Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Can LLMs Simulate Human Behavioral Variability? A Case Study in the Phonemic Fluency Task

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 ↗

Pith's one-line read No LLM, individually or in an ensemble, reproduces the scope of human behavioral variability in a phonemic fluency task.

desk verdict Broad model sweep with a real design flaw: the count-constrained prompt makes the headline claim about LLM variability weaker than the abstract suggests. read the letter →

arxiv 2505.16164 v2 pith:HXZVXJNQ submitted 2025-05-22 cs.CL

Mengyang Qiu , Zoe Brisebois , Siena Sun This is my paper
classification cs.CL
keywords largelanguagemodelsphonemicfluencybehavioralvariabilitycognitivesimulationlexicaldiversityverbalZipf'slaw
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

This paper asks whether large language models can replace human participants in a cognitive experiment, using the phonemic fluency task in which people list words starting with F. It compares 34 models across 45 configurations with responses from 106 human participants. The central claim is that no model, alone or combined with others, reproduces the scope of human variability: humans produced 476 distinct words, the best model only 226, and every LLM's rank-frequency distribution was steeper than the human one. If this is right, LLMs systematically produce more rigid and convergent output than human behavior, which would undercut their use as stand-ins in behavioral research.

What carries the argument

The argument is carried by the phonemic fluency task itself paired with distributional indices of variability. Phonemic fluency requires an effortful, form-based search of the mental lexicon, which is unnatural in everyday language use, so it is a stringent probe of whether LLMs can mimic individual differences rather than central tendencies. The key measurements are the type-to-token ratio and the share of idiosyncratic types, the Zipf scaling exponent of the rank-frequency distribution, and word co-occurrence networks filtered with the Triangulated Maximally Filtered Graph. These measures are what separate human output's long tail of rare words from the LLMs' steeper, more convergent distributions.

What would settle it

Run the same 106 demographic prompts but omit the 'number of correct responses' line; if any single model or ensemble then produces a unique-type count near 476 or a Zipf slope statistically indistinguishable from 0.89, the claim that LLMs cannot reproduce the scope of human variability would be falsified.

Watch

Extended reading notes

Core claim

The paper's core discovery is that the variability of human word production in a phonemic fluency task is not approximated by any tested LLM. Claude 3.7 Sonnet came closest on averages and on which linguistic features predicted production, but it generated fewer than half of the human unique word types (226 vs. 476), a lower type-to-token ratio (0.13 vs. 0.27), and fewer idiosyncratic words (73 vs. 201). Every LLM followed Zipf's law with a steeper scaling exponent than humans (alpha 1.19 to 1.53 vs. 0.89), meaning high-frequency words dominated more and the long tail of rare words was thinner. Word co-occurrence networks also differed structurally: human output formed tighter local clusters with weaker global integration, while Claude's network was more evenly connected, and the two networks' similarity matrices correlated only modestly. Ensembling outputs from a random mix of models across 1,000 simulations did not recover human diversity, because model vocabularies overlapped so heavily (mean overlap 0.74).

Load-bearing premise

The conclusion rests on the assumption that giving the model the participant's exact number of correct responses does not itself shrink the variety of words it produces; if models were more diverse when free to choose their own stopping point, the reported gap would be partly an artifact of the prompt.

Editorial extensions

If this is right

  • Model selection matters: newer models and thinking-enabled modes can reduce lexical diversity, so evaluations should cover multiple providers and versions.
  • Ensemble sampling across diverse models is not a shortcut to human-like variability, because providers share a largely common word pool.
  • LLMs may be better treated as baselines that capture central tendencies, against which human uniqueness and flexibility can be measured.
  • Using LLMs as substitutes for human participants in fluency-based or similar open-ended cognitive tasks would likely underrepresent population variability.

Reading between the lines

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

  • Because the prompt supplied each participant's exact response count, the observed rigidity could partly be an artifact of that constraint; a version that lets models choose their own stopping point could yield more diverse output.
  • The high vocabulary overlap across providers suggests training-data and alignment pressures converge on high-frequency, prototypical responses, which may affect any open-ended generation task where rare or idiosyncratic content is valued.
  • A concrete extension would be to condition prompts on richer participant traits such as profession or reading history, and test whether that restores variability without sacrificing adherence to response counts.
  • If the network difference reflects true retrieval dynamics, one could test it behaviorally by comparing human switch costs or reaction times with LLM token-level latencies, though such a comparison is not in this paper.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper asks whether large language models can reproduce the inter-participant behavioral variability seen in a phonemic fluency task. Using 106 human participants who generated F-words, the authors prompt 34 LLMs in 45 configurations to role-play each participant, providing age, education, and the participant's exact number of correct responses in the prompt. They then compare participant-level response counts, lexical diversity (TTR, ITTTR), item-level Zipf distributions and linguistic correlates, word co-occurrence networks, and ensemble mixtures. The headline finding is that no single model or ensemble matches human-level variability; Claude 3.7 Sonnet is the most human-like but still produces fewer than half the unique types. The paper concludes that LLMs are systematically more rigid and convergent than humans and cautions against using them as substitutes for human participants.

