acceptodds
Under review as a conference paper at ICLR 2027

HIdden in the DEcodes: Recovering Clinical Findings without Retraining Medical MLLMs

13 views

Abstract

Automatic radiology report generation commonly commits to a single decoded report from a medical multimodal large language model (MLLM). However, clinically correct findings omitted from this report may be hidden in other stochastic decodes produced by the same frozen model, while unsupported findings may also recur across decodes. Existing approaches address these errors by modifying training objectives, model architectures, supervision, or refinement modules, but typically require model updates or additional training and do not ensure that a single decode captures all supported findings. We introduce HIDE, a model-agnostic, training-free framework for recovering clinical findings without retraining medical MLLMs. HIDE treats the frozen generator as a proposal distribution for verification-guided inference: stochastic decoding broadens candidate coverage, cross-sample consistency provides model-side proposal support, localized visual support from a prototype bank verifies image compatibility, and source-constrained minimal edits compose only jointly supported findings. Given an initial draft and additional stochastic samples, HIDE extracts finding-level candidates and selects among three conservative actions: adding a supported finding missing from the draft, removing an unsupported draft claim, or abstaining when the evidence is inconclusive. Across six frozen MLLM report generators, HIDE achieves mean relative gains of 27.3%, 54.1%, and 16.0% on CheXbert-F1-14, CheXbert-F1-5, and RadGraph-F1 on MIMIC-CXR, respectively. On CheXpert Plus, the corresponding gains reach 41.2%, 239.7%, and 19.4%. The correction signal also transfers to IU-Xray when fitted only on MIMIC-CXR or CheXpert Plus. HIDE further achieves state-of-the-art performance among training-free correction methods.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Radiology Report Generation, Multimodal Large Language Model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_gAgZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes HIDE, which corrects chest X-ray reports without updating the report generator. It extracts findings from a greedy draft and several stochastic samples, combining cross-sample support with local prototype-based visual evidence. Branch-specific predictors then authorize additions, deletions, or no change. The authors report improvements across six generators, two in-domain datasets, and transfer to IU-Xray.

Critical strengths
  1. The problem is important and well motivated. A finding omitted by one decode may still appear in other samples. The method also recognizes that repeated claims are not necessarily correct.
  2. Finding-level edits can combine useful content from different reports. Positive, normal, and confounder prototypes provide complementary evidence.
Critical weaknesses

W1. Key implementation details are missing

Predictor fitting. Section 3.4 and Appendices D/F specify the features and thresholds, but not the model used for TbT_b. The target labels, fitting objective, fitting procedure, and exact definition of kbk_b also need clarification. Appendix F confirms that the predictor has fitted parameters. The thresholds alone are therefore insufficient to reproduce it. Please report the actual size of the 0.5% subset and explain how fitting and threshold selection are organized.

Visual verification and missing prototype roles. Section 3.3 does not specify JJ, ββ, prototype counts, or how small and empty prototype groups are handled. Table B.2 states that no sufficient negative rule is defined for pleural other, so how are SnormS^{norm} and SconfS^{conf} in Eq. (4) computed for this category? Listing pleural effusion or atelectasis as alternatives does not establish that pleural other is absent. Please provide any additional label mapping and the fallback rule for groups with fewer than JJ samples.

Text editing. Section 3.5 describes copying source sentences and deleting matched spans, but does not specify the editor. Does it use string operations, sentence-rewriting rules, or a language model? If a model is used, please provide its identity, prompt, and constraint checks. A source sentence may contain several findings, while only some are authorized. How does the editor avoid adding the others? Pseudocode and an edit trace would make the procedure much easier to assess.

W2. Baseline implementations and resource use need clarification

The baselines have different designs. Please document their full implementations and parameter settings. In particular, report CCD’s expert model and control parameters. For CCS, specify the utility, embedding checkpoint, candidate count, and sampling settings. Please also explain how DeCo, ICD, and VCD are adapted to the tested generators.

In addition, a matched-size or shared candidate pool would be a useful additional comparison for CCS. Reporting auxiliary data, pretrained components, and end-to-end costs would clarify the performance–resource trade-off. It is recommended to provide the above points as well.

