Breaking Homogeneity: Diversifying Persona Sets for Creative LLM Outputs

Sang Bin Moon; Nicole Cho; Daniel Borrajo; Sumitra Ganesh; Abolfazl Hashemi

cs.CLSep 24, 2026

Breaking Homogeneity: Diversifying Persona Sets for Creative LLM Outputs

Authors: Sang Bin Moon, Nicole Cho, Daniel Borrajo, Sumitra Ganesh, Abolfazl Hashemi

Organizations: School of Electrical and Computer Engineering Purdue University West Lafayette, IN 47909, USA · J.P. Morgan AI Research New York, NY 10172, USA

Abstract

Language models often produce homogeneous responses to open-ended tasks; such homogeneity can spawn groupthink-the convergence of ideas toward a singular and potentially suboptimal decision. We formulate persona diversification as a set-level conditioning problem and study two orthogonal design choices: selecting versus generating personas, and space-filling versus frontier-seeking diversity. We instantiate this design space with four methods spanning coverage and dispersion subset selections, uniform-coverage sampling, and evolutionary persona generation. Evaluations on the Alternative Uses Task (AUT), Infinity-Chat, and Divergent Association Task (DAT) show the benefits of the proposed methods across tasks and creativity objectives. On AUT, evolutionary persona generation increases response diversity by 78.8%, originality by 26.1%, flexibility by 49.5%, and holistic creativity by 13.9% over task-only prompting, while maintaining 98.5% validity; on Infinity-Chat, it nearly doubles persona-induced response separation relative to random personas. Moreover, evolutionary personas compose with creativity-optimized prompting, further increasing its response diversity by 18.6% and creativity by 6.3%. These results establish persona-set geometry as a task-agnostic mechanism for eliciting divergent LLM outputs, and support persona diversification as a reusable complement to prompt optimization.

Figures & tables

Figure 1: Persona diversification design space. Two principles of diversity (columns) are each instantiated at two levels of control (rows). The space-filling principle is implemented by Coverage and UC-MCMC, whereas the frontier-seeking principle is instantiated by Dispersion and evolutionary TextGrad . Three algorithms have formal guarantees for persona-space objectives and sampling targets, while evolutionary generation is empirical.

Figure 2: Persona conditioning diversifies Alternative Uses Task responses. Left: Example uses for a tin can under task-only prompting (top) and persona conditioning (bottom), illustrating a broader range of ideas. Right: Mahalanobis-whitened PCA projection of response embeddings across persona conditions, with convex hulls showing their semantic spread.

Figure 4: Persona set and response dispersion on AUT (left) and Infinity-Chat (right). Every points represent different configurations. Arrows connect six matched Coverage

\rightarrow

Dispersion contrasts while holding candidate-pool and distance metric fixed. Dashed lines show the linear fits and annotations report configuration-level Pearson correlations and 90% confidence intervals.

Variant	Diversity	Originality	Flexibility	Creativity	Utility	Validity
Alternative-Use	$1.42_{\pm 0.09}$	$10.79_{\pm 0.32}$	$.288_{\pm.022}$	$2.72_{\pm 0.09}$	$4.30_{\pm 0.05}$	$.984_{\pm.014}$
Random	$1.54_{\pm 0.12}$	$11.89_{\pm 0.48}$	$.331_{\pm.024}$	$2.96_{\pm 0.04}$	$4.28_{\pm 0.02}$	$.993_{\pm.006}$
Dispersion (selection)	$1.96_{\pm 0.17}$	$12.00_{\pm 0.28}$	$.357_{\pm.048}$	$2.92_{\pm 0.04}$	$4.23_{\pm 0.09}$	$.986_{\pm.010}$
Evolution (generation)	$\mathbf{2.53}_{\pm 0.20}$	$\mathbf{13.61}_{\pm 0.17}$	$\mathbf{.431}_{\pm.045}$	$\mathbf{3.10}_{\pm 0.03}$	$4.08_{\pm 0.06}$	$.985_{\pm.012}$
Creativity-enhanced	$5.15_{\pm 0.16}$	$17.15_{\pm 0.40}$	$.503_{\pm.056}$	$3.19_{\pm 0.09}$	$3.38_{\pm 0.07}$	$.968_{\pm.011}$
DMAD	$2.79_{\pm 0.22}$	$13.90_{\pm 0.49}$	$.372_{\pm.025}$	$2.89_{\pm 0.07}$	$4.17_{\pm 0.11}$	$.977_{\pm.009}$

Table 2: Response creativity evaluation on AUT. Diversity is the hull extent in a 5-D PCA projection of Mahalanobis-whitened response embeddings, Originality is the mean Mahalanobis distance to the common-use reference set, Flexibility is the number of distinct categories per valid use, and Creativity is the rank-normalized holistic score ( Góes et al., 2023 ) from an independent LLM judge. Subscripts give 95% confidence-interval half-widths across five random seeds. Utility ( Stevenson et al., 2022 ) and validity are controls and are not bolded. Bold = best results within each group.