Significance. The question of whether LLMs can simulate human behavioral variability is timely and important for cognitive modeling claims. The paper's strengths include unusually broad model coverage (34 models, 45 configurations), grounding in real participant metadata, and multiple complementary outcome measures (TTR, Zipf slopes, network metrics, ensemble sampling). The empirical finding that LLM Zipf slopes are steeper than human slopes and that lexicon overlap across models is very high is a concrete, falsifiable regularity. However, the central claim currently rests on a design that injects the exact human response count into every prompt and then uses adherence to that injected count as both a success criterion and a selection filter. If the diversity gap persists when models are allowed to choose their own stopping point or are otherwise not constrained by the exact count, the paper's conclusion would be substantially strengthened; if not, the observed rigidity may be an artifact of the instruction rather than an intrinsic LLM limitation.

major comments (3)
  1. [Section 3.1 and Appendix A] The design is circular in a way that is load-bearing for the central claim. The prompt (Figure A1) includes 'Number of correct responses: {num_correct}', and Section 3.1 evaluates adherence using MAE against that same number, retaining only configurations with MAE≤1.69 and no outlier where |LLM−Human|>5. Low MAE is therefore a measure of instruction-following, not evidence of human-like production rates, and the retention filter selects on the injected value. Because every analyzed output is count-matched, the observed low TTR and steep Zipf slopes could be a direct consequence of count-constrained generation rather than an intrinsic property of LLMs. Appendix A reports that without the count, pilot models overproduced; yet overproduction may reflect deeper lexical exploration that yields more unique and idiosyncratic types. To support the conclusion that no LLM reproduces human variability, the paper needs a control condition without the count injection (or with a different target count) comparing TTR, ITTTR, and Zipf slopes, or an empirically grounded argument that the count constraint does not suppress diversity.
  2. [Section 3.1 and Table A2] The conclusion that 'no LLM, individually or in ensemble, reproduces the scope of human variability' is only demonstrated for the 21 configurations that adhered exactly to the injected counts. The 12 overproducing configurations (e.g., GPT-4 Turbo with MAE=57.17, o3 with MAE=35.39) are excluded from all participant-level and item-level analyses. If overproduction tends to generate longer, more exploratory lists with more unique and idiosyncratic types, then excluding those configurations biases the diversity comparison in the direction of the paper's conclusion. The authors should either include these configurations in the variability analyses (e.g., by token-matching or by analyzing their diversity directly) or explicitly restrict the abstract and Section 7 claims to count-constrained, instruction-following configurations. As written, the claim 'no LLM' is too broad given the exclusions.
  3. [Section 2.2 and Appendix A] The paper does not report any sampling parameters: temperature, top-p, or random seed are absent, and the 'thinking' / 'reasoning effort' modes are named but not defined numerically. Lexical diversity in LLM generation is strongly sensitive to sampling temperature; with zero temperature, repeated prompts can yield nearly identical outputs, which would artificially lower TTR and idiosyncratic type counts. Without reporting these parameters, the results are not reproducible, and the reader cannot determine whether the diversity gap reflects a model's capability or the authors' choice of decoding strategy. The authors should provide the exact API parameters used, or at least state that default parameters were used, and ideally vary temperature to show that the diversity gap is robust across sampling settings.
minor comments (5)
  1. [Section 5.2] The z-tests for deviation from random networks report negative z-values (e.g., human: z=−2460.1 for CC) while the text states that the networks exhibit 'substantially higher CCs and slightly higher ASPLs than random networks.' The sign is internally inconsistent; either the sign convention is reversed or the interpretation is incorrect. Please correct this discrepancy.
  2. [Section 6.2] The ensemble type range is reported as 143–214, but Claude 3.7 Sonnet alone produced 226 types, which is above the upper bound. The phrase 'values that fall within or near the upper tail of the ensemble distribution for types' is inaccurate for a value that exceeds the maximum; this should be reworded.
  3. [Table 1] Table 1 appears to list only a subset of the 21 retained configurations (the table shows 11 model rows), but the text refers to all retained configurations. The selection criterion for inclusion in Table 1 is not stated. If the table is intended to be representative or to show only the most diverse models, that should be explicitly indicated.
  4. [Section 2.1] The human data come from Qiu and Johns (2021), which involved two experiments; the paper should clarify whether the 106 participants are pooled across both experiments and whether experiment membership was included as a covariate or otherwise controlled for, since the two experiments may have had different procedures.
  5. [Appendix A] The prompt uses 'Highest degree: {education}' while Section 2.1 describes education in years of schooling (M=14.92, SD=2.01). These are different operationalizations; please clarify how the 'education' variable was mapped to the prompt field, since the mapping could affect the fidelity of the role-play.

