JustLLMGRPO: Radiographic Control for Chest X-Ray Generation

Anglin Liu; Jintai Chen; Pengxiang Cai; Qingyuan Zeng; Xiaohan Li; Zexun Li

REVIEW 4 major objections 4 minor 52 references

JustLLMGRPO: Radiographic Control for Chest X-Ray Generation

T0 review · 4 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read The paper claims that for a frozen, CXR-adapted Sana generator, the prompt itself is a major optimization axis: one-pass LLM rewriting lowers RadDINO-FID from 54.225 to 27.572, and adding GRPO to the prompt policy reaches 26.780 while…

desk verdict Prompt reformulation alone roughly halves RadDINO-FID on a frozen CXR generator; the claim is real but the 'preserves clinical findings' part needs better evidence than the paper currently gives. read the letter →

arxiv 2608.08046 v1 pith:I26F3ZWJ submitted 2026-08-08 cs.AI

Pengxiang Cai , Xiaohan Li , Anglin Liu , Qingyuan Zeng , Zexun Li , Jintai Chen This is my paper · ORCID

classification cs.AI

keywords chestX-raygenerationpromptoptimizationgrouprelativepolicytext-to-imagediffusionradiologyreportreformulationCheXGenBenchSanareinforcementlearning

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 establishes that, for a frozen chest-X-ray-adapted Sana generator, the wording of the prompt is a large optimization axis. Rewriting report-derived prompts once with an unmodified LLM lowers RadDINO-FID from 54.225 to 27.572, nearly halving the metric, by removing temporal comparisons, uncertainty, and report framing while emphasizing visible findings. Because unconstrained rewriting drops source-prompt alignment from 0.695 to 0.609, the paper adds standard group relative policy optimization (GRPO) to the LLM prompt policy with group-relative radiology-aware image rewards, reaching RadDINO-FID 26.780 and restoring alignment to 0.696. On CheXGenBench this yields the best point estimates for fidelity, coverage, and downstream classification among the compared generators, although per-category fidelity is not uniform: relative to direct Sana prompting, five of fourteen CheXpert categories improve and nine worsen.

What carries the argument

The load-bearing machinery is a prompt-policy loop: a Qwen3-4B prompt policy samples candidate prompts; frozen Sana renders each candidate and a matched-seed source-prompt control sharing the same initial noise; a radiology-aware reward $Q(I,c,y)=0.45\,b(I,c)+0.10\,\ell(I,y)+0.45\,d(I)$ measures each candidate's improvement over the control; and standard group relative policy optimization (GRPO) — a critic-free reinforcement-learning update that derives advantages from a group of sampled responses — updates only the language model with a clipped surrogate and KL regularization. The matched-seed relative reward is what isolates the causal effect of wording from stochastic rendering variation.

What would settle it

A decisive check would be to start from a set of source prompts that already contain only renderable visual language; if optimizing those prompts still lowers RadDINO-FID by the same margin, the reported gain is not caused by removing report discourse. Independently, have radiologists label generated images for the source CheXpert findings; if agreement with source labels does not improve relative to direct prompting, the maintained source-alignment claim would not mean faithful depiction.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that an already CXR-adapted text-to-image generator leaves a large quality margin in how the prompt is phrased. With Sana frozen, an unmodified LLM that rewrites a report-derived prompt into a concise visual description removes temporal and uncertain language while keeping explicit findings, and RadDINO-FID falls from 54.225 to 27.572. Because unconstrained rewriting lowers BioViL-T alignment with the source prompt from 0.695 to 0.609, the paper adds GRPO to the LLM prompt policy, using a matched-seed relative reward that compares each candidate rendering against a source-prompt control. The trained prompt policy reaches RadDINO-FID 26.780 and alignment 0.696, and the paper reads this as proof that prompt expression is a distinct and substantial axis for text-conditioned chest X-ray generation, not a substitute for generator-side adaptation.

Load-bearing premise

The method assumes that stripping temporal, uncertain, and report-framing language from a radiology prompt removes only content that cannot be drawn, and keeps the important findings; the paper's own audit finds only 60.0% of positive CheXpert labels preserved and 29.9% exact CheXbert matches, so the fidelity gain could partly come from simpler images rather than truer ones.