	Infinity-Chat				DAT
Variant	Homog. $\downarrow$	Sep. $\uparrow$	Flex. $\uparrow$	Val.	Div. $\uparrow$	Flex. $\uparrow$	Val.
Baseline Task	$0.946_{\pm 0.002}$	–	$1.39_{\pm 0.02}$	$.950_{\pm.015}$	$89.13_{\pm 0.49}$	$1.79_{\pm 0.01}$	$.996_{\pm.005}$
Random	$0.864_{\pm 0.005}$	$0.068_{\pm 0.003}$	$2.17_{\pm 0.05}$	$1.000_{\pm.000}$	$89.15_{\pm 1.24}$	$1.83_{\pm 0.04}$	$.993_{\pm.011}$
Dispersion (selection)	$0.840_{\pm 0.004}$	$0.085_{\pm 0.006}$	$2.40_{\pm 0.05}$	$.986_{\pm.007}$	$89.39_{\pm 0.26}$	$1.84_{\pm 0.01}$	$.995_{\pm.006}$
Evolution (generation)	$\mathbf{0.798}_{\pm 0.006}$	$\mathbf{0.126}_{\pm 0.006}$	$\mathbf{2.81}_{\pm 0.08}$	$.992_{\pm.005}$	$\mathbf{89.48}_{\pm 0.32}$	$\mathbf{1.85}_{\pm 0.01}$	$.996_{\pm.004}$
Creative	–	–	–	–	$86.05_{\pm 0.43}$	$1.69_{\pm 0.01}$	$.977_{\pm.010}$

Table 3: Response creativity evaluation on Infinity-Chat (IC) and DAT. Homogeneity is mean pairwise cosine similarity, Separation is mean between-persona distance minus mean within-persona distance, Flexibility is the Vendi score, measuring effective diversity under the embedding similarity kernel, and Divergence is the DAT semantic score. Subscripts give 95% confidence-interval half-widths across five random seeds. Bold = best results within each group. Control (validity) and undefined combinations (–) are not bolded.

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

Variant	Persona	Description
Common-Use	–	Góes et al. (2023) common-use (nn)
Alternative-Use	–	Góes et al. (2023) creative (nc)
Creativity-Enhanced	–	Góes et al. (2023) expert (bs)
Random	✓	$5$ chosen uniformly at random
Typical	✓	$5$ density-ranked modes
Selection (ours)	✓	maximize dispersion on base pool

Appendix

Table 4: Variants, from persona-free baselines to our diversity-aware selection and generation. Persona variants all inject into the Alternative-Use prompt; Selection and Evolution share one selection operator and differ only in the candidate pool.

Metric	Basis	Description
Validity	Human / LLM judge	Proportion of admissible, well-formed uses (gate; control)
Diversity	Embedding	Mean pairwise distance among a variant’s uses ( $\uparrow$ )
Originality	Embedding	Mean semantic or directed Chamfer distance from uses
		to the object concept or common uses ( $\uparrow$ )
Flexibility	Category	(Effective) number (ratio) of distinct use categories ( $\uparrow$ )
Surprise	Category	(Effective) share of uses in categories beyond common uses ( $\uparrow$ )

Appendix

Table 5: Creativity and diversity metrics. Embedding metrics use the fixed encoder

\phi

(EmbeddingGemma) with cosine, L2, or Mahalanobis dissimilarity; category metrics use both human clustering and Vendi-based effective counts; LLM-judge metrics use an independent Qwen3.6-27B judge.

\uparrow

: higher indicates a larger value. Validity is a gate. Utility is the effectiveness dimension required alongside originality for creativity ( Runco & Jaeger, 2012 ) and is reported separately to expose trade-offs.

Dim.	Pair	$n$	ICC $(C,1)$	95% CI	QWK	$\rho$	Exact	Within-1	Paper ICC
Orig.	LLM vs. rater 1	139	0.480	$[0.341,0.609]$	0.447	0.553	0.547	0.878	0.78
Orig.	LLM vs. rater 2	139	0.575	$[0.452,0.693]$	0.574	0.599	0.504	0.942	0.78
Orig.	LLM vs. mean	139	0.568	$[0.443,0.682]$	0.481	0.635	0.302	0.878	0.78
Orig.	Human ceiling	139	0.677	$[0.577,0.754]$	0.609	0.669	0.475	0.993	0.78
Util.	LLM vs. rater 1	139	0.355	$[0.211,0.504]$	0.344	0.509	0.475	0.813	0.70
Util.	LLM vs. rater 2	139	0.430	$[0.292,0.568]$	0.406	0.597	0.532	0.806	0.70

Appendix

Table 6: AUT LLM-as-a-judge agreement on the valid Stevenson subset (

n{=}139

). Paper ICC is the human–human ICC reported by Stevenson et al.; Human ceiling is the reproduced rater01 – rater02 agreement on the released file. LLM scores are from Qwen3.6-27B with the faithful Stevenson prompts (including 0/99 codes).

