Mine, Sample, Render: Test-Time Diversity Alignment Of Generative Models
Kostas Triaridis, Morris E. Sharp, Nebojsa Jojic
cs.AI, cs.LG
2026-10-15
MSR mines attributes from a real corpus, samples configurations from corpus statistics, and renders them with any generator, lifting DINOv2 Vendi from 4.3 to 6.7 without training.
Generators today deliver excellent single outputs and narrow batches. Ask GPT-5.5 for ten webpages from one prompt and the layouts converge; let a GPT-6 Astra agent independently build a folding chair twenty times and it keeps landing on the same design decisions. Raising the temperature or writing "be diverse" into the prompt says nothing about what should vary, how often, or which combinations are plausible. Reward-maximizing post-training makes it worse by concentrating samples around high-reward modes. A single coherent output can still look obviously machine-made when placed in a crowd.
This matters wherever you need a population, not a specimen: synthetic datasets, simulated users for agent evaluation, and text-to-image test-time scaling, where diversity caps the payoff from extra samples. The variation almost certainly appeared during pretraining; the deployed model just does not reliably recover it at sampling time. The paper splits "what can be generated" from "how often it should be generated" and hands the second question to classical statistics.
MSR is a training-free, three-stage test-time pipeline that treats the generator as a pure conditional renderer.
The design point is the division of labor: the corpus acts as external memory, the foundation model does what it is reliable at (naming differences, realizing an explicit attribute), and plain statistics decides how often each choice appears. In the controlled weather experiment, mining almost exactly recovered the axes used to build the dataset, decent evidence that the split works.
Weather reports, scored against held-out 2017 observations:
| Method | Forward KL ↓ | JS ↓ | Coverage ↑ |
| Prompt only | 0.570 | 0.484 | 21% |
| Corpus in context | 0.394 | 0.342 | 24% |
| MSR | 0.181 | 0.203 | 68% |
| True-distribution reference | 0.150 | 0.166 | 83% |
Stuffing the whole corpus into context helps far less than sampling from its estimated statistics. MSR approaches the true-schema reference, with coverage still 15 points short.
User simulation on 5,000 real coding conversations: MSR shrinks the gap to real users on all five simulators. GPT-4o drops from 0.605 to 0.353 Forward KL with coverage rising from 29.9% to 73.0%; GPT-5.5 goes from 0.526 to 0.310. PersonaHub personas come out worse than prompt-only on four of five models.
| Task | Metric | Baseline | MSR |
| Webpages (GPT-5.5) | Vendi, structure | 2.158 | 3.956 |
| Text-to-image (GPT Image 2) | DINOv2 Vendi | 4.32 (coordinated rewrite) | 6.73 |
| 3D folding chair (GPT-6 Astra) | Quality score | 82.35 (independent) | 83.55 |
| 3D folding chair | Odd-one-out removal rank ↓ | 48.8 (independent) / 25.1 (asked to differ) | 17.1 |
Text-to-image replicates across GPT Image 2, MAI-Image-2.5, and FLUX.2-flex with the same sampled configurations reused; in blinded odd-one-out pools, 99.8-100% of MSR-Resolved images get removed early. The 3D study carries the sharpest comparison: telling the agent to make each chair different from the last does raise variety, but quality falls from 82.35 to 70.80, while MSR posts the best diversity at 83.55, slightly above independent generation.
The pipeline is training-free and generator-agnostic; schemas and sampled configurations port across models, as the three-generator image experiment shows. Concrete fits: synthetic dataset construction, user simulators for agent evaluation, and batch design work anywhere the set has to resemble a real population rather than one polished template. Attributes are readable natural language, explicit constraints can be pinned into a configuration, and failures trace back to a specific attribute. The honest framing: attribute-guided generation, discrete density estimation, and hierarchical generation all have precedents, with fLSA as the closest relative. The contribution is estimating a corpus-level joint distribution over compositional attributes and sampling it as the diversity source, validated across five domains.
The authors flag that mining quality is bound by the miner: the weather schema lost "windy" despite 24 supporting examples, and the controlled setting cannot separate mining failure from rendering failure. Gains on Kimi K2.6 are markedly smaller than on GPT-5.5, attributed to instruction-following gaps.
Unreported costs stand out: mining and labeling 10,000 webpages or a 28,875-attribute image schema is a real preprocessing bill the paper never totals. Odd-one-out judging relies on GPT-5.6 Sol and GPT-6 Astra, the same family as several generators, with no external calibration of judge taste. User-simulation coverage tops out at 81%, so the absolute gap to real users remains large. The weather corpus is itself GPT-5-mini-generated text, so the cleanest experiment has a synthetic flavor. Vendi and odd-one-out measure spread, not whether the spread is useful; per-output quality checks exist only for some tasks.