{"id":"5838368c-4669-4d9d-ab56-5ebe874265f8","arxiv_id":"2608.08046","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Rewriting report-derived prompts to remove non-imageable clinical language and optimizing that rewriting with GRPO reduces RadDINO-FID by about half for a frozen chest X-ray generator while keeping aggregate report alignment.","lead":"This paper shows that rewriting chest X-ray report text as compact visual prompts, without changing the image generator, dramatically improves the realism of generated X-rays on a standard benchmark. The authors then apply reinforcement learning to the text rewriting model to recover alignment with the original reports.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The FID gain may come from dropped findings: only 60% of positive labels survive rewriting, and per-condition FID worsens in 9/14 categories; the alignment recovery is partly reward-driven.","rationale":"I read the paper in good faith. The matched-seed relative reward is a sound control for stochastic rendering, the instruction-only baseline cleanly isolates rewriting from policy learning, and releasing code is a real asset. The central claim, however, hinges on 'faithful depiction.' The paper's own Table C.1 shows that positive labels are preserved only 60% and CheXbert exact match 29.9%, which the text defends as surface-form compression. That defense is weakened because CheXbert is semantic, not lexical. Table 3's 9/14 per-condition FID regressions make the concern concrete: the global FID gain is not mirrored in most pathologies, suggesting the metric is rewarding overall distributional similarity, not per-finding fidelity. The BioViL-T recovery is also partially circular because BioViL-T is directly optimized with weight 0.45 in Eq. 2. These points do not refute the existence of an optimization axis; they mean the evidence does not yet establish that the axis preserves clinical content. A conditional verdict is therefore appropriate, conditioned on a label-preservation check. I agree with the reader's weakest assumption; this stress test adds the per-condition FID pattern and the reward-weight circularity.","tokens_in":21072,"tokens_out":7395,"duration_ms":77231,"concrete_test":"Take the 5,034 CheXGenBench evaluation rows and split them by whether the rewritten prompt preserves all positive CheXpert labels from the source (using CheXbert or the paper's CheXbert audit). Recompute RadDINO-FID and per-condition FID (Table 3) on the subset with full label preservation, comparing direct Sana vs JustLLMGRPO. If the 50.6% global FID improvement shrinks substantially or the per-condition regressions remain, the headline gain is confounded by dropped findings rather than re-expression.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim requires that rewriting removes only non-renderable report discourse and preserves clinically renderable findings. Appendix C.1 (Table C.1) does not support this: positive CheXpert label preservation is 60.0% and exact CheXbert label match between source and candidate prompts is 29.9%. CheXbert is a semantic label classifier, so the 29.9% figure indicates label-level divergence, not just synonym substitution. If 40% of positive findings are absent from the conditioning text, the improved global RadDINO-FID (54.225 to 26.780) could reflect a simpler, more central image distribution rather than more faithful per-finding control. Table 3 is consistent with that reading: conditional RadDINO-FID worsens in 9 of 14 CheXpert categories relative to direct Sana even though aggregate FID improves. The alignment recovery after GRPO (0.609 to 0.696) is not independent evidence of faithfulness because BioViL-T is a 0.45-weight term in the training reward (Eq. 2); the policy is optimized against it. The paper's own caveat that 'competitive rankings do not imply uniform improvement' concedes that the global metric can mask per-condition regressions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":21354,"tokens_out":6272,"duration_ms":61104,"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":[{"comment":"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.","section":"§4 (Further Analysis) and Appendix C.1, Table C.1"},{"comment":"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.","section":"§4 (Condition-level fidelity), Table 3"},{"comment":"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.","section":"Equation (2), Table 4"},{"comment":"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.","section":"§4 (Experimental Setup), Tables 1, 3, 4"}],"minor_comments":[{"comment":"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.","section":"Abstract and Section 1"},{"comment":"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'.","section":"§4 (Prompt-Language Analysis), Figure 3c"},{"comment":"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...'