Dim.	Pair	$n$	ICC $(C,1)$	95% CI	QWK	$\rho$	Exact	Within-1	Paper ICC
Orig.	LLM vs. rater 1	147	0.386	$[0.248,0.525]$	0.378	0.488	0.524	0.844	0.57
Orig.	LLM vs. rater 2	147	0.479	$[0.341,0.608]$	0.474	0.533	0.476	0.905	0.57
Orig.	LLM vs. mean	147	0.467	$[0.332,0.601]$	0.410	0.555	0.286	0.837	0.57
Orig.	Human ceiling	147	0.630	$[0.498,0.734]$	0.579	0.642	0.476	0.986	0.57
Util.	LLM vs. rater 1	147	0.336	$[0.199,0.472]$	0.316	0.499	0.449	0.769	0.68
Util.	LLM vs. rater 2	147	0.400	$[0.264,0.534]$	0.367	0.572	0.503	0.769	0.68

Appendix

Table 7: AUT LLM-as-a-judge agreement on all doubly rated Stevenson responses (

n{=}147

), including rows with invalid codes. Metrics as in Table 6 .

Metric	Value
$n$ scored pairs	750
Quality prompt	hivemind_quality_absolute_v1
Spearman $\rho$ (judge vs. human mean)	0.158
Pearson $r$ (judge vs. human mean)	0.278
ICC $(C,1)$	0.225
ICC 95% CI (bootstrap)	$[0.146,\,0.296]$

Appendix

Table 8: Infinity-Chat quality-judge agreement against the

25

-rater human mean on the Artificial Hivemind dense absolute-rating set. Exact / within-one / QWK compare the integer LLM score to the human mean rounded to the nearest integer.

Human mean bin	$n$	Invalid rate	Prompt
Good ( $\geq 4$ )	405	0.002	hivemind_validity_screen_v1
Mid ( $(2,4)$ )	341	0.023
Bad ( $\leq 2$ )	4	0.250
Overall	750	0.013

Appendix

Table 9: Infinity-Chat validity-gate tripwire on the same

750

pairs. The automated gate fails the protocol if it marks

{>}5\%

of human-good responses invalid; the observed rate is

0.25\%

Persona set	Dispersion			Vendi score
Persona set	Cosine	L2	Mahal.	Cosine	L2	Mahal.
Baseline personas
Task-conditioned	$0.130$	$0.510$	$13.42$	$1.81$	$2.21$	$1.21$
Individualized personas (base pool $=997$ )
Typical	$0.270$	$0.734$	$12.04$	$2.34$	$2.73$	$1.21$
Random	$0.248$	$0.703$	$13.67$	$2.60$	$2.92$	$1.25$

Appendix

Table 10: Persona diversity of the variants measured in Dispersion (equation 8 ) and Vendi score (equation 53 ) of each persona set (

k=5

), computed on persona embeddings under cosine, L2, and Mahalanobis dissimilarity. Dispersion represents how personas are spread far from each other (minimum pairwise distance within the set; higher is more spread). Vendi score represents the effective number of distinct clusters of personas (higher is more distinct). Methods are grouped by their algorithm categories: MPAQ is the baseline ( Jin et al., 2025 ) that generated persona for output diversity, Individualized personas represent baseline selection methods from the base pool of size 997, Selected personas represent our methods of persona selection according to various objective functions, and Generated personas represent our methods of persona generation, each selected from its expanded pool by exact Mahalanobis max–min dispersion. Bold = best per column within the selected and generated groups; baseline groups are not bolded.

Figure 5: Persona-to-response relationship on (a) the Alternative Uses Task and (b) Infinity-Chat. Points represent 14 analogous task-agnostic method configurations per benchmark; response hull extent and vendi score are averaged over seven AUT objects and five Infinity-Chat queries. Arrows connect six matched Coverage

\rightarrow

Dispersion contrasts while holding candidate-pool family and selector metric fixed. Dashed lines show configuration-level linear fits; annotations report Pearson correlations and descriptive 95% intervals.

Variant	All seven objects	Human-reference objects
Standard persona
Common-Use	$0.971_{\pm 0.018}$	$0.981_{\pm 0.015}$
Alternative-Use	$0.984_{\pm 0.014}$	$0.968_{\pm 0.025}$
Expert	$0.997_{\pm 0.006}$	$0.997_{\pm 0.007}$
Creativity-enhanced	$0.968_{\pm 0.011}$	$0.981_{\pm 0.022}$
ZS-CoT	$0.978_{\pm 0.009}$	$0.997_{\pm 0.007}$

Appendix

Table 11: Response validity: the proportion of generated uses admitted by the validity gate of Section D.2 , reported over all seven AUT objects and, for comparability with Stevenson , over the three objects with human reference data (

175

and

75

generated uses per variant, respectively). Every variants beyond Alternative-Use build prompts upon the generic Alternative-Use prompt unless its group states otherwise; the final two groups re-run our selection and generation methods on the Creativity-enhanced prompt of Góes et al. (2023) , showing that persona diversification composes with a stronger prompting strategy rather than competing with it. Cells give the admitted proportion averaged over objects. Subscripts denote the 95% confidence interval across 5 independent random seeds. Stevenson exists only for the human-reference objects, so its seven-object cell is omitted (—). Validity is a control rather than a target: values near

1

are expected and the informative signal is degradation rather than rank, so no value is bolded. Generated variants report their expanded pool size in parentheses.

