Modern generative models produce high-quality individual outputs but struggle to sample diversely, even when explicitly instructed to do so. We introduce MINE-SAMPLE-RENDER (MSR), a test-time approach that externalizes the diversity into an explicit, corpus-derived sampling procedure. Our pipeline (i) mines an interpretable categorical schema from a corpus of real artifacts and supplies statistics for sampling; (ii) samples attribute configurations from the corpus-derived statistics; (iii) uses a generative model purely as a conditional renderer of a sampled configuration. Across weather report generation, user simulation, webpage design, text-to-image generation, and agent-generated 3D assets, we evaluate the diversity and output quality of MSR-guided generation and show that MSR improves alignment with corpus distributions and diversity of generated outputs.
Figures & tables
Figure 1: Different generations, familiar choices. Five baseline outputs per task . Top: GPT-5.5 webpages generated from the same simple prompt and a given website content. Middle: GPT Image 2 outputs for “A cat is curled up next to a light teddy bear.” Bottom: Independent GPT-6 Astra agent runs asked to build a functional 3D folding chair. See Figures 4 , 5 , and 6 for comparison with MSR, with protocols and results in Sections 3.3 , 3.4 , and 3.5 .
Figure 2: Overview of Mine–Sample–Render. (Top) We use LLMs to mine attributes to create a schema A(D) by comparing small cohorts of artifacts. We then label our source data with the schema to create an empirical distribution of attributes. (Bottom) At inference time, we sample attribute configurations and a model uses them, together with the prompt, to generate diverse outputs.
Figure 3: (Left) LLMs generate weather reports that collapse to a specific mode and do not produce samples that follow conditions that are common but do not belong to the dominant mode , even when given the source dataset in the context (Right) . Generated-text distributional gap against the held-out real distribution of weather conditions.
User simulator
KLfwd↓
KLbwd↓
JS ↓
Coverage ↑
GPT-5-mini (prompt only)
0.5668
0.5764
0.5866
0.4646
+ PersonaHub (coding-filtered)
0.5971
0.6033
0.6157
0.4013
+ MSR (compact)
0.5236
0.5556
0.5319
0.5649
+ MSR (fine-grained)
0.5184
0.5339
0.5373
0.5106
GPT-5.5 (prompt only)
0.5258
0.5272
0.5167
0.6345
+ PersonaHub (coding-filtered)
0.5395
0.5505
0.5371
0.5850
Table 1: Distributional gap between real and simulated user behaviors in coding, following the evaluation protocol of Mehri et al. (2026) : an annotator describes each real and simulated conversation along behavioral facets. These descriptions are embedded and clustered, and the resulting real and simulated cluster distributions are compared using Forward and Backward KL and Jensen–Shannon divergence. Coverage is calculated as the percentage of real clusters with at least one simulated user.
Figure 4: Webpages generated by GPT-5.5 with a fixed prompt ( top ) and additionally MSR ( bottom ).
Webpage generator
Vendi Score
Odd-one-out
text ↑
image ↑
% removed ↑
mean removal rank ↓
GPT-5.5 (prompt only)
2.158
2.446
-
-
+ MSR (WebSight)
3.956
4.116
95.075%
6.20 vs 19.30
+ MSR (Design2Code)
3.273
3.457
97.650%
5.77 vs 19.73
Kimi K2.6 (prompt only)
2.426
3.150
-
-
+ MSR (WebSight)
3.842
4.077
75.575%
9.16 vs 16.33
Table 2: Diversity evaluation for generated webpages.
Figure 5: Images generated by GPT Image 2 using the prompt “A cat is curled up next to a light teddy bear.” (Row 1) Original prompt only. (Row 2) Independent rewrite with GPT-5.6 Sol. (Row 3) MSR –Resolved, using GPT-5.6 Sol for semantic resolution.
Generator / conditioning
Vendi Score
Odd-one-out
CLIP ↑
DINOv2 ↑
% removed ↑
mean removal rank ↓
GPT Image 2 (prompt only)
2.37
2.16
13.80%
37.97
+ independent rewrite
3.02
3.19
88.80%
25.26
+ coordinated rewrite
3.86
4.32
97.40%
16.45
+ MSR –Direct
5.14
4.97
-
-
+ MSR –Resolved
6.68
6.73
100.00%
6.82
Table 3: Image diversity across three generators. Vendi scores are averaged over 50 ten-image sets per arm. Odd-one-out compares the three baselines with MSR –Resolved, removing 30 of 40 images per request; survivors receive rank 40. Dashes indicate that Direct was not included in this evaluation.
