level estimates from national polls.Political Analysis, 12(4):375–385,
[24]

Cai, Meredith Ringel Morris, Percy Liang, and Michael S

Joon Sung Park, Lindsay Popowski, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Social simulacra: Creating populated prototypes for social computing systems. InProceedings of the 35th Annual ACM Symposium on User Interface Software and Technolog...

2022
[25]

O’Brien, Carrie J

Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST), pages ...

2023
[26]

Zou, Aaron Shaw, Benjamin Mako Hill, Carrie Cai, Meredith Ringel Morris, Robb Willer, Percy Liang, and Michael S

Joon Sung Park, Carolyn Q. Zou, Aaron Shaw, Benjamin Mako Hill, Carrie Cai, Meredith Ringel Morris, Robb Willer, Percy Liang, and Michael S. Bernstein. Generative agent simulations of 1,000 people.arXiv preprint arXiv:2411.10109, 2024. URL https://arxiv.org/abs/2411. 10109

2024 arXiv
[27]

Morgan Kaufmann, San Mateo, CA, 1988

Judea Pearl.Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann, San Mateo, CA, 1988

1988
[28]

Merlau, Grace Jiarui Fan, Malek Ben Sliman, Melanie Brucks, Eric J

Tianyi Peng, George Gui, Daniel J. Merlau, Grace Jiarui Fan, Malek Ben Sliman, Melanie Brucks, Eric J. Johnson, Vicki Morwitz, et al. A mega-study of digital twins reveals strengths, weaknesses and opportunities for further improvement.arXiv preprint arXiv:2509.19088, 2025. UR...

2025 arXiv
[29]

Bachelier, Paris, 1835

Adolphe Quételet.Sur l’homme et le développement de ses facultés, ou Essai de physique sociale. Bachelier, Paris, 1835. Published in 2 volumes
[30]

Language models are unsupervised multitask learners.OpenAI technical report,

Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners.OpenAI technical report,
[31]

Rivera-Soto, Olivia Elizabeth Miano, Juanita Ordonez, Barry Y

Rafael A. Rivera-Soto, Olivia Elizabeth Miano, Juanita Ordonez, Barry Y . Chen, Aleem Khan, Marcus Bishop, and Nicholas Andrews. Learning universal authorship represen- tations. InProceedings of the 2021 Conference on Empirical Methods in Natural Lan- guage Processing (EMNLP),...

2021 doi
[32]

Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. Whose opinions do language models reflect? InProceedings of the 40th In- ternational Conference on Machine Learning (ICML), volume 202 ofProceedings of Machine Learning Research, pa...

2023
[33]

Schelling

Thomas C. Schelling. Dynamic models of segregation.The Journal of Mathematical Sociology, 1(2):143–186, 1971. doi:10.1080/0022250X.1971.9989794

1971
[34]

Smaldino

Paul E. Smaldino. Models are stupid, and we need more of them. In Robin R. Vallacher, Andrzej Nowak, and Stephen J. Read, editors,Computational Social Psychology. Psychology Press, New York, 2017. URL https://smaldino.com/wp/wp-content/uploads/2017/ 01/Smaldino2017-ModelsAreStupid.pdf

2017
[35]

Why the 1936 Literary Digest poll failed.Public Opinion Quarterly, 52(1): 125–133, 1988

Peverill Squire. Why the 1936 Literary Digest poll failed.Public Opinion Quarterly, 52(1): 125–133, 1988. doi:10.1086/269085

1936 doi
[36]

sigma 3/2024: World insurance: Strengthening global resilience with a new lease of life

Swiss Re Institute. sigma 3/2024: World insurance: Strengthening global resilience with a new lease of life. Technical report, Swiss Re, 2024. URL https://www.swissre.com/institute/research/sigma-research/ sigma-2024-03-world-insurance-global-resilience.html

2024
[37]

Gui, Tianyi Peng, Daniel J

Olivier Toubia, George Z. Gui, Tianyi Peng, Daniel J. Merlau, Ang Li, and Haozhe Chen. Twin- 2K-500: A dataset for building digital twins of over 2,000 people based on their answers to over 500 questions.Marketing Science, 44(6):1446–1455, 2025. doi:10.1287/mksc.2025.0262

2025
[38]

Prince- ton University Press, Princeton, NJ, 1944

John von Neumann and Oskar Morgenstern.Theory of Games and Economic Behavior. Prince- ton University Press, Princeton, NJ, 1944

1944
[39]