	Mean pairwise distance			Convex hull volume
Variant	Cosine	L2	Mahal.	Cosine/L2	Mahal.
Standard persona
Common-Use	$0.155_{\pm 0.004}$	$0.526_{\pm 0.006}$	$15.50_{\pm 0.35}$	$0.103_{\pm 0.004}$	$2.17_{\pm 0.23}$
Alternative-Use	$0.188_{\pm 0.003}$	$0.590_{\pm 0.006}$	$11.97_{\pm 0.28}$	$0.172_{\pm 0.014}$	$1.42_{\pm 0.09}$
Expert	$0.179_{\pm 0.004}$	$0.580_{\pm 0.006}$	$14.57_{\pm 0.29}$	$0.161_{\pm 0.004}$	$2.36_{\pm 0.13}$
Creativity-enhanced	$0.179_{\pm 0.002}$	$0.592_{\pm 0.004}$	$\mathbf{20.24}_{\pm 0.56}$	$0.144_{\pm 0.010}$	$\mathbf{5.15}_{\pm 0.16}$

Appendix

Table 12: Response diversity across all seven AUT objects, reported two ways. Mean pairwise distance (equation 55 ) is the average cosine, L2 or Mahalanobis distance among the valid uses of a variant; convex hull volume (equation 55 ) is the volume of the convex hull those uses span in the same geometry, so it rewards a set that occupies a region rather than merely separating in the mean. Both are averaged over objects. Subscripts denote the 95% confidence interval across 5 independent random seeds, and higher is more diverse. Hull extent under cosine and L2 coincides to nine decimal places, the two metrics are monotone transforms of one another on unit-normalized embeddings, so it is reported once. Baseline and our method variants are grouped by the personas: standard persona represents LLMs default persona when no persona is given explicitly, baseline personas represent existing persona injection methods, individualized personas represent broad population of personas from PersonaHub Ge et al. (2024) ; Jiang et al. (2025) , while selected personas and generated personas represent our specific persona selection and generation algorithms. The final two groups re-run our selection and generation methods on the Creativity-enhanced prompt, showing that persona diversification composes with a stronger prompting strategy. Bold marks the best value per column within a method category, with tied values bolded jointly. All personas, including gibberish, are injected to the alternative use prompt unless their group states otherwise.

	Mean pairwise distance			Convex hull extent
Variant	Cosine	L2	Mahal.	Cosine/L2	Mahal.
Standard persona
Common-Use	$0.182_{\pm 0.005}$	$0.583_{\pm 0.010}$	$15.49_{\pm 0.66}$	$0.124_{\pm 0.009}$	$2.16_{\pm 0.38}$
Alternative-Use	$\mathbf{0.195}_{\pm 0.003}$	$\mathbf{0.601}_{\pm 0.005}$	$11.38_{\pm 0.39}$	$0.173_{\pm 0.012}$	$1.20_{\pm 0.13}$
Expert	$0.185_{\pm 0.005}$	$0.590_{\pm 0.005}$	$13.76_{\pm 0.32}$	$\mathbf{0.170}_{\pm 0.012}$	$2.01_{\pm 0.26}$
Creativity-enhanced	$0.174_{\pm 0.006}$	$0.583_{\pm 0.009}$	$\mathbf{19.54}_{\pm 0.31}$	$0.138_{\pm 0.006}$	$\mathbf{5.06}_{\pm 0.40}$

Appendix

Table 13: Response diversity restricted to the three objects with human reference data (book, fork, tin can), so that all variants are directly comparable to Stevenson . Cells, bolding, and column conventions are as in Table 12 .

	To object			To common uses
Variant	Cosine	L2	Mahal.	Cosine	L2	Mahal.
Standard persona
Common-Use	$0.142_{\pm 0.002}$	$0.525_{\pm 0.004}$	$15.28_{\pm 0.16}$	—	—	—
Alternative-Use	$0.170_{\pm 0.001}$	$0.580_{\pm 0.002}$	$13.60_{\pm 0.14}$	$0.114_{\pm 0.008}$	$0.439_{\pm 0.019}$	$10.79_{\pm 0.32}$
Expert	$0.164_{\pm 0.002}$	$0.568_{\pm 0.004}$	$14.72_{\pm 0.17}$	$0.104_{\pm 0.003}$	$0.417_{\pm 0.010}$	$11.25_{\pm 0.40}$
Creativity-enhanced	$0.173_{\pm 0.004}$	$0.586_{\pm 0.007}$	$\mathbf{18.13}_{\pm 0.35}$	$\mathbf{0.174}_{\pm 0.006}$	$\mathbf{0.585}_{\pm 0.011}$	$\mathbf{17.15}_{\pm 0.40}$