Editorial extensions

If this is right

Prompt-policy optimization and generator-side adaptation are complementary levers; gains from one do not exhaust the other.
Report-derived prompts for frozen medical generators should be compressed into concise visual descriptions before rendering, because temporal, uncertain, and framing language dilutes the conditioning signal.
Group-relative image feedback with matched seeds can recover source-prompt alignment lost by aggressive rewriting, without retraining the generator.
Better aggregate fidelity and downstream classification can coexist with worse per-category fidelity: compared with direct Sana, JustLLMGRPO improves five of fourteen CheXpert categories and worsens nine.
Optimizing the prompt policy can outperform directly fine-tuning the generator on the same reward: the Sana-GRPO control raises alignment to 0.816 but FID worsens to 59.145, whereas prompt-policy optimization keeps alignment at 0.696 and FID at 26.780.

Reading between the lines

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

A decisive extension is to test the method on source prompts that already contain only renderable visual language; if the FID gain persists, the mechanism is not discourse removal but a deeper rephrasing effect, and if it vanishes, the gain is mostly compression.
Because the paper's own text audit preserves only 60.0% of positive CheXpert labels and 29.9% exact CheXbert matches, an independent clinical reader study is needed to confirm that the improved images actually depict the source findings rather than a statistically plausible but simpler chest X-ray.
The same matched-seed, group-relative reward design is a transferable recipe for any text-conditioned image generator faced with verbose, narrative prompts, in or beyond radiology.
The composite reward's RadDINO component could be gamed by producing images that match the reference feature distribution while dropping rare findings; adding an independent pathology classifier or human preference signal to the reward would test robustness.

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

Quick take: this paper demonstrates something real and worth building on. With the Sana generator frozen, instruction-only LLM rewriting of report-derived prompts cuts RadDINO-FID from ~54 to ~27.6, and the GRPO step adds a small further gain while recovering source alignment. The strongest part is that the largest gain comes from a plain rewrite with no reward, which rules out the worry that the headline is just reward hacking. The matched-seed relative reward is a clean design choice, and the prompt-language analysis is useful. Code is released.

The soft spots are where you'd expect. All headline numbers are single-run point estimates with no error bars, and the 'state-of-the-art' claim rests on a development evaluation while baselines come from CheXGenBench. More importantly, the preservation audit in Appendix C.1 undercuts the 'maintained alignment means faithful depiction' reading: positive CheXpert label preservation is 60%, exact CheXbert label match is 29.9%, and conditional RadDINO-FID worsens in 9 of 14 categories relative to direct Sana. The paper is honest about these numbers and says surface-form changes are not clinical correctness, but the abstract and intro still lean on the preservation story. It is entirely possible that part of the FID gain comes from simpler, more central images rather than more faithful per-finding control. The fact that GRPO restores BioViL-T alignment from 0.609 to 0.696 is not independent evidence, since BioViL-T is a 45% weight in the training reward.

If I were refereeing, I would ask for: (a) multiple seeds or bootstrapped CIs on the headline metrics, (b) a human or at least a radiology-aware image classifier evaluation of whether dropped positives are actually absent from the generated images, and (c) an analysis of the 9/14 per-condition regressions. But I would not desk-reject this. The core phenomenon—that expression matters as much as generator choice—is credible, and the ablation isolating instruction-only rewriting is exactly the right experiment. It deserves a serious referee and a revise-and-resubmit, not a rejection.

Referee Report

4 major / 4 minor

Summary. The paper studies whether the performance of a frozen, CXR-adapted text-to-image generator can be improved by optimizing only the input prompt policy. With Sana frozen, the authors show that a single reformulation of LLaVA-Rad source prompts by an unmodified Qwen3-4B LLM reduces RadDINO-FID from 54.225 to 27.572. They then apply Group Relative Policy Optimization (GRPO) to the LLM prompt policy, using a matched-seed relative reward that combines BioViL-T alignment, CheXpert label consistency, and RadDINO fidelity. The final JustLLMGRPO system reports RadDINO-FID 26.780, BioViL-T alignment 0.696 versus 0.695 for direct prompting, state-of-the-art PRDC coverage (0.588), and downstream macro AUC 0.746. The central claim is that prompt expression is a distinct and substantial optimization axis for text-conditioned chest X-ray generation, even after generator-side domain adaptation.