Figure 6: 3D folding chairs generated with GPT-6 Astra.
Method
Mean odd-one-out removal rank ↓
Quality ↑
GPT-6 Astra independent generations
48.800
82.35
GPT-6 Astra coordinated generations
25.125
70.80
GPT-6 Astra + MSR
17.100
83.55
Table 4: Evaluation for 3D folding chair generation.
Appendix figures & tables24 assets
Supplementary material from the paper’s appendix.
Appendix
Axis
Data-generating attribute
Recovered attribute
Source support
Temperature
temp_cold
temperature_cold
64
temp_mild
temperature_mild
97
temp_warm
temperature_warm
99
temp_hot
temperature_hot
40
Sky
sky_clear
sky_clear
74
sky_partly
sky_partly_cloudy
84
Appendix
Table 5: Data-generating and recovered weather schemas. Names are the exact retained identifiers; the correspondence is a semantic alignment, not an exact string match. Source support counts the intended configurations of the 300 rendered source reports, not labels extracted from their text. A dash means no distinct corresponding attribute was recovered. The ten city indicators are supplied metadata and are not included in these 16 weather rows.
Condition
Generated
Complete labels
Excluded
Prompt only
3,000
2,838
162
Append reports
3,000
3,000
0
Mine–Sample–Render recovered schema
3,000
3,000
0
Known-schema control
3,000
3,000
0
Appendix
Table 6: Weather reports generated and retained for distributional evaluation. All conditions generate 300 reports per city. Prompt-only exclusions are incomplete four-axis annotations after retries, not ungenerated reports; its retained counts range from 244 to 300 per city.
Schema
Attribute
Count
Frequency
Compact
tone_no_greeting_direct
539
89.83%
Compact
expectation_of_code_full_snippet
415
69.17%
Compact
request_imperative_command
366
61.00%
Compact
no_or_small_code_sample
283
47.17%
Compact
context_provides_short_example_or_none
274
45.67%
Fine-grained
request_imperative_command
409
68.17%
Appendix
Table 7: The five most frequent non-absent attributes per user schema after back-applying the frozen vocabulary to all 600 training conversations. Identifiers are retained verbatim.
Schema
Attribute A
Attribute B
Joint count
Lift
Compact
expectation_of_code_full_snippet
tone_no_greeting_direct
376 (62.67%)
1.009
Compact
request_imperative_command
tone_no_greeting_direct
343 (57.17%)
1.043
Compact
expectation_of_code_full_snippet
request_imperative_command
297 (49.50%)
1.173
Fine-grained
request_imperative_command
wants_full_runnable_code_or_complete_artifact
258 (43.00%)
1.198
Fine-grained
evidence_pastes_existing_code_snippet
provides_code_or_artifact_example
182 (30.33%)
1.841
Fine-grained
provides_code_or_artifact_example
request_imperative_command
182 (30.33%)
0.927
Appendix
Table 8: The three most frequent unordered cross-axis attribute pairs per user schema. Joint frequency is the fraction of the 600 source conversations containing both attributes. Lift divides joint frequency by the product of the two marginal frequencies.
User simulator
Inference control
Effective N
GPT-5-mini
Low reasoning
4,960
GPT-5.5
Low reasoning
4,903
GPT-4o
No explicit reasoning control
4,946
Kimi K2.6
No explicit reasoning control
4,935
Qwen3.8-27B
Thinking enabled; private state removed
4,956
Appendix
Table 9: User simulators and final matched evaluation sizes.
Schema
Attribute
Count
Frequency
WebSight
type_grotesque_sans
8,134
81.34%
WebSight
color_light_neutral_background
6,653
66.53%
WebSight
header_present
6,192
61.92%
WebSight
content_left_aligned
5,811
58.11%
WebSight
top_h1_heading_present
4,314
43.14%
Design2Code
site_header_present
418
87.08%
Appendix
Table 10: The five most frequent attributes per schema in the empirical source-page matrices. Identifiers are retained verbatim. Counts use 10,000 WebSight pages and 480 Design2Code pages.