Appendix

Table 14: Response originality across all seven AUT objects, measured both as distance to the prompted object and as directed distance to the common-use set, following equation 57 . Cells give the mean over valid uses, averaged over objects; higher is more original. Common-Use defines the common-use reference set

C_{o}

, so its distance to that set is zero by construction and is omitted (—). Subscripts denote the 95% confidence interval across 5 independent random seeds. The final two groups re-run our selection and generation methods on the Creativity-enhanced prompt, showing that persona diversification composes with a stronger prompting strategy. Bold marks the best value per column within the standard, selected and generated groups separately, with tied values bolded jointly; baseline and individualized personas are not bolded. Generated variants report their expanded pool size in parentheses.

	To object			To common uses
Variant	Cosine	L2	Mahal.	Cosine	L2	Mahal.
Standard persona
Common-Use	$0.182_{\pm 0.005}$	$0.583_{\pm 0.010}$	$15.49_{\pm 0.66}$	—	—	—
Alternative-Use	$\mathbf{0.195}_{\pm 0.003}$	$\mathbf{0.601}_{\pm 0.005}$	$11.38_{\pm 0.39}$	$0.173_{\pm 0.012}$	$1.20_{\pm 0.13}$	$10.79_{\pm 0.32}$
Expert	$0.185_{\pm 0.005}$	$0.590_{\pm 0.005}$	$13.76_{\pm 0.32}$	$\mathbf{0.170}_{\pm 0.012}$	$2.01_{\pm 0.26}$	$11.25_{\pm 0.40}$
Creativity-enhanced	$0.174_{\pm 0.006}$	$0.583_{\pm 0.009}$	$\mathbf{19.54}_{\pm 0.31}$	$0.138_{\pm 0.006}$	$\mathbf{5.06}_{\pm 0.40}$	$\mathbf{17.15}_{\pm 0.40}$

Appendix

Table 15: Response originality restricted to the three objects with human reference data (book, fork, tin can), so that all variants are directly comparable to Stevenson . Cells, bolding, and column conventions are as in Table 14 . Common-Use defines the common-use reference set

C_{o}

, so its distance to that set is zero by construction and is omitted (—).

	All seven objects		Human-reference objects
Variant	Flexibility	Vendi	Flexibility	Vendi
Standard persona
Common-Use	$0.160_{\pm 0.011}$	$0.086_{\pm 0.003}$	$0.214_{\pm 0.025}$	$0.094_{\pm 0.004}$
Alternative-Use	$0.288_{\pm 0.022}$	$0.096_{\pm 0.002}$	$0.284_{\pm 0.012}$	$0.099_{\pm 0.003}$
Expert	$0.300_{\pm 0.026}$	$0.094_{\pm 0.001}$	$0.320_{\pm 0.017}$	$0.096_{\pm 0.001}$
Creativity-enhanced	$\mathbf{0.503}_{\pm 0.056}$	$0.104_{\pm 0.002}$	$\mathbf{0.467}_{\pm 0.071}$	$0.101_{\pm 0.002}$

Appendix

Table 16: Response flexibility, reported over all seven AUT objects and, for comparability with Stevenson , over the three objects with human reference data (

175

and

75

generated uses per variant, respectively). Both measures are size-controlled: Flexibility is the number of distinct use categories and Vendi is the Vendi score equation 53 , each divided by the number of valid uses, so that variants admitting different numbers of uses remain comparable. Cells give the mean over objects. Subscripts denote the 95% confidence interval across 5 independent random seeds; higher is more flexible. Stevenson exists only for the human-reference objects, so its seven-object cells are omitted (—). The final two groups re-run our selection and generation methods on the Creativity-enhanced prompt, showing that persona diversification composes with a stronger prompting strategy. Bold marks the best value per column within the standard, selected and generated groups separately, with tied values bolded jointly; baseline, individualized and human rows are not bolded. Generated variants report their expanded pool size in parentheses.

	All seven objects		Human-reference objects
Variant	Surprise	Wtd. Vendi	Surprise	Wtd. Vendi
Standard persona
Common-Use	—	—	—	—
Alternative-Use	$0.506_{\pm 0.062}$	$0.094_{\pm 0.002}$	$0.397_{\pm 0.087}$	$0.093_{\pm 0.004}$
Expert	$0.415_{\pm 0.020}$	$0.093_{\pm 0.002}$	$0.317_{\pm 0.014}$	$0.098_{\pm 0.002}$
Creativity-enhanced	$\mathbf{0.837}_{\pm 0.067}$	$0.105_{\pm 0.002}$	$\mathbf{0.717}_{\pm 0.131}$	$0.102_{\pm 0.002}$

Appendix

Table 17: Response surprise, reported over all seven AUT objects and, for comparability with Stevenson , over the three objects with human reference data (

175

and

75

generated uses per variant, respectively). Both measures are size-controlled by the number of valid uses: Surprise is the proportion of uses judged surprising relative to the common-use reference set, and Wtd. Vendi is the novelty-weighted Vendi score equation 53 under the same normalization. Common-Use defines the reference set against which surprise is measured, so its surprise is zero by construction and is omitted (—); Stevenson exists only for the human-reference objects. Cells give the mean over objects. Subscripts denote the 95% confidence interval across 5 independent random seeds; higher is more surprising. The final two groups re-run our selection and generation methods on the Creativity-enhanced prompt, showing that persona diversification composes with a stronger prompting strategy. Bolding conventions are as in Table 16 .