Significance. If the results hold, the paper identifies a practically useful and underexplored axis: for a fixed, already-adapted generator, the linguistic form of the conditioning prompt can yield large aggregate fidelity gains. The experimental design is strong in several respects: the generator and reward models are frozen; candidates and source controls share deterministic seeds; the evaluation is performed on held-out source prompts; and the paper is unusually candid about the limits of point estimates and the non-uniformity of per-category results. The public code release is a further strength. However, the preservation audit in Appendix C.1 substantially weakens the 'faithful depiction' interpretation. The headline alignment maintenance after GRPO is also confounded by the fact that BioViL-T is a 0.45-weight term in the training reward. These issues are load-bearing for the paper's main claim, though they are addressable with additional analyses or tempered conclusions.

major comments (4)

[§4 (Further Analysis) and Appendix C.1, Table C.1] The preservation audit directly contradicts the central assumption that rewriting removes only non-renderable report discourse while preserving clinically renderable findings. Table C.1 reports only 60.0% positive CheXpert label preservation and 29.9% exact CheXbert label match between source and candidate prompts. CheXbert is a semantic label classifier, so the 29.9% figure indicates actual label-level divergence, not merely synonym substitution. If 40% of positive findings are absent or altered in the conditioning text, the improved global RadDINO-FID (54.225 to 26.780) may partly reflect simpler or more central image distributions rather than more faithful per-finding control. The paper's caveat that the audit 'does not measure clinical correctness' does not resolve this, because the CheXbert comparison is a semantic diagnostic. To support the faithful-depiction claim, the authors should either (a) measure preservation on the generated images (e.g., CheXpert-classifier agreement between images generated from source versus rewritten prompts), or (b) explicitly restrict their claims to aggregate distributional fidelity and coverage, and substantially temper statements such as 'faithfully depict specified findings' in the abstract and Section 1.
[§4 (Condition-level fidelity), Table 3] The per-condition results show that the global FID improvement is not uniform: compared with direct Sana, JustLLMGRPO improves five categories and worsens nine, and it falls outside the top two in Fracture, Lung Lesion, Pleural Other, and Pneumothorax. The paper's own acknowledgment that 'competitive rankings do not imply uniform improvement' is buried after the headline aggregate result. This is load-bearing because the abstract and conclusion claim state-of-the-art fidelity and faithful depiction, which are not supported for these categories. The authors should present the per-condition table in the main text (or at least prominently summarize the nine regressions), discuss plausible mechanisms for the regressions, and adjust the fidelity claims to reflect the distribution of per-category behavior rather than only the aggregate point estimate.
[Equation (2), Table 4] The GRPO training reward Q in Eq. (2) is a weighted sum of BioViL-T (0.45), CheXpert label consistency (0.10), and RadDINO (0.45). The evaluation metrics in Table 4 are BioViL-T alignment and RadDINO-FID, i.e., the same or closely related functions as the training reward components. Consequently, the recovery of BioViL-T alignment from 0.609 (instruction-only rewriting) to 0.696 (after GRPO) is expected because the policy is directly optimized against that scorer. While the evaluation set contains held-out source prompts, the score functions are identical to the training objectives, so the alignment maintenance is not independent evidence of preserved source-prompt semantics. The authors should either evaluate alignment with a reward model or metric not used in training, or explicitly frame the recovery as reward optimization and provide a separate semantic-preservation analysis (e.g., with a held-out CheXpert classifier on generated images) to support the claim that GRPO restores alignment without sacrificing findings.
[§4 (Experimental Setup), Tables 1, 3, 4] All headline results are single-run point estimates without error bars, confidence intervals, or significance tests. The paper itself labels the evaluation a 'development evaluation' and repeatedly writes 'not statistical significance'. This is particularly problematic for the GRPO-specific contributions: the FID improvement from 27.572 to 26.780 is only 2.9%, and the alignment difference (0.696 versus 0.695) is within plausible noise. Additionally, the comparison to CheXGenBench baselines appears to compare the authors' development-set evaluation against published baseline values that may have been computed on a different split; the paper does not state that the baselines were rerun on the same development set. This undermines the state-of-the-art claim in Table 1. Please provide uncertainty estimates (multiple seeds or bootstrapping) and confirm that all compared numbers are obtained on the same evaluation split, or rerun the baselines under the authors' protocol.

minor comments (4)

[Abstract and Section 1] The phrase 'anunderexplored' in the abstract paragraph of Section 1 is missing a space ('an underexplored'). More substantively, the abstract and Section 1 claim that prompts 'faithfully depict specified findings'; this should be qualified given the preservation audit in Table C.1 and the per-condition regressions in Table 3.
[§4 (Prompt-Language Analysis), Figure 3c] The caption for Figure 3c says 'preserving explicit findings', but the underlying statistic is the prevalence of at least one visual-finding term, not per-finding preservation. The text later clarifies this, but the figure caption is misleading and should be reworded to say 'prevalence of visual-finding terms'.
[§4 (Further Analysis)] The sentence 'These results indicate that JustLLMGRPO preserves generation-critical constraints while focusing the frozen generator on renderable findings' overstates the evidence, since the cited audit preserves laterality and negative polarity at high rates but positive findings at only 60.0%. Recommend softening to 'preserves laterality and negative-polarity constraints while...'.
[§5 (Conclusion)] The conclusion states that JustLLMGRPO 'achieves state-of-the-art distribution coverage and downstream classification utility'. Given the single-run development-evaluation status and the split-matching question raised in the major comments, this should be phrased as 'achieves competitive point estimates on the development evaluation' unless uncertainty quantification and matching baselines are added.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the headline FID gain comes from reward-free LLM rewriting, and the GRPO metric overlap is transparent rather than a reduction.