Schema
Attribute A
Attribute B
Joint count
Joint frequency
Lift
WebSight
color_light_neutral_background
type_grotesque_sans
5,462
54.62%
1.009
WebSight
header_present
type_grotesque_sans
5,148
51.48%
1.022
WebSight
content_left_aligned
type_grotesque_sans
4,821
48.21%
1.020
Design2Code
site_footer_present
site_header_present
372
77.50%
1.065
Design2Code
primary_navigation_present
site_header_present
337
70.21%
1.118
Design2Code
body_typography_sans_serif
site_header_present
325
67.71%
0.998
Appendix
Table 11: The three most frequent unordered co-occurring attribute pairs per schema. Joint frequency is the fraction of source pages containing both attributes. Lift divides joint frequency by the product of the two marginal frequencies; values near one indicate little departure from marginal independence.
Table 12: Highest-lift attribute pairs with joint support of at least 1%. The three rows per schema are ranked by lift, not frequency. Several Design2Code pairs tie at lift 96; alphabetical tie-breaking determines the displayed subset.
#
Category
Prompt ID
Exact prompt
1
baseball bat
baseball_bat_403675
A guy holding a baseball bat is ready to play ball
2
cow
cow_537695
A cow has a ring in its nose near two men.
3
wine glass
wine_glass_285773
A table with a wine glass and wine bottles on it
4
dining table
dining_table_106338
a big table that has some plates of food on it
5
spoon
spoon_306638
A bowl of food set on a table with a spoon.
6
scissors
scissors_223362
a close up of a person holding a pair of scissors
Appendix
Table 19
Valid images (%)
Attributes realized (%)
Generator
Direct
Resolved
Direct
Resolved
GPT Image 2
99
97
61.6
80.2
MAI-Image-2.5
96
97
65.3
78.5
FLUX.2-flex
95
90
71.1
77.9
Appendix
Table 14: Model-based audits of 100 images per MSR variant and generator. Each attribute-realization rate uses 2,000 originally sampled attribute occurrences.
λ
CLIP Vendi
DINOv2 Vendi
0.0
4.23
4.40
0.5
5.14
4.97
1.0
5.34
5.21
Appendix
Table 21
Representation
λ=.5 Resolved
λ=1 Resolved
CLIP
6.68
6.62
DINOv2
6.73
6.42
Appendix
Table 22
Source
Retained records
Amazon Berkeley Objects
65
Wikimedia Commons
83
Art Institute of Chicago
1
Cleveland Museum of Art
1
Metropolitan Museum of Art
10
Appendix
Table 23
Property A
Property B
P(B ∣ A)
P(B ∣ not A)
Phi
Feet: metal
Frame: tubing
95.3%
6.3%
0.890
Back: sling
Back: visible textile weave
95.8%
8.5%
0.876
Back: light
Seat: light
91.3%
3.8%
0.879
Back: edge binding
Seat: edge binding
95.7%
11.8%
0.841
Back: geometric
Seat: geometric
71.4%
2.2%
0.750
Appendix
Table 24
Post-hoc topic description
Two enriched properties: topic probability / corpus prevalence
Figure 7: (Top) Example of a report generated by GPT-6 Astra for a generated folding chair. (Bottom) Illustration of the folding mechanism, originally provided as a video.
Figure 8: From sampled attributes to an articulated chair: open, intermediate, and folded states of the same generated Blender model. The programmed sequence includes a lift and tilt during folding; the folded pose is held rather than claimed to stand unsupported. These renders demonstrate modeled folding motion, not physically verified performance.
Quality criterion
Weight
MSR
Coordinated
Coherent deployed chair and support
25%
4.40
3.70
Resolved, meaningful folding transformation
30%
4.40
3.85
Human use, handling, and folded-state usefulness
15%
3.90
3.35
Integrated form and detailed resolution
20%
4.20
3.35
Compelling and purposeful design idea
10%
3.90
3.75
Weighted total / 100
100%
84.70
72.55
Appendix
Table 15: Evaluation of chair quality. Each criterion is scored out of 5 and assigned a weight by an investigator agent, different from the final judge.
Figure 9: Webpages generated by GPT-5.5 using the prompt “Create a training hub page using the grounding file for the manufacturing onboarding program. Use orange and yellow as the main colors – I want it to feel energetic and welcoming. Include all content and images from the grounding doc.” (Left) Original prompt only. (Middle) Mine–Sample–Render using the Websight schema (Right) Mine–Sample–Render using the Design2Code schema.