	Stevenson et al. (2022)			Góes et al. (2023)
Variant	Originality	Surprise	Utility	Creativity (rank)
Standard persona
Common-Use	$1.68_{\pm 0.03}$	$1.34_{\pm 0.01}$	$4.86_{\pm 0.04}$	$2.02_{\pm 0.08}$
Alternative-Use	$2.63_{\pm 0.05}$	$2.23_{\pm 0.06}$	$4.30_{\pm 0.05}$	$2.72_{\pm 0.09}$
Expert	$2.28_{\pm 0.05}$	$2.01_{\pm 0.05}$	$4.39_{\pm 0.05}$	$2.71_{\pm 0.05}$
Creativity-enhanced	$\mathbf{3.17}_{\pm 0.06}$	$\mathbf{2.66}_{\pm 0.06}$	$3.38_{\pm 0.07}$	$\mathbf{3.19}_{\pm 0.09}$

Appendix

Table 18: LLM-judged creativity on the AUT across all seven objects. A Qwen3.6-27B judge, independent of the Gemma-4-31B generator, rates every use for Originality, Surprise and Utility on a

1

–

5

scale following Stevenson et al. (2022) , and for holistic creativity following Góes et al. (2023) ; we report the rank-normalized form of that holistic judgment, which correlates with its raw form at

r=0.996

. Cells give the mean over judged uses (at most

175

per variant). The judged set is every use not rejected outright by the validity gate of Section D.2 , admitted and uncertain uses, so it is slightly larger than the admitted set of Table 11 . Subscripts denote the 95% confidence interval across 5 independent random seeds. The final two groups re-run our selection and generation methods on the Creativity-enhanced prompt, showing that persona diversification composes with a stronger prompting strategy. Originality, Surprise and Creativity are target metrics, bolded per column within the standard, selected and generated groups separately (and within each creativity-enhanced group), with tied values bolded jointly. Utility is a control : conventional uses are the most useful by construction, so it necessarily peaks at Common-Use and is never bolded. Baseline, individualized and human rows are not bolded. Generated variants report their expanded pool size in parentheses.

	Stevenson et al. (2022)			Góes et al. (2023)
Variant	Originality	Surprise	Utility	Creativity (rank)
Standard persona
Common-Use	$1.84_{\pm 0.04}$	$1.45_{\pm 0.03}$	$4.84_{\pm 0.03}$	$2.29_{\pm 0.12}$
Alternative-Use	$2.37_{\pm 0.09}$	$2.07_{\pm 0.05}$	$4.39_{\pm 0.08}$	$2.78_{\pm 0.16}$
Expert	$2.18_{\pm 0.04}$	$2.00_{\pm 0.06}$	$4.40_{\pm 0.08}$	$2.83_{\pm 0.07}$
Creativity-enhanced	$\mathbf{3.05}_{\pm 0.18}$	$\mathbf{2.66}_{\pm 0.06}$	$3.52_{\pm 0.11}$	$\mathbf{3.19}_{\pm 0.09}$

Appendix

Table 19: LLM-judged creativity restricted to the three objects with human reference data (book, fork, tin can), so that all variants are directly comparable to Stevenson (at most

75

judged uses per variant). Scales, judge, bolding and control conventions are as in Table 18 .

Variant	Valid	Quality
Standard persona
Open-Ended Query	$0.950_{\pm 0.015}$	$4.84_{\pm 0.02}$
ZS-CoT	$1.000_{\pm 0.000}$	$4.99_{\pm 0.01}$
Step-Back	$1.000_{\pm 0.000}$	$4.98_{\pm 0.02}$
DMAD	$1.000_{\pm 0.000}$	$4.96_{\pm 0.04}$
Gibberish	$0.123_{\pm 0.013}$	$1.01_{\pm 0.01}$

Appendix

Table 20: Infinity-Chat 100 response validity and quality ( Jiang et al., 2026 ) . Each condition contributes

250

responses,

50

per open-ended item for prompt conditions and ten per persona per item for persona conditions over

k=5

personas and five items. Valid is the proportion of responses admitted by the validity gate; Quality is the mean judge rating on a

1

–

5

scale. Subscripts denote the 95% confidence interval across 5 independent random seeds. The final two groups re-run our selection and generation methods on the DMAD prompt. Both are controls rather than targets, validity near

1

is expected, and quality is the effectiveness dimension against which diversity gains are traded, so no value is bolded. Generated variants report their expanded pool size in parentheses.