full rationale

The paper's central empirical claim is that instruction-only rewriting by an unmodified Qwen3-4B lowers RadDINO-FID from 54.225 to 27.572 with Sana frozen (Table 4). This step involves no reward fitting, no evaluation-set optimization, and no parameter update, so it does not reduce to the paper's own inputs. The subsequent GRPO stage does optimize a reward, Eq. 2, that is a weighted sum of BioViL-T (0.45), CheXpert label consistency (0.10), and a RadDINO-derived score (0.45), and the benchmark then reports BioViL-T alignment and RadDINO-FID; hence the reported alignment recovery from 0.609 to 0.696 is partly an expected consequence of optimizing the same scoring function. However, the paper explicitly states the reward definition, evaluates on a held-out set, and does not present that alignment recovery as an independent prediction; it is a training outcome. The large fidelity improvement is not attributable to this reward because the one-pass rewriting baseline already achieves it. Appendix C.1 is candid that the audit 'is sensitive to surface-form changes and does not measure clinical correctness', and the low positive-label preservation rate (60.0%) is a real clinical-fidelity limitation but not a circularity. The paper also explicitly concedes that 'competitive rankings do not imply uniform improvement' (conditional FID worsens in 9/14 categories). There is no load-bearing self-citation chain, no imported uniqueness theorem, and no ansatz smuggled in via citation; the benchmark, Sana checkpoint, LLaVA-Rad prompts, and reward models are external resources. Therefore the claimed derivation chain is not circular, and the appropriate score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The listed assumptions are all domain-level or protocol-level; none are derived in the paper. The matching-seed and reward-validity assumptions are the ones whose failure would most directly undermine the central claim. No new physical or architectural entities are introduced; the matched-seed relative reward is a protocol, not an entity.

free parameters (5)

Reward component weights (b, l, d) = 0.45, 0.10, 0.45
Eq. 2 combines BioViL-T, CheXpert label consistency, and RadDINO into the per-image score; weights are chosen by hand, not learned or justified by a selection procedure.
RadDINO internal weights = 0.60, 0.30, 0.10
Appendix B.2: global feature-density, local top-k similarity, and memorization penalty weights inside the RadDINO component; hand-set.
Reward transformation thresholds = 0.1, -0.03 (Eq. 3)
Piecewise mapping from relative improvement to reward uses manually chosen thresholds; central to GRPO credit assignment.
Candidate group size K = 5
Group size for GRPO advantage normalization in Eq. 6; chosen by the authors; affects variance of advantage estimates.
GRPO regularization constants = clip ratio 0.2, KL beta 1e-3, learning rate 1e-6
Standard GRPO hyperparameters, set by hand rather than derived; they influence stability but are not fitted to the target performance.

assumptions (4)

domain assumption Reward models BioViL-T, CheXpert classifier, and RadDINO are valid proxies for image fidelity, label consistency, and source alignment.
Eq. 2 and the evaluation use these models; if they are biased or gameable, both training and reported gains are suspect.
domain assumption Removing temporal, uncertainty, and report-framing language preserves the renderable findings in the source prompt.
Core of the compression claim; Appendix C.1 shows only 60.0% positive-label preservation, so this assumption is partially contradicted.
domain assumption A matched seed neutralizes stochastic renderer variation so relative rewards isolate the prompt effect.
Eq. 1 and Appendix B.1; plausible but assumes the seed controls all rendering confounding.
domain assumption CheXGenBench development evaluation is comparable to the published baseline values used for comparison.
Tables 1-3 compare against benchmark-provided baselines while noting the evaluation is on the development set; the assumption is not fully verified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of JustLLMGRPO: Radiographic Control for Chest X-Ray Generation." pith.science (2026). https://pith.science/paper/I26F3ZWJ