Circularity Check

1 steps flagged · score 6.0 of 10

Count-constrained prompting and retention filters make low response-count error trivial and confine variability claims to count-matched outputs.

  1. fitted input called prediction [Section 3.1 (Number of responses) and Appendix A, Figure A1 prompt]
    "To evaluate whether LLMs adhered to the performance constraints specified in the prompt, we computed the Mean Absolute Error (MAE) between LLM-generated and human response counts for each configuration. Configurations with MAE≤1.69 (i.e., within 10% of the human mean) were considered to have successfully simulated human-like production rates. ... [Prompt includes:] Number of correct responses: {num_correct}"

    The MAE success criterion compares each model's response count against the exact count already inserted into the prompt as 'Number of correct responses: {num_correct}'. A model that simply outputs the stated number of words achieves MAE=0 by construction, so 'successfully simulated human-like production rates' reduces to instruction-following. The same injected count is then used as the outlier filter (|LLM−Human|>5) that retains only 21 of 33 count-matched configurations for all diversity, Zipf, network, and ensemble analyses. Overproducing configurations (e.g., GPT-4 Turbo, MAE=57.17; o3, MAE=35.39) are excluded, so the central claim that no LLM or ensemble reproduces the scope of human behavioral variability is demonstrated only for outputs already forced to match the target counts.

full rationale

The paper's lexical-diversity, Zipf, and network comparisons are not directly derived from the prompt inputs and retain independent content; a count-matched design does not by itself force TTR=0.13 versus 0.27 or alpha=1.19 versus 0.89. However, the participant-level production-rate comparison is target leakage: the human response count is both the independent variable in the prompt and the dependent variable in the MAE criterion, and the same count-based filter selects the configurations used for every downstream analysis. Appendix A confirms that the count was added deliberately to prevent overproduction, making the low MAE a measure of prompt adherence rather than evidence of human-like variability. Self-citations (e.g., Qiu and Johns 2021 data) are not load-bearing here because the dataset is externally available and the cited results are not used to forbid alternatives. Overall, a core component of the variability claim is circular, but the lexical and structural findings are not fully reduced to the input, so the appropriate score is partial circularity rather than complete equivalence.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a public human dataset and several analytical choices. The main ledger items are hand-picked inclusion thresholds and the domain assumption that feeding the outcome count into the prompt is a valid simulation strategy. No new theoretical entities are introduced; the paper's contribution is empirical.

free parameters (3)
  • MAE inclusion threshold = 1.69 (10% of the human mean of 16.89)
    Configurations with mean absolute error below this threshold were called human-like and retained; the value is chosen by hand and determines which models enter the diversity comparisons.
  • Outlier exclusion threshold = 5 responses
    Configurations with any simulation deviating by more than 5 from the prompted human count were excluded from participant and item analyses, which removes 12 configurations from the core comparisons.
  • Minimum unique types for item-level analyses = 100 unique types
    Only configurations producing at least 100 unique types were included in the Zipf and linguistic regression analyses in Section 4, narrowing the set of models compared.
assumptions (4)
  • domain assumption The Qiu and Johns (2021) MTurk dataset, including Google Speech-to-Text transcription with manual review, accurately represents human phonemic fluency variability.
    Section 2.1; all human benchmarks, 476 types, TTR 0.27, and Zipf slope 0.89, come from this dataset, so any transcription or sampling error propagates into every comparison.
  • ad hoc to paper Age, education, and the participant's number of correct responses are sufficient information to make an LLM simulate that participant.
    Section 2.2 and Appendix A; the number of correct responses is the outcome variable, so this assumption is load-bearing and partly circular.
  • standard math Co-occurrence based cosine networks filtered with TMFG reveal the retrieval structure of fluency behavior.
    Section 5.1; the network difference conclusion depends on this established but nontrivial modeling choice.
  • standard math Zipf's law, estimated by log-log linear regression, is a valid summary of production frequency distributions.
    Section 4.1; the steeper-slope conclusion depends on this model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can LLMs Simulate Human Behavioral Variability? A Case Study in the Phonemic Fluency Task." pith.science (2026). https://pith.science/paper/HXZVXJNQ

@misc{pith2026250516164,
  author       = {Pith},
  title        = {Pith review of: Can LLMs Simulate Human Behavioral Variability? A Case Study in the Phonemic Fluency Task},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HXZVXJNQ}},
  note         = {Machine review of arXiv:2505.16164}
}
read the original abstract