Variant	Similarity $\downarrow$	Separation $\uparrow$	Trigram $\downarrow$
Standard persona
Open-Ended Query	$0.946_{\pm 0.002}$	—	$0.247_{\pm 0.004}$
ZS-CoT	$0.949_{\pm 0.003}$	—	$0.096_{\pm 0.012}$
Step-Back	$0.962_{\pm 0.003}$	—	$0.276_{\pm 0.007}$
DMAD	$\mathbf{0.929}_{\pm 0.002}$	—	$\mathbf{0.055}_{\pm 0.004}$
Gibberish †	$0.894_{\pm 0.012}$	$0.045_{\pm 0.008}$	$0.055_{\pm 0.022}$

Appendix

Table 21: Infinity-Chat 100 response homogeneity ( Jiang et al., 2026 ) . Similarity is the mean pairwise embedding similarity among a condition’s responses to the same item, averaged over items; Separation is the gain in mean between-persona distance over within-persona distance, and is undefined for conditions without personas (—); Trigram is mean trigram Jaccard overlap. Arrows give the favourable direction; bold marks the best per column within the standard, selected and generated groups, and within each DMAD group. † Gibberish is excluded from bolding because

90\%

of its responses fail the validity gate, leaving statistics computed on a tenth of the sample. Subscripts denote the 95% confidence interval across 5 independent random seeds. The final two groups re-run our selection and generation methods on the DMAD prompt. Generated variants report their expanded pool size in parentheses.

		Fluency		Flexibility
Variant	Tokens	Entropy $\uparrow$	Top- $10$ $\downarrow$	TTR $\uparrow$	Vendi $\uparrow$
Standard persona
Open-Ended Query	$31{,}170$	$\mathbf{5.98}_{\pm 0.02}$	$0.244_{\pm 0.002}$	$0.068_{\pm 0.001}$	$1.39_{\pm 0.02}$
ZS-CoT	$12{,}536$	$5.80_{\pm 0.03}$	$\mathbf{0.232}_{\pm 0.004}$	$0.099_{\pm 0.003}$	$1.40_{\pm 0.03}$
Step-Back	$11{,}900$	$5.69_{\pm 0.02}$	$0.241_{\pm 0.004}$	$0.093_{\pm 0.001}$	$1.28_{\pm 0.02}$
DMAD	$10{,}443$	$5.85_{\pm 0.02}$	$0.251_{\pm 0.002}$	$\mathbf{0.129}_{\pm 0.003}$	$\mathbf{1.56}_{\pm 0.01}$

Appendix

Table 22: Infinity-Chat 100 response fluency and flexibility ( Jiang et al., 2026 ) . Tokens is the number of word tokens a condition produced, given because the three lexical measures are corpus statistics and are biased downward by corpus size. Fluency : vocabulary entropy

H(\mathcal{W})

in nats and top-

10

concentration

C_{10}(\mathcal{W})

, the share of occurrences taken by the ten most frequent words. Flexibility : type–token ratio

TTR(\mathcal{W})

and the Vendi score equation 53 of the response embeddings, the effective number of distinct concepts; the latter is computed over response sets of equal size and so is unaffected by the token counts. Arrows give the favourable direction; bold marks the best per column within the standard, selected and generated groups, and within each DMAD group. Token counts are not bolded. † Gibberish is excluded from bolding because

90\%

of its responses fail the validity gate. Subscripts denote the 95% confidence interval across 5 independent random seeds. The final two groups re-run our selection and generation methods on the DMAD prompt. Generated variants report their expanded pool size in parentheses.

Variant	Valid words
Standard persona
Non-Divergent Association	$0.998_{\pm 0.003}$
Divergent Association	$0.996_{\pm 0.005}$
Base-Instruction	$1.000_{\pm 0.000}$
Random-Instruction	$1.000_{\pm 0.000}$
Creative	$0.977_{\pm 0.010}$

Appendix

Table 23: DAT response validity: the proportion of the

n_{\mathrm{raw}}=350

words generated per condition that are admitted by the validity gate, which discards words repeated within a completion or absent from the reference vocabulary. The final two groups re-run our selection and generation methods on the DMAD prompt, showing that persona diversification composes with a stronger prompting strategy rather than competing with it. Validity is a control rather than a target, values near

1

are expected and the informative signal is degradation rather than rank, so no value is bolded. No condition trades word validity for divergence. Subscripts denote the 95% confidence interval across 5 independent random seeds. Generated variants report their expanded pool size in parentheses.