Questions for authors
  1. See weaknesses.
  2. Suggestions: 1)The gains across models and datasets support the method's overall effectiveness. An edit-level analysis would help explain these gains. Useful statistics including action frequencies, correct and incorrect edits, and the rate of unchanged reports will help in understanding how overall benefits are generated and which cases remain at risk. 2)Please check Figure 4. There is a strange subfigure in it.
3 comments

We sincerely thank the reviewer for recognizing that our paper is well motivated, the importance of the research problem, the feasibility of recovering omitted findings from other samples, and the value of finding level edits and complementary prototype evidence.We have uploaded the revised paper.


The reviewer’s concerns focus on several implementation details, which we clarify and supplement below.

Weaknesses

[W1.1] Predictor Details.

Response:We have added details to Appendix D (Lines 1024-1040). We clarify this below.

  • Model. For each generator and branch, we fit a CART classifier TbT_b on [onehot(fg),kb(g),m(g)][\mathrm{onehot}(f_g), k_b(g), m(g)]. Each split maximizes Gini impurity reduction, with a maximum tree depth of 7 and a minimum of 10 samples per leaf.
  • Cross-sample support. kb(g)k_b(g) counts additional sampled reports matching the candidate’s finding and laterality. ADD+ and ADD− require matching polarity. For DEL, a report counts if it describes the target as present or uncertain.
  • Targets. Let yg=1y_g=1 if the CheXbert-14 reference label is positive and 00 otherwise. The target is ygy_g for ADD+ and 1−yg1-y_g for ADD− and DEL.
  • Fitting. Five-fold grouped cross-fitting on each dataset’s 0.5% training fitting subset yields out-of-fold probabilities. The subsets contain 784 image–report pairs for MIMIC-CXR and 245 for CheXpert Plus.
  • Thresholds. We use these probabilities to select the threshold triplet that maximizes overall F1 on the edited reports, then freeze the predictor and thresholds for testing.

[W1.2] Prototype Details.

Response: We have added details to Appendix B (Lines 901-917). We clarify this below.

  • Prototype counts. The bank contains 12,267 prototypes across 129 shards per role, with a maximum of C=32C=32 prototypes per shard: 4,128 positive, 4,108 normal, and 4,031 confounder. For pleural other, the bank contains 160 normal and 128 confounder prototypes.
  • Retrieval and scoring. We use J=5J=5 and β=50\beta=50 in Eq. (4). For pleural other, Snorm(g)S_{\mathrm{norm}}(g) and Sconf(g)S_{\mathrm{conf}}(g) are the mean cosine similarities to the five nearest prototypes in the matching normal and confounder shards, respectively.
  • Negative-label mapping for pleural other. Both roles require an explicit, conflict-free annotation that pleural or parenchymal scarring is absent in the ipsilateral whole-lung region. Among regions meeting this condition, we assign the confounder role if pleural effusion, costophrenic angle blunting, atelectasis (including linear or patchy atelectasis), or lobar or segmental collapse is explicitly present. Otherwise, we assign the normal role.
  • Small or empty groups. A role group must represent at least 20 distinct patients to be eligible. If a required role is unavailable or has fewer than 5 prototypes at retrieval, HIDE applies no edit and preserves the original report. No group was empty, and no case triggered this fallback in our experiments.

[W1.3] Editing Module Details.

Response: We have added details to Appendix H (Lines 1242-1315). After locating the target clause and sentence (c,s)(c,s), HIDE applies one of two rules:

  • If the sentence contains only the authorized finding, HIDE directly adds or deletes that sentence.
  • If the sentence also contains other findings, HIDE uses the same frozen model that generated the report (temperature 0, top-p=1p=1, maximum 256 tokens) to make the local text edits needed to add or remove only the authorized finding.

Prompt

You are a local radiology sentence editor, not a diagnostician. You receive one sentence from an existing radiology report and an immutable list of exact target clauses. Perform only the requested local operation. Preserve every unrelated finding, device, comparison, negation, uncertainty, laterality, and factual statement. Do not introduce a new finding, diagnosis, cause, recommendation, or medical fact. Do not rewrite the whole report.

For DELETE_REPAIR, remove the listed target clauses and their target specific modifiers, then repair only the grammar of the remaining sentence. For ADD_SUBTRACTION, retain the listed clauses and remove the listed unwanted finding clauses. Preserve the retained finding’s original medical wording, polarity, and laterality. Repair only the local grammar.