@misc{pith2026260808046,
  author       = {Pith},
  title        = {Pith review of: JustLLMGRPO: Radiographic Control for Chest X-Ray Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I26F3ZWJ}},
  note         = {Machine review of arXiv:2608.08046}
}

read the original abstract

Text-conditioned chest X-ray generation aims to synthesize realistic radiographs that faithfully depict specified findings. Existing work has primarily improved quality by updating image generators, implicitly treating prompts as fixed after CXR-domain adaptation. We show that this generator-centric view leaves a substantial optimization dimension underexplored. With a CXR-adapted Sana generator frozen, one-pass reformulation by an unmodified LLM reduces RadDINO-FID from 54.225 to 27.572. Prompt analysis shows that the LLM suppresses temporal comparisons, uncertainty, and other non-renderable report content while emphasizing visible radiographic findings. However, unconstrained reformulation reduces BioViL-T alignment with source prompts from 0.695 to 0.609. We therefore introduce JustLLMGRPO, which applies standard Group Relative Policy Optimization (GRPO) only to the LLM prompt policy while keeping Sana frozen. Group-relative radiology-aware image feedback retains visual focus while preserving source-prompt alignment. On CheXGenBench, JustLLMGRPO reduces RadDINO-FID to 26.780, a 50.6% improvement over direct prompting, while maintaining alignment (0.696 versus 0.695). It also achieves state-of-the-art distribution coverage and downstream classification utility. These results show that substantial performance can remain latent in how radiographic information is expressed to an adapted generator. Code is publicly available at https://github.com/pxcai/JustLLMGRPO.

Figures

Figures reproduced from arXiv: 2608.08046 by the authors.

**Figure 1.** Prompt optimization refocuses a frozen CXR generator. With Sana frozen, the source prompt produces an incorrect CXR with right-dominant effusion, whereas the JustLLMGRPO prompt restores the requested laterality. architecture, parameters, or learned conditioning representations. This practice implicitly assumes that the remaining bottleneck lies within the generator. Yet the conditioning text determines which findin… view at source ↗

**Figure 2.** Lexical shift induced by LLM rewriting. Word clouds show the 18 words with the largest decreases and increases in the proportion of prompts containing them after rewriting. Word size reflects the magnitude of change. ator that remains frozen throughout. Direct source prompting yields a RadDINO-FID of 54.225. Reformulating each prompt once with an unmodified Qwen3-4B reduces it by 49.2% to 27.572 without updating ei… view at source ↗

**Figure 4.** JustLLMGRPO overview. The prompt policy samples candidate prompts, and frozen Sana produces matched-seed candidate and source renderings. A radiology-aware relative reward scores each candidate, and GRPO updates only the LLM prompt planner. The optimized prompt conditions Sana at inference. GRPO for prompt policies. GRPO estimates advantages within sampled groups without a separate critic (Shao et al. 2024; Guo et a… view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Representative qualitative comparisons under matched source prompts. Each row starts from the same LLaVA-Rad source prompt; cyan italics mark the requested findings. JustLLMGRPO more clearly depicts bilateral perihilar opacities, a right pleural effusion, and a left r…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 46 canonical work pages

[1]

Bannur, S.; Hyland, S.; Liu, Q.; et al. 2023. Learning To Exploit Temporal Structure for Biomedical Vision-Language Processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 15016--15027

work page 2023
[2]

Black, K.; Janner, M.; Du, Y.; Kostrikov, I.; and Levine, S. 2024. Training Diffusion Models with Reinforcement Learning. International Conference on Learning Representations, 2024: 4965--4987

work page 2024
[3]

M.; Abraham, T

Bluethgen, C.; Chambon, P.; Delbrouck, J.-B.; van der Sluijs , R.; Po acin, M.; Zambrano Chaves, J. M.; Abraham, T. M.; Purohit, S.; Langlotz, C. P.; and Chaudhari, A. S. 2025. A Vision--Language Foundation Model for the Generation of Realistic Chest X-Ray Images. Nature Biomedical Engineering, 9(4): 494--506

work page 2025
[4]

Cao, T.; Wang, C.; Liu, B.; Wu, Z.; Zhu, J.; and Huang, J. 2023. BeautifulPrompt: Towards Automatic Prompt Engineering for Text-to-Image Synthesis. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track, 1--11

work page 2023
[5]

Deng, M.; Wang, J.; Hsieh, C.-P.; et al. 2022. RLPrompt: Optimizing Discrete Text Prompts with Reinforcement Learning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 3369--3391

work page 2022
[6]

A.; and Hospedales, T

Dutt, R.; Sanchez, P.; Yao, Y.; McDonagh, S.; Tsaftaris, S. A.; and Hospedales, T. 2026. CheXGenBench: A Unified Benchmark For Fidelity, Privacy and Utility of Synthetic Chest Radiographs. Transactions on Machine Learning Research

work page 2026
[7]