	Diversity		Fluency		Flexibility
Variant	Disp. $\uparrow$	Hull $\uparrow$	Entropy $\uparrow$	Top- $10$ $\downarrow$	TTR $\uparrow$	Vendi $\uparrow$
Standard persona
Non-Divergent Association	$7.05_{\pm 0.68}$	—	$2.89_{\pm 0.05}$	$0.68_{\pm 0.05}$	$0.07_{\pm 0.00}$	$1.42_{\pm 0.01}$
Divergent Association	$12.30_{\pm 0.75}$	$1.28_{\pm 0.06}$	$3.63_{\pm 0.04}$	$0.53_{\pm 0.02}$	$0.19_{\pm 0.01}$	$1.79_{\pm 0.01}$
Base-Instruction	$13.30_{\pm 0.44}$	$0.59_{\pm 0.06}$	$2.96_{\pm 0.02}$	$0.76_{\pm 0.01}$	$0.10_{\pm 0.01}$	$1.54_{\pm 0.00}$
Random-Instruction	$14.19_{\pm 0.70}$	$0.79_{\pm 0.11}$	$3.06_{\pm 0.06}$	$0.71_{\pm 0.03}$	$0.11_{\pm 0.01}$	$1.63_{\pm 0.01}$

Appendix

Table 24: DAT response creativity over valid words. Diversity : mean Mahalanobis dispersion and Mahalanobis hull extent. Fluency : vocabulary entropy

H(\mathcal{W})

in nats, and top-

10

concentration

C_{10}(\mathcal{W})

, the share of occurrences taken by the ten most frequent words. Flexibility : type–token ratio

TTR(\mathcal{W})

, and the Vendi score equation 53 of the word embeddings, the effective number of distinct concepts. The final two groups re-run our selection and generation methods on the DMAD prompt, showing that persona diversification composes with a stronger prompting strategy rather than competing with it. Subscripts denote the 95% confidence interval across 5 independent random seeds. Arrows give the favorable direction; bold marks the best value per column within the standard, selected and generated groups separately, and within each DMAD group. Generated variants report their expanded pool size in parentheses.

Variant	DAT score	Percentile
Standard persona
Non-Divergent Association	$46.35_{\pm 2.30}$	$0.6$
Divergent Association	$89.13_{\pm 0.49}$	$96.5$
Base-Instruction	$77.46_{\pm 0.29}$	$38.9$
Random-Instruction	$81.02_{\pm 0.35}$	$63.9$
Creative	$86.05_{\pm 0.43}$	$88.9$

Appendix

Table 25: DAT score and human-reference percentile. The score is the mean over completions of

100\times

the average pairwise semantic distance among the first seven valid unique nouns, following Olson et al. (2021) ; the subscript is the 95% confidence interval across 5 independent random seeds. Percentile is the mean rank of a condition’s completions within the

8{,}572

human completions of Olson et al. (2021) , which by construction places Olson at

50

. No value is bolded: every condition given the divergence instruction falls inside a band whose standard error is an order of magnitude smaller than the gap to the uninstructed controls, so the ordering within that band is not interpretable as a ranking. Generated variants report their expanded pool size in parentheses.

Output Diversity in Language Models

See all 32 papers

Are LLMs becoming similarly creative? Evidence from three years of models
Aug 19, 2026Nirav Patel, Josiah Crossman, Eva Aggarwal +1Homogenization in Language Models Computational Creativity
Department of Computer Science, Duke University · Department of Electrical & Computer Engineering, Duke University
Examining and Addressing Barriers to Diversity in LLM-Generated Ideas
Feb 23, 2026Yuting Deng, Melanie Brucks, Olivier ToubiaDiverse Text Generation LLM Prompting
Columbia Business School
PERSONAWEAVER: Controllable Diversity Beyond Conventional Archetypes in Procedural Character Generation
Sep 22, 2026Maan Qraitem, Kate Saenko, Bryan A. PlummerPersonality Modeling in Language Models Procedural Content Generation
Boston University

Aug 19, 2026J/K move · Enter open · S save

Are LLMs becoming similarly creative? Evidence from three years of models

Nirav Patel, Josiah Crossman, Eva Aggarwal, Emily Wenger

Department of Computer Science, Duke University · Department of Electrical & Computer Engineering, Duke University

Homogenization in Language Models Computational Creativity Language Model Generation Evaluation Creativity Assessment

Open paper AskPDF arXiv

Many benchmarks track Large Language Model (LLM) performance on tasks with verifiable answers, but less is known about how LLM performance is evolving on open-ended tasks, where creativity, originality and diversity may matter as much as quality. As LLMs increasingly support human ideation and creative work, understanding trends in LLM performance on open-ended tasks is critical. This paper presents a preliminary analysis of LLM creative outputs spanning three years of model releases, examining model responses to Infinity-Chat100, a real-world collection of open-ended user queries, and the Alternate Uses Task, an established psychometric creativity assessment. Using sentence-embedding similarity, we examine trends in LLM responses to these prompts. Our findings show a statistically significant decrease in model output diversity over time, suggesting that LLM outputs may be converging in creative substance across models. If this trend persists, LLM-driven homogenization may progressively diminish human agency in human-AI co-creative work, demanding careful consideration of LLMs' role in the human creative process.

Abstract

Output Diversity in Language Models

Explore similar work

Are LLMs becoming similarly creative? Evidence from three years of models

Examining and Addressing Barriers to Diversity in LLM-Generated Ideas

PERSONAWEAVER: Controllable Diversity Beyond Conventional Archetypes in Procedural Character Generation