Return one JSON object only with exactly one key named revised_sentence. Do not return Markdown or explanations.

OPERATION
{DELETE_REPAIR or ADD_SUBTRACTION}

ORIGINAL SENTENCE
{original sentence}

CLAUSES TO RETAIN
· {clause to retain} or none

CLAUSES TO REMOVE
· {clause to remove} or none

Apply exactly this operation and return JSON only.

Due to space limitations, we will continue our response in the next window.

Below is our continued response to [W1.3].

Pseudocode

For each authorized ADD or DEL:
    Locate target clause c and sentence s.
    If s contains only the target finding:
        Copy s into the report for ADD, or remove s for DEL.
    Otherwise:
        ADD: retain authorized clauses and mark all other finding clauses for removal.
        DEL: mark the target clause for removal and retain unrelated clauses.
        Pass s and both clause lists to the frozen local editor.
        Validate the returned sentence before applying the edit.

Edit trace example

StepExample
Sampled sentence“Bibasilar atelectasis and a small left pleural effusion are present.”
AuthorizationADD bibasilar atelectasis only.
Editor inputADD_SUBTRACTION · Retain “Bibasilar atelectasis” · Remove “a small left pleural effusion”
Sentence added“Bibasilar atelectasis is present.”

[W2.1] Baseline implementation details

Response: We have added the implementation to Appendix I (Lines 1316-1381). We followed the official implementations. The first table lists each method’s settings, and the second gives the candidate counts and generator-specific adaptations.

BaselineImplementationParameter settings
CCDUses the TorchXRayVision DenseNet expert. Symptom-grounded contrastive decoding (SCD) uses the expert’s symptom labels, and expert-informed contrastive decoding (ECD) adjusts token probabilities.Expert threshold 0.5, SCD weight α=0.5\alpha=0.5, ECD weight β=0.5\beta=0.5, diagnostic plausibility constraint γ=10\gamma=10. Fixed across generators.
CCSUses Image-Grounded Utility with the CCS authors’ CXR-adapted Qwen3-VL-Embed-2B checkpoint. CCS scores all candidate pairs and selects the report with the highest mean utility against the other N−1N-1 reports.Sampling temperature τ=0.5\tau=0.5. The candidate count NN matches HIDE’s sampling budget for each generator, as listed below.
DeCoAt each decoding step, the generator’s native adapter selects an anchor layer from the specified range and uses its logits to calibrate the final-layer logits.Top-p=0.9p=0.9, correction strength α=0.6\alpha=0.6. Anchor-layer indices are listed below.
ICDEach generator receives the same image with the original and negative prompts. ICD subtracts the weighted negative-prompt distribution, applies truncation, and samples the next token.Contrast weight λ=1.0\lambda=1.0, truncation coefficient α=0.1\alpha=0.1, top-p=1.0p=1.0, repetition penalty 1.01.0, beam size 11.
VCDContrasts the original image with a perturbed view generated by the official add_diffusion_noise function. Each generator’s visual interface is listed below.Contrast strength α=0.1\alpha=0.1, truncation threshold β=0.1\beta=0.1, noise step 500500.

CCS and HIDE use the same number of candidate reports for each generator. Both methods use the same fixed pool of reports for each study.

GeneratorDeCo anchor layersVCD adaptation
Lingshu-7B20–28Native Qwen2.5-VL processor
MedGemma-4B-IT20–28Gemma3Processor, with noise applied directly to pixel_values
MedVLM-R1-2B20–28Native Qwen2-VL processor, preserving image_grid_thw alignment and RoPE state
Med-R1-2B20–28Native Qwen2-VL processor, preserving image_grid_thw alignment and RoPE state
LLaVA-RAD-7B20–28Native LLaVA wrapper with raw image tensors
CheXAgent-2-3B23–31CheXAgent visual mode with the Phi forward pass

For CheXAgent, the DeCo layer range follows its hidden-state indexing, where hidden_states[0] contains the input embeddings. The ICD negative prompt used across generators was: “You are an unreliable medical image observer, only give a vague impression of this image and do not rely strictly on the visual content.”

[W2.2] CCS needs matched-pool comparison and resource documentation.