Fan, Y.; Watkins, O.; Du, Y.; Liu, H.; Ryu, M.; Boutilier, C.; Abbeel, P.; Ghavamzadeh, M.; Lee, K.; and Lee, K. 2023. DPOK: Reinforcement Learning for Fine-Tuning Text-to-Image Diffusion Models. In Advances in Neural Information Processing Systems, volume 36, 79858--79885

work page 2023
[8]

Guo, D.; Yang, D.; Zhang, H.; et al. 2025. DeepSeek-R1 Incentivizes Reasoning in LLMs through Reinforcement Learning. Nature, 645(8081): 633--638

work page 2025

Show all 52 references

[9]

Han, W.; Kim, C.; Ju, D.; Shim, Y.; and Hwang, S. J. 2024. Advancing Text-Driven Chest X-Ray Generation with Policy-Based Reinforcement Learning. In Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024: 27th International Conference, Marrakesh, Morocco, Oc...

2024
[10]

Hao, Y.; Chi, Z.; Dong, L.; and Wei, F. 2023. Optimizing Prompts for Text-to-Image Generation. In Advances in Neural Information Processing Systems, volume 36, 66923--66939

2023
[11]

Irvin, J.; Rajpurkar, P.; Ko, M.; et al. 2019. CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01): 590--597

2019
[12]

Johnson, A. E. W.; Pollard, T. J.; Berkowitz, S. J.; Greenbaum, N. R.; Lungren, M. P.; Deng, C.-y.; Mark, R. G.; and Horng, S. 2019. MIMIC-CXR, a de-Identified Publicly Available Database of Chest Radiographs with Free-Text Reports. Scientific Data, 6(1): 317

2019
[14]

J.; Chang, J.; and Ye, J

Lee, S.; Kim, W. J.; Chang, J.; and Ye, J. C. 2024 b . LLM-CXR: Instruction-Finetuned LLM for CXR Image Understanding and Generation. International Conference on Learning Representations, 2024: 29745--29765

2024
[16]

I.; de Moura, J.; Novo, J.; and Ortega, M

Mor \'i s, D. I.; de Moura, J.; Novo, J.; and Ortega, M. 2024. Adapted Generative Latent Diffusion Models for Accurate Pathological Analysis in Chest X-Ray Images. Medical & Biological Engineering & Computing, 62(7): 2189--2212

2024
[17]

P.; et al

P\'erez-Garc\'ia, F.; Bond-Taylor, S.; Sanchez, P. P.; et al. 2025 a . RadEdit: Stress-Testing Biomedical Vision Models via Diffusion Image Editing. In Computer Vision -- ECCV 2024, 358--376

2025
[18]

P\'erez-Garc\'ia, F.; Sharma, H.; Bond-Taylor, S.; et al. 2025 b . Exploring Scalable Medical Image Encoders beyond Text Supervision. Nature Machine Intelligence, 7(1): 119--130

2025
[21]

Shentu, J.; and Al Moubayed, N. 2024. CXR-IRGen: An Integrated Vision and Language Model for the Generation of Clinically Accurate Chest X-Ray Image-Report Pairs. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 5212--5221

2024
[22]

Weber, T.; Ingrisch, M.; Bischl, B.; and R \"u gamer, D. 2023. Cascaded Latent Diffusion Models for High-Resolution Chest X-Ray Synthesis. In Advances in Knowledge Discovery and Data Mining, 180--191

2023
[23]

Wu, M.; Wang, L.; Zhao, P.; et al. 2026 a . RePrompt: Reasoning-Augmented Reprompting for Text-to-Image Generation via Reinforcement Learning. International Conference on Learning Representations, 2026: 14030--14057

2026
[24]

Wu, S.; Sun, M.; Wang, W.; Wang, Y.; and Liu, J. 2026 b . VisualPrompter: Semantic-Aware Prompt Optimization with Visual Feedback for Text-to-Image Synthesis. International Conference on Learning Representations, 2026: 30578--30600

2026
[26]

Xu, Y.; Sun, L.; Peng, W.; Jia, S.; Morrison, K.; Perer, A.; Zandifar, A.; Visweswaran, S.; Eslami, M.; and Batmanghelich, K. 2024. MedSyn: Text-Guided Anatomy-Aware Synthesis of High-Fidelity 3-D CT Images. IEEE Transactions on Medical Imaging, 43(10): 3648--3660

2024
[28]

M.; Huang, S.-C.; Xu, Y.; et al

Zambrano Chaves, J. M.; Huang, S.-C.; Xu, Y.; et al. 2025. A Clinically Accessible Small Multimodal Radiology Model and Evaluation Metric for Chest X-Ray Findings. Nature Communications, 16(1): 3108