Dickerson

Angelina Wang, Jamie Morgenstern, and John P. Dickerson. Large language models that replace human participants can harmfully misportray and flatten identity groups.Nature Machine Intelligence, 7:400–411, 2025. doi:10.1038/s42256-025-00986-z. 13

2025 doi
[40]

Empirical validation of agent-based models: Alternatives and prospects.Journal of Artificial Societies and Social Simulation, 10(2): 8, 2007

Paul Windrum, Giorgio Fagiolo, and Alessio Moneta. Empirical validation of agent-based models: Alternatives and prospects.Journal of Artificial Societies and Social Simulation, 10(2): 8, 2007. URLhttps://www.jasss.org/10/2/8.html

2007
[41]

HumanLM: Simulating users with state alignment beats response imitation.arXiv preprint arXiv:2603.03303, 2026

Shirley Wu, Evelyn Choi, Arpandeep Khatua, Zhanghan Wang, Joy He-Yueya, Tharindu Cyril Weerasooriya, Wei Wei, Diyi Yang, Jure Leskovec, and James Zou. HumanLM: Simulating users with state alignment beats response imitation.arXiv preprint arXiv:2603.03303, 2026. URLhttps://arxi...

2026
[2004]

doi:10.1093/pan/mph024

doi
[2019]

URL https://cdn.openai.com/better-language-models/language_models_ are_unsupervised_multitask_learners.pdf
[2024]

URLhttps://openreview.net/forum?id=PXD3FAVHJT

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Christopher H. Achen. Mass political attitudes and the survey response.American Political Science Review, 69(4):1218–1231, 1975. doi:10.2307/1955282

work page doi:10.2307/1955282 1975

[2] [2]

Alwin and Jon A

Duane F. Alwin and Jon A. Krosnick. The reliability of survey attitude measurement: The influence of question and respondent attributes.Sociological Methods & Research, 20(1): 139–181, 1991. doi:10.1177/0049124191020001005

work page doi:10.1177/0049124191020001005 1991

[3] [3]

Argyle, Ethan C

Lisa P. Argyle, Ethan C. Busby, Nancy Fulda, Joshua R. Gubler, Christopher Rytting, and David Wingate. Out of one, many: Using language models to simulate human samples.Political Analysis, 31(3):337–351, 2023. doi:10.1017/pan.2023.2

doi:10.1017/pan.2023.2 2023

[4] [4]

Bettman, Mary Frances Luce, and John W

James R. Bettman, Mary Frances Luce, and John W. Payne. Constructive consumer choice processes.Journal of Consumer Research, 25(3):187–217, 1998. doi:10.1086/209535

doi:10.1086/209535 1998

[5] [5]

Eckstein, Noémi Élteto, Thomas L

Marcel Binz, Elif Akata, Matthias Bethge, Franziska Brändle, Fred Callaway, Julian Coda-Forno, Peter Dayan, Can Demircan, Maria K. Eckstein, Noémi Élteto, Thomas L. Griffiths, Susanne Haridi, Akshay K. Jagadish, Li Ji-An, Alexander Kipnis, Sreejan Kumar, Tobias Ludwig, Marcel Mathony, Marcelo Mattar, and Eric Schulz. A foundation model to predict and capt...

doi:10.1038/s41586-025-09215-4 2025

[6] [6]

Clinton, Cassy Dorff, Brenton Kenkel, and Jennifer M

James Bisbee, Joshua D. Clinton, Cassy Dorff, Brenton Kenkel, and Jennifer M. Larson. Synthetic replacements for human survey data? the perils of large language models.Political Analysis, 32(4):401–416, 2024. doi:10.1017/pan.2024.5

doi:10.1017/pan.2024.5 2024

[7] [7]

An appraisal-based chain-of-emotion architecture for affective language model game agents.PLoS ONE, 19(5): e0301033, 2024

Maximilian Croissant, Madeleine Frister, Guy Schofield, and Cade McCall. An appraisal-based chain-of-emotion architecture for affective language model game agents.PLoS ONE, 19(5): e0301033, 2024. doi:10.1371/journal.pone.0301033

doi:10.1371/journal.pone.0301033 2024

[8] [8]

QLoRA: Efficient fine- tuning of quantized LLMs

Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. QLoRA: Efficient fine- tuning of quantized LLMs. InAdvances in Neural Information Processing Systems (NeurIPS), volume 36, 2023. URL https://proceedings.neurips.cc/paper_files/paper/2023/ hash/1feb87871436031bdc0f2beaa62a049b-Abstract-Conference.html

work page 2023