Response: We evaluated HIDE and CCS on MIMIC-CXR in the same controlled environment using Med-R1-2B, a shared fixed candidate pool, and the same sampling budget. HIDE uses MAIRA-2 for grounding, RAD-DINO for visual encoding, a prototype bank, and a fitted action predictor. CCS uses a pretrained RRG-Qwen-VL-Embedding selector adapted on MIMIC-CXR image–report pairs.

Appendix I reports runtime and FLOPs per study with all stages run sequentially. HIDE took 23.80 s for candidate generation, 6.32 s for parsing, 0.36 s for grounding, 0.09 s for prototype evidence, and less than 0.001 s for editing. CCS took 23.37 s for candidate generation and 1.55 s for embedding and selection. End-to-end costs were 30.57 s and 4.74×10134.74 \times 10^{13} FLOPs per study for HIDE, compared with 24.92 s and 2.79×10132.79 \times 10^{13} FLOPs for CCS.

Due to space limitations, we will continue our response in the next window.

[Q1(Suggestion 1)] Edit-level analysis is recommended to explain observed gains.

Response: To clarify the sources of HIDE’s gains, we directly matched the finding–polarity tuple of each authorized edit against the CheXbert-14 labels extracted from the physician-authored reference report.

DatasetEdit TypeCorrect AdditionsIncorrect AdditionsAccuracy
MIMIC-CXRPositive finding addition7,5291,74681.2%
MIMIC-CXRNegative finding addition78312786.0%
CheXpert PlusPositive finding addition1582785.4%
CheXpert PlusNegative finding addition18481.8%

On MIMIC-CXR, 11,845 edits were evaluable. Positive finding additions accounted for 78.3% of them and achieved 81.2% label consistency (7,529/9,275); negative finding additions accounted for 7.7% and achieved 86.0% consistency (783/910).

On CheXpert Plus, 242 edits were evaluable. Positive finding additions accounted for 76.4% and achieved 85.4% label consistency (158/185); negative finding additions accounted for 9.1% and achieved 81.8% consistency (18/22).

These results identify positive finding additions as the largest source of label-consistent corrections and show that both positive and negative finding additions are generally reliable.

[Q2(Suggestion 2)] Check Figure 4.

Response: We confirmed that the values in this panel of Fig. 4 are correct. Med-R1-2B’s low baseline score makes a given absolute improvement larger in percentage terms.


We hope these clarifications and additional analyses address your concerns and better convey the scope and contribution of our work. We would be grateful if you would consider them in your final assessment.

Reviewer_avjqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

HIDE improves chest X-ray reports without retraining medical MLLM. It extracts complementary findings from multiple stochastic generations, combines cross-sample consistency with localized visual prototype evidence, and applies constrained additions or deletions. Experiments across six generators show improvements in reference-based clinical metrics, including transfer to IU-Xray.

Critical strengths
  • Practical problem and sensible design. The method addresses both omissions and unsupported findings, while recognizing that repeated agreement alone does not establish correctness.
  • Broad evaluation of applicability. Results across multiple generators and datasets, together with component ablations, support the usefulness of combining sampling-based evidence with visual verification.
Critical weaknesses
  • Limited differentiation from prior work. Sampling-based factuality checking, visual verification, and minimally invasive post-editing are established ideas [1–3]. The contribution appears primarily to be their integration for finding recovery. Moreover, Section 2’s characterization of prior methods as operating at the report or decoding-path level overlooks the sentence-level capabilities of RadFlag and SCUQ-RRG.
  • Insufficiently controlled comparisons. Direct comparisons with competing methods cover only three generators on MIMIC-CXR. HIDE additionally uses multiple samples, labeled prototypes, auxiliary models, and corrector fitting, but resource and supervision matched comparisons are not sufficiently documented.
  • Claims exceed the demonstrated scope. “Without retraining the MLLM” is supported, whereas an unqualified “training-free framework” obscures the fitted auxiliary predictor and thresholds. Report-level metric gains also do not directly establish the precision of recovered findings or the risk of erroneous additions and deletions.