Large language models (LLMs) are increasingly explored as substitutes for human participants in cognitive tasks, but their ability to simulate human behavioral variability remains unclear. This study examines whether LLMs can approximate individual differences in the phonemic fluency task, where participants generate words beginning with a target letter. We evaluated 34 distinct models across 45 configurations from major closed-source and open-source providers, and compared outputs to responses from 106 human participants. While some models, especially Claude 3.7 Sonnet, approximated human averages and lexical preferences, none reproduced the scope of human variability. LLM outputs were consistently less diverse, with newer models and thinking-enabled modes often reducing rather than increasing variability. Network analysis further revealed fundamental differences in retrieval structure between humans and the most human-like model. Ensemble simulations combining outputs from diverse models also failed to recover human-level diversity, likely due to high vocabulary overlap across models. These results highlight key limitations in using LLMs to simulate human cognition and behavior.

Figures

Figures reproduced from arXiv: 2505.16164 by the authors.

Figure 1
Figure 1. Production frequency distribution by rank for human participants (red) and [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Clustering coefficient and average short [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Distributions of variability metrics across [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Analysing Moral Bias in Finetuned LLMs through Mechanistic Interpretability

    cs.CL 2025-10 conditional novelty 5.0 of 10

    The Knobe effect in fine-tuned LLMs is localized to mid-to-late transformer layers and can be removed by patching in pretrained activations at a single layer.

Reference graph

Works this paper leans on

8 extracted references · 7 canonical work pages · cited by 1 Pith paper

  1. [1]

    Can LLMs Simulate Human Behavioral Variability? A Case Study in the Phonemic Fluency Task

    Introduction Large language models (LLMs) have rapidly ad- vanced in recent years, achieving impressive per- formance across a wide range of natural language tasks. As a result, researchers have become in- creasingly interested in using LLMs as experimen- tal tools in cognitive and behavioral science. Some even propose that LLMs could replace human par- t...

    work page Pith review arXiv 2022

  2. [2]

    Method 2.1. Human data source We used phonemic fluency data collected by Qiu and Johns (2021) 1, who investigated noun- and verb-based semantic fluency across two experi- ments (see also Qiu et al., 2021). In both experi- ments, participants first completed a phonemic flu- ency task using the letterFas a familiarization trial before proceeding to semantic...

    work page 2021

  3. [3]

    Number of responses Human participants produced an average of 16.89 correct responses within the one-minute time con- straint (SD= 4.84 )

    Participant-Level Analysis 3.1. Number of responses Human participants produced an average of 16.89 correct responses within the one-minute time con- straint (SD= 4.84 ). To evaluate whether LLMs adhered to the performance constraints specified in the prompt, we computed the Mean Absolute Error (MAE) between LLM-generated and human response counts for eac...

    work page 2021

  4. [4]

    long tail

    Item-Level Analysis 4.1. Distribution of production frequency Previous studies have observed that word produc- tion in verbal fluency tasks typically follows Zipf’s law, a type of power-law relationship where word frequency is inversely proportional to its rank, ex- pressed as f(r)∝1/r α, where f(r) denotes the frequency of a word at rank r and α is the s...

    work page 2020

  5. [5]

    Claude consistently exhibited the most human-like patterns–response count, variability, production frequency distribution, and linguistic predictors of word choice

    Network Analysis Building on our earlier findings, we focused the net- work analysis exclusively on comparing human and Claude 3.7 Sonnet outputs. Claude consistently exhibited the most human-like patterns–response count, variability, production frequency distribution, and linguistic predictors of word choice. 5.1. Network construction approach To better ...

    work page 2016

  6. [6]

    LLM Ensemble Analysis The preceding analyses demonstrate that no single LLM configuration captures the full scope of human behavioral variability. A natural question is whether combining outputs from multiple models—each po- tentially drawing on different architectures, training data, and alignment strategies—could better ap- proximate the distributional ...

    work page 2024

  7. [7]

    Discussion and Conclusion This study evaluated whether LLMs can simu- late the behavioral variability observed in human phonemic fluency performance. Across 34 models and 45 configurations, no model—individually or in ensemble—captured the full scope of human vari- ability, despite many producing plausible average response counts and showing sensitivity t...

    arXiv 2021

  8. [1999]

    In a verbal fluency task , you will be asked to say as many words as you can think of that conform to a specified criterion within 1 minute

    Normative data stratified by age and edu- cation for two measures of verbal fluency: FAS and animal naming.Archives of Clinical Neu- ropsychology, 14(2):167–177. Sean Trott. 2024. Large language models and the wisdom of small crowds.Open Mind, 8:723–738. Dolores Villalobos, Lucía Torres-Simón, Javier Pa- cios, Nuria Paul, and David Del Río. 2023. A system...

    work page 2025

Pith tools

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