Figure 10: Webpages generated by Kimi K2.6 using the prompt “Create a recap page for the Axium annual hackathon using the grounding doc, green color scheme.” (Left) Original prompt only. (Middle) Mine–Sample–Render using the Websight schema (Right) Mine–Sample–Render using the Design2Code schema.
Figure 11: Images generated by GPT Image 2 using the prompt “A closeup on some chocolate cake has a fork by it.” (Row 1) Original prompt only. (Row 2) Independent rewrite with GPT-5.6 Sol. (Row 3) Coordinated rewrite with GPT-5.6 Sol. (Row 4) MSR –Resolved.
Figure 12: Images generated by GPT Image 2 using the prompt “A table with a wine glass and wine bottles on it.” (Row 1) Original prompt only. (Row 2) Independent rewrite with GPT-5.6 Sol. (Row 3) Coordinated rewrite with GPT-5.6 Sol. (Row 4) MSR –Resolved.
Figure 13: Images generated by GPT Image 2 using the prompt “Apples and oranges sitting next to each other on a shelf.” (Row 1) Original prompt only. (Row 2) Coordinated rewrite with GPT-5.6 Sol. (Row 3) Independent rewrite with GPT-5.6 Sol. (Row 4) MSR –Resolved.
Figure 14: Images generated by GPT Image 2 using the prompt “A dragon swinging a bat with a dog watching.” (Row 1) Original prompt. (Row 2) Mine–Sample–Render .
Open-ended generation tasks often require a set of meaningfully different outputs, yet large language models often produce similar generations. Existing test-time diversity methods operate at different stages of generation with varying effectiveness, but it remains unclear what design choices lead to meaningful diversity in the output. We introduce a framework that characterizes test-time diverse generation methods by the diversity source introduced during generation and provide a transmission score for measuring how effectively variation in the source reaches the final output. Guided by this framework, we propose fully automated specification-level generation methods that first generate diverse intermediate specifications and then condition on them to produce final responses. Across five open-ended tasks and four backbone models, specification-level injection improves output diversity over test-time baselines while maintaining comparable quality. Our analysis shows that successful diversity injection depends on both the diversity of the sources and their transmission to the output, highlighting source design and source-to-output realization as two key levers for building more diverse generation systems.
Deep generative models have achieved great success in producing high-quality samples, making them a central tool across machine learning applications. Beyond sample quality, an important yet less systematically studied question is whether trained generative models faithfully capture the diversity of the underlying data distribution. In this work, we address this question by directly comparing the diversity of samples generated by state-of-the-art models with that of test samples drawn from the target data distribution, using recently proposed reference-free entropy-based diversity scores, Vendi and RKE. Across multiple benchmark datasets, we find that test data consistently attains substantially higher Vendi and RKE diversity scores than the generated samples, suggesting a systematic downward diversity bias in modern generative models. To understand the origin of this bias, we analyze the finite-sample behavior of entropy-based diversity scores and show that their expected values increase with sample size, implying that diversity estimated from finite training sets could inherently underestimate the diversity of the true distribution. As a result, optimizing the generators to minimize divergence to empirical data distributions would induce a loss of diversity. Finally, we discuss potential diversity-aware regularization and guidance strategies based on Vendi and RKE as principled directions for mitigating this bias, and provide empirical evidence suggesting their potential to improve the results.
Farzan Farnia, Mohammad Jalali, Azim Ospanov
Department of Computer Science and Engineering, The Chinese University of Hong Kong.
Modern generative models still lack human-level creativity, particularly in multi-branch diversity. Prior approaches to address this problem often incur heavy computation or strong dependency on model architecture. Therefore, we introduce UAG(Universal Avoidance Generation), a model-agnostic and computationally efficient generation strategy that penalizes similarity among previously generated outputs. Thus, UAG can enhance multi-branch diversity across both diffusion and transformer models, with minimal additional computation. In experiments, our method achieves up to 1.9 times higher diversity, runs 4.4 times faster, and requires only 1/64 of the FLOPs compared to state-of-the-art methods. The full code is https://anonymous.4open.science/r/2026_ACL_Universal/.