[1] Manakul, P., Liusie, A., and Gales, M. J. F. (2023). SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models. Proceedings of EMNLP 2023. [2] Yin, S., et al. (2023). Woodpecker: Hallucination Correction for Multimodal Large Language Models. arXiv preprint arXiv:2310.16045. [3] Gao, L., et al. (2023). RARR: Researching and Revising What Language Models Say, Using Language Models. Proceedings of ACL 2023.

Questions for authors
  1. What is the essential methodological advance over existing approaches? Please discuss [1–3] and clarify the distinction from RadFlag, SCUQ-RRG, and CCS beyond the use of finding-level processing.
  2. Were the comparisons resource matched? Did CCS use the same sampled candidate pool and sampling budget? Can you include finding-level voting and a simple verifier using the same visual features, fitting data, and editor?
6 comments

We sincerely thank the reviewer for recognizing HIDE’s practical design, its joint treatment of omitted and unsupported findings, and its use of visual verification beyond repeated agreement. We also appreciate the positive assessment of our broad evaluation and component ablations across multiple generators and datasets.We have uploaded the revised paper.


The reviewer’s concerns focus on our positioning relative to prior work and supplementary experimental details, as clarified below.

Weaknesses

[W1,Q1] Insufficient clarification of differences from prior work

Response: Prior methods primarily assess or revise an existing response, or select one complete report from a candidate pool. In contrast, HIDE focus on recovering omitted findings across multiple reports sampled from the same frozen generator and using localized image evidence to decide which individual findings to add to the greedy draft.

MethodProblem addressedContent under considerationRole of multiple decodesUnit of decision
SelfCheckGPTAssessment of existing contentStatements in the initial responseCheck consistency across samplesSentence / passage
RadFlagAssessment of existing contentStatements in the candidate reportMeasure support across sampled reportsSentence / report
SCUQ-RRGAssessment of existing contentThe generated report and its sentencesEstimate uncertainty from cross-sample consistencySentence / report
WoodpeckerRevision of existing contentObjects and attributes mentioned in the responseUses visual checks prompted by the initial responseClaim / passage
RARRRevision of existing contentClaims in the initial passageUses external evidence retrieved for those claimsClaim / passage
CCSSelection among candidate reportsComplete sampled reportsRank candidates by pairwise consensusReport
HIDERecovery of omitted contentFindings across the greedy draft and sampled reportsDiscover complementary findings and measure their recurrenceFinding

We have revised Sec. 2 to better articulate the difference of our method (Lines 139-144, 146-154).

[W2.1] Direct comparisons cover a subset of the evaluated models and datasets.

Response: We added comparisons with all competing methods for the remaining three generators on MIMIC-CXR and all six on CheXpert Plus (See below).

HIDE achieves competitive clinical and NLG scores, with more consistent gains across generators and datasets. These gains reflect its ability to recover omissions across samples and filter unsupported findings using local image evidence.

We have added these results in Appendix J (Lines 1382-1389, 1404-1457).

Clinical and NLG metrics on MIMIC-CXR (first two base models).

MethodCheXbert₁₄ ↑CheXbert₅ ↑RadGraph-F1 ↑RaTEScore ↑RadEval-BERT ↑ROUGE-L ↑BLEU-1 ↑BLEU-4 ↑METEOR ↑CIDEr ↑BERTScore ↑
Lingshu-7B0.34390.37080.20210.51220.20750.27470.17370.02320.16830.98220.5037
+ICD0.34380.37040.20800.51540.20170.27590.18220.02510.17240.98700.5083
+VCD0.38710.39780.20250.50000.19450.26940.17090.02090.16290.94190.4965
+DECO0.32870.34100.20740.51100.19900.28020.19210.02550.17800.97930.5074
+CCS0.35160.36510.19420.50160.20120.26220.15020.01600.15190.91810.4935
+CCD0.35370.37850.16810.49920.16270.18060.10250.01390.10960.76240.4395
+HIDE0.45910.48630.21960.52610.21310.29090.22400.02560.19471.01220.5137
MedVLM-R1-2B0.21700.13860.10440.44190.19300.21480.22140.00570.18350.41640.3600
+ICD0.24030.19390.10760.44440.18790.21590.21820.00570.18890.44010.3661
+VCD0.21850.14130.09870.43960.18690.20960.22060.00560.17960.39940.3580
+DECO0.21220.15930.09540.43470.17860.19630.19550.00240.15630.37670.3339
+CCS0.21280.17790.09070.44320.17530.19820.20710.00540.16150.40120.3661
+CCD0.24570.25620.11490.48500.18260.22820.21740.00540.18980.49060.3862
+HIDE0.31650.31650.11930.46180.19740.22060.22200.00570.19180.48050.3717

Due to space limitations, we will continue our response in the next window.