.","section":"§4 (Further Analysis)"},{"comment":"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.","section":"§5 (Conclusion)"}],"recommendation":"major_revision","confidential_remarks":"The paper is well structured and unusually transparent about its own limitations. The main scientific contribution—demonstrating a large aggregate fidelity gain from prompt rewriting for a frozen generator—is likely real, and the matched-seed protocol is a methodological strength. However, the reviewer concerns are load-bearing: the preservation audit undermines the faithful-depiction narrative, the alignment recovery after GRPO is confounded by using the same scorer in the training reward, and the state-of-the-art comparison rests on single-run point estimates and possibly mismatched splits. These issues are fixable through additional experiments (e.g., generated-image-level clinical preservation, uncertainty estimates, rerunning baselines on the same split) and a more careful framing. I do not see grounds for rejection, but the revision requirements go beyond local presentation fixes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nThe 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.\n\nIf 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.","headline":"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.","tokens_in":21864,"tokens_out":2496,"would_cite":true,"duration_ms":24929,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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…","keywords":["chest X-ray generation","prompt optimization","group relative policy optimization","text-to-image diffusion","radiology report reformulation","CheXGenBench","Sana","reinforcement learning"],"falsifier":"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.","tokens_in":20914,"feed_emoji":"🩻","tokens_out":10357,"duration_ms":98083,"temperature":0.7,"pith_summary":"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.","feed_headline":"Rewriting prompts halves X-ray FID with the generator frozen","feed_subtitle":"Prompt wording is a real lever: rewriting cuts X-ray FID from 54.2 to 26.8 with frozen generator.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies CheXGenBench, the benchmark protocol, baseline values, and MIMIC-CXR split that every reported score is measured against.","marker":"Dutt et al. 2026"},{"why":"Introduces Sana, the CXR-adapted generator that remains frozen while only prompts are optimized.","marker":"Xie et al. 2024"},{"why":"Introduces Qwen3-4B, the language model used both as the unmodified rewriter and as the trained prompt policy.","marker":"Yang et al. 2025"},{"why":"Introduces GRPO, the group-relative policy-optimization algorithm applied to the prompt policy.","marker":"Shao et al. 2024"},{"why":"Demonstrates large-language-model reinforcement learning with GRPO-style updates, motivating the choice of policy optimization.","marker":"Guo et al. 2025"},{"why":"Provides BioViL-T, the vision-language model used for source-prompt alignment in both the reward and evaluation.","marker":"Bannur et al. 2023"},{"why":"Provides CheXpert labels, which ground the label-consistency component of the reward and the per-category fidelity analysis.","marker":"Irvin et al. 2019"},{"why":"Provides LLaVA-Rad source prompts and the clinical alignment evaluation used to judge generated images.","marker":"Zambrano Chaves et al. 2025"},{"why":"Applies GRPO to prompt rewriting for a fixed general-purpose image generator, the closest prior setting this work differentiates against.","marker":"Wu et al. 2026a"}],"fun_headline_variants":["Prompt rewriting halves chest X-ray FID with frozen generator","LLM-optimized prompts cut X-ray FID by 50% without touching generator","Better prompts, not better generators: X-ray FID halved","GRPO-tuned prompts slash X-ray FID while preserving alignment","Frozen generator, rewritten prompts: X-ray FID cut in half"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Prompt rewriting halves chest X-ray FID with frozen generator","LLM-optimized prompts cut X-ray FID by 50% without touching generator","Better prompts, not better generators: X-ray FID halved","GRPO-tuned prompts slash X-ray FID while preserving alignment","Frozen generator, rewritten prompts: X-ray FID cut in half"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000529,"raw_usage":{"total_tokens":2583,"prompt_tokens":1013,"completion_tokens":1570,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":1475}},"tokens_in":629,"tokens_out":1570,"duration_ms":14587,"temperature":1.0,"reasoning_tokens":1475,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:30:05.107193+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"A.; and Hospedales, T","cited_arxiv_id":null,"evidence_quote":"Supplies CheXGenBench, the benchmark protocol, baseline values, and MIMIC-CXR split that every reported score is measured against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates large-language-model reinforcement learning with GRPO-style updates, motivating the choice of policy optimization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides BioViL-T, the vision-language model used for source-prompt alignment in both the reward and evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides CheXpert labels, which ground the label-consistency component of the reward and the per-category fidelity analysis."},{"cited_title":"M.; Huang, S.-C.; Xu, Y.; et al","cited_arxiv_id":null,"evidence_quote":"Provides LLaVA-Rad source prompts and the clinical alignment evaluation used to judge generated images."}],"review_version":1}