2025
[29]

I.; Han, Z.; Paster, K.; Pitis, S.; Chan, H.; and Ba, J

Zhou, Y.; Muresanu, A. I.; Han, Z.; Paster, K.; Pitis, S.; Chan, H.; and Ba, J. 2022. Large Language Models Are Human-Level Prompt Engineers. In The Eleventh International Conference on Learning Representations

2022
[30]

Learning To Exploit Temporal Structure for Biomedical Vision-Language Processing , booktitle =
[31]

International Conference on Learning Representations , volume =

Training Diffusion Models with Reinforcement Learning , author =. International Conference on Learning Representations , volume =
[32]

Nature Biomedical Engineering , volume =

A Vision--Language Foundation Model for the Generation of Realistic Chest X-Ray Images , author =. Nature Biomedical Engineering , volume =
[33]

BeautifulPrompt: Towards Automatic Prompt Engineering for Text-to-Image Synthesis , booktitle =

Cao, Tingfeng and Wang, Chengyu and Liu, Bingyan and Wu, Ziheng and Zhu, Jinhui and Huang, Jun , year = 2023, pages =. BeautifulPrompt: Towards Automatic Prompt Engineering for Text-to-Image Synthesis , booktitle =

2023
[34]

RLPrompt: Optimizing Discrete Text Prompts with Reinforcement Learning , booktitle =
[35]

Transactions on Machine Learning Research , langid =

CheXGenBench: A Unified Benchmark For Fidelity, Privacy and Utility of Synthetic Chest Radiographs , author =. Transactions on Machine Learning Research , langid =
[36]

DPOK: Reinforcement Learning for Fine-Tuning Text-to-Image Diffusion Models , booktitle =

Fan, Ying and Watkins, Olivia and Du, Yuqing and Liu, Hao and Ryu, Moonkyung and Boutilier, Craig and Abbeel, Pieter and Ghavamzadeh, Mohammad and Lee, Kangwook and Lee, Kimin , year = 2023, volume =. DPOK: Reinforcement Learning for Fine-Tuning Text-to-Image Diffusion Models ...

2023
[37]

DeepSeek-R1 Incentivizes Reasoning in LLMs through Reinforcement Learning , year = 2025, journal =

2025
[38]

Advancing Text-Driven Chest X-Ray Generation with Policy-Based Reinforcement Learning , booktitle =

Han, Woojung and Kim, Chanyoung and Ju, Dayun and Shim, Yumin and Hwang, Seong Jae , year = 2024, pages =. Advancing Text-Driven Chest X-Ray Generation with Policy-Based Reinforcement Learning , booktitle =

2024
[39]

Optimizing Prompts for Text-to-Image Generation , booktitle =

Hao, Yaru and Chi, Zewen and Dong, Li and Wei, Furu , year = 2023, volume =. Optimizing Prompts for Text-to-Image Generation , booktitle =

2023
[40]

CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison , year = 2019, journal =

2019
[41]

Scientific Data , volume =

MIMIC-CXR, a de-Identified Publicly Available Database of Chest Radiographs with Free-Text Reports , author =. Scientific Data , volume =
[42]

International Conference on Learning Representations , volume =

LLM-CXR: Instruction-Finetuned LLM for CXR Image Understanding and Generation , author =. International Conference on Learning Representations , volume =
[43]

2302.12172 , keywords =

Vision-Language Generative Model for View-Specific Chest X-Ray Generation , author =. 2302.12172 , keywords =

arXiv
[44]

2403.17804 , keywords =

Improving Text-to-Image Consistency via Automatic Prompt Optimization , author =. 2403.17804 , keywords =

arXiv
[45]

Medical & Biological Engineering & Computing , volume =

Adapted Generative Latent Diffusion Models for Accurate Pathological Analysis in Chest X-Ray Images , author =. Medical & Biological Engineering & Computing , volume =
[46]

Exploring Scalable Medical Image Encoders beyond Text Supervision , year = 2025, journal =

2025
[47]

RadEdit: Stress-Testing Biomedical Vision Models via Diffusion Image Editing , booktitle =
[48]

2310.03739 , keywords =

Aligning Text-to-Image Diffusion Models with Reward Backpropagation , author =. 2310.03739 , keywords =

arXiv
[49]

arXiv preprint arXiv:2402.03300 , author =

DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , year = 2024, eprint =. arXiv preprint arXiv:2402.03300 , author =

2024 arXiv
[50]