Clinical and NLG metrics on MIMIC-CXR (remaining four base models).

MethodCheXbert₁₄ ↑CheXbert₅ ↑RadGraph-F1 ↑RaTEScore ↑RadEval-BERT ↑ROUGE-L ↑BLEU-1 ↑BLEU-4 ↑METEOR ↑CIDEr ↑BERTScore ↑
MedGemma-4B-IT0.47830.45840.20080.52520.20900.25550.22880.02790.20300.81070.4936
+ICD0.49170.48590.22310.53880.20970.26740.23370.03200.20940.89870.4831
+VCD0.52100.49510.20560.51640.20820.24740.20900.02390.19050.82140.4840
+DECO0.49620.49770.20600.52370.21070.25630.21470.02250.19570.78310.4572
+CCS0.48710.45940.20420.52730.20960.25820.22960.02790.20400.82250.4970
+CCD0.51430.50910.14270.54820.19920.16880.07260.00940.17700.69210.2343
+HIDE0.51430.51730.22690.55130.20970.28280.28130.03200.22990.90760.5092
LLaVA-RAD-7B0.56340.57410.27470.54940.25430.32480.29130.06430.25981.26950.5412
+ICD0.56000.57460.27590.54730.25510.32440.28920.06340.25881.27120.5422
+VCD0.55050.54470.25960.54300.25820.31020.28370.05040.24891.14990.5324
+DECO0.55530.55800.27780.54770.24260.31020.26880.05080.24181.18160.5264
+CCS0.57260.58550.27600.55100.25470.32190.28590.05760.25361.24790.5402
+CCD0.51630.52530.24730.53930.24300.29790.27120.05270.23521.11050.4721
+HIDE0.58580.59770.27980.55450.25760.33030.30360.06550.26471.28550.5439
Med-R1-2B0.20980.13570.08830.43830.18990.20570.22670.00370.17400.38080.3455
+ICD0.23910.18320.09970.44380.18750.20720.22420.00400.18030.41320.3501
+VCD0.22040.13330.08830.43810.18690.20430.22290.00420.17730.38110.3523
+DECO0.19800.13140.08960.43430.18630.19280.19190.00230.15600.34410.3296
+CCS0.21470.15280.07910.43550.17740.19040.21160.00400.16250.35490.3603
+CCD0.24740.26130.09980.45070.17950.22280.25290.00770.18390.42130.3552
+HIDE0.30530.29650.11000.45410.19310.21520.22580.00400.18410.42550.3616
CheXAgent-2-3B0.41040.44110.12890.42780.16390.15210.05070.00520.07840.44250.3180
+ICD0.29950.35620.09070.36530.14410.09720.01280.00100.04220.33720.2734
+VCD0.46350.48930.13930.42880.16900.15730.06800.00700.08770.45640.3285
+DECO0.41740.47250.12960.45300.16240.15790.06970.00430.08540.38130.2963
+CCS0.42550.44840.12700.42360.16350.14610.04960.00450.07480.43450.3162
+CCD0.40780.46160.11140.39800.16450.13780.04240.00380.06790.41470.3130
+HIDE0.52200.57310.17220.46780.17090.19340.12400.00940.12080.49300.3457

Clinical and NLG metrics on CheXpert-Plus (Lingshu-7B).

MethodCheXbert₁₄ ↑CheXbert₅ ↑RadGraph-F1 ↑RaTEScore ↑RadEval-BERT ↑ROUGE-L ↑BLEU-1 ↑BLEU-4 ↑METEOR ↑CIDEr ↑BERTScore ↑
Lingshu-7B0.28070.29700.15060.46530.19790.22100.12720.01030.12790.74400.4438
+ICD0.30160.30090.13620.45850.17100.14880.02540.00080.06170.56250.3085
+VCD0.32410.24780.15540.46800.18350.16690.03270.00250.06660.62240.3180
+DECO0.35250.33930.18010.46370.19500.19210.05320.00240.08550.55040.2874
+CCS0.33730.36210.18010.48920.17790.15860.02520.00090.06770.63970.2827
+CCD0.28350.32730.16370.47160.18120.18640.04340.00190.07750.67370.3224
+HIDE0.37690.42280.18150.47420.20000.23520.15740.01090.14340.79310.4569

Due to space limitations, we will continue our response in the next window.

Clinical and NLG metrics on CheXpert-Plus (remaining five base models).

MethodCheXbert₁₄ ↑CheXbert₅ ↑RadGraph-F1 ↑RaTEScore ↑RadEval-BERT ↑ROUGE-L ↑BLEU-1 ↑BLEU-4 ↑METEOR ↑CIDEr ↑BERTScore ↑
MedGemma-4B-IT0.43720.33680.18450.50700.21360.23910.20970.01800.17730.76650.4510
+ICD0.48330.38100.19980.50850.20990.18630.06820.00350.09720.71980.3150
+VCD0.47450.33640.19280.50140.20120.18070.05640.00300.08960.69420.3192
+DECO0.47620.42480.19630.50220.20170.18900.06390.00390.10070.72150.3135
+CCS0.47480.40380.19420.50300.17890.16790.05410.00340.09460.65960.2860
+CCD0.48330.37210.15790.50850.21420.19290.10770.00670.16090.77150.2893
+HIDE0.48670.38100.19980.51430.21930.24780.22460.01810.19160.82210.4535
MedVLM-R1-2B0.19730.13790.09900.43710.20200.19790.16820.00230.15010.40380.3309
+ICD0.21180.17700.11160.44920.20690.19180.14030.00240.11900.51190.2757
+VCD0.18180.16840.08840.44360.21220.18750.13750.00230.11680.47200.2582
+DECO0.20080.12000.08240.44400.21220.18460.10430.00200.09920.44170.2544
+CCS0.18180.16670.09030.45330.21310.17300.11140.00070.09660.49920.2738
+CCD0.22630.26670.11880.46210.20690.19810.13650.00220.11340.56010.3014
+HIDE0.30830.31150.12170.45810.21430.20230.17740.00300.16310.46830.3405
LLaVA-RAD-7B0.51130.51610.22300.52120.22510.26620.19760.01850.17400.98170.4806
+ICD0.47350.51240.22710.52110.18970.17830.04500.00350.08640.81540.3140
+VCD0.47140.44620.22160.50890.18150.16930.04270.00240.08140.66240.2964
+DECO0.52170.50000.22510.52170.18990.17780.03700.00150.08030.79850.3231
+CCS0.51530.53150.23170.51440.18100.18060.03840.00190.08510.79280.3056
+CCD0.32700.38980.17690.51550.15540.13390.03690.00340.06640.57350.1215
+HIDE0.58220.51850.23960.52400.23310.27650.20710.02040.17801.00080.4884
Med-R1-2B0.14680.02500.08390.42690.21040.19110.19010.00150.15180.37910.3423
+ICD0.24800.23640.10320.45020.21300.18910.13000.00150.11410.53580.2751
+VCD0.15810.04300.07790.43240.19960.17630.12600.00240.10750.45910.2599
+DECO0.14220.08420.07560.43380.20800.17340.09670.00160.09870.36220.2348
+CCS0.14720.08600.05680.43620.21160.16300.11010.00190.09920.40250.2475
+CCD0.16670.20470.10500.47740.20800.19670.14700.00200.11500.57480.3012
+HIDE0.31750.33330.11550.45250.21770.20560.19110.00240.16550.45020.3656
CheXAgent-2-3B0.48510.45310.20890.48550.20940.19310.10140.00780.10600.63530.3222
+ICD0.42640.48780.16110.45830.18240.09140.00890.00030.03220.56820.1930
+VCD0.51900.52240.22080.50390.21080.13840.02630.00390.05830.62100.2378
+DECO0.51320.54900.20350.49700.22000.12960.02930.00070.05260.46710.2125
+CCS0.50520.45590.23010.51570.21080.14190.02100.00230.05570.59370.2401
+CCD0.47480.48000.18350.49300.18790.11870.01240.00130.04540.57470.2198
+HIDE0.55970