CXR-IRGen: An Integrated Vision and Language Model for the Generation of Clinically Accurate Chest X-Ray Image-Report Pairs , booktitle =

Shentu, Junjie and Al Moubayed, Noura , year = 2024, pages =. CXR-IRGen: An Integrated Vision and Language Model for the Generation of Clinically Accurate Chest X-Ray Image-Report Pairs , booktitle =

2024
[51]

Cascaded Latent Diffusion Models for High-Resolution Chest X-Ray Synthesis , booktitle =

Weber, Tobias and Ingrisch, Michael and Bischl, Bernd and R. Cascaded Latent Diffusion Models for High-Resolution Chest X-Ray Synthesis , booktitle =
[52]

RePrompt: Reasoning-Augmented Reprompting for Text-to-Image Generation via Reinforcement Learning , year = 2026, journal =

2026
[53]

International Conference on Learning Representations , volume =

VisualPrompter: Semantic-Aware Prompt Optimization with Visual Feedback for Text-to-Image Synthesis , author =. International Conference on Learning Representations , volume =
[54]

arXiv preprint arXiv:2410.10629 , author =

SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers , year = 2024, eprint =. arXiv preprint arXiv:2410.10629 , author =

2024 arXiv
[55]

IEEE Transactions on Medical Imaging , volume =

MedSyn: Text-Guided Anatomy-Aware Synthesis of High-Fidelity 3-D CT Images , author =. IEEE Transactions on Medical Imaging , volume =
[56]

arXiv preprint arXiv:2505.09388 , author =

Qwen3 Technical Report , year = 2025, eprint =. arXiv preprint arXiv:2505.09388 , author =

2025 arXiv
[57]

A Clinically Accessible Small Multimodal Radiology Model and Evaluation Metric for Chest X-Ray Findings , year = 2025, journal =

2025
[58]

Large Language Models Are Human-Level Prompt Engineers , booktitle =

Zhou, Yongchao and Muresanu, Andrei Ioan and Han, Ziwen and Paster, Keiran and Pitis, Silviu and Chan, Harris and Ba, Jimmy , year = 2022, langid =. Large Language Models Are Human-Level Prompt Engineers , booktitle =

2022

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

Bannur, S.; Hyland, S.; Liu, Q.; et al. 2023. Learning To Exploit Temporal Structure for Biomedical Vision-Language Processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 15016--15027

work page 2023

[2] [2]

Black, K.; Janner, M.; Du, Y.; Kostrikov, I.; and Levine, S. 2024. Training Diffusion Models with Reinforcement Learning. International Conference on Learning Representations, 2024: 4965--4987

work page 2024

[3] [3]

M.; Abraham, T

Bluethgen, C.; Chambon, P.; Delbrouck, J.-B.; van der Sluijs , R.; Po acin, M.; Zambrano Chaves, J. M.; Abraham, T. M.; Purohit, S.; Langlotz, C. P.; and Chaudhari, A. S. 2025. A Vision--Language Foundation Model for the Generation of Realistic Chest X-Ray Images. Nature Biomedical Engineering, 9(4): 494--506

work page 2025

[4] [4]

Cao, T.; Wang, C.; Liu, B.; Wu, Z.; Zhu, J.; and Huang, J. 2023. BeautifulPrompt: Towards Automatic Prompt Engineering for Text-to-Image Synthesis. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track, 1--11

work page 2023

[5] [5]

Deng, M.; Wang, J.; Hsieh, C.-P.; et al. 2022. RLPrompt: Optimizing Discrete Text Prompts with Reinforcement Learning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 3369--3391

work page 2022

[6] [6]

A.; and Hospedales, T

Dutt, R.; Sanchez, P.; Yao, Y.; McDonagh, S.; Tsaftaris, S. A.; and Hospedales, T. 2026. CheXGenBench: A Unified Benchmark For Fidelity, Privacy and Utility of Synthetic Chest Radiographs. Transactions on Machine Learning Research

work page 2026

[7] [7]

Fan, Y.; Watkins, O.; Du, Y.; Liu, H.; Ryu, M.; Boutilier, C.; Abbeel, P.; Ghavamzadeh, M.; Lee, K.; and Lee, K. 2023. DPOK: Reinforcement Learning for Fine-Tuning Text-to-Image Diffusion Models. In Advances in Neural Information Processing Systems, volume 36, 79858--79885

work page 2023

[8] [8]

Guo, D.; Yang, D.; Zhang, H.; et al. 2025. DeepSeek-R1 Incentivizes Reasoning in LLMs through Reinforcement Learning. Nature, 645(8081): 633--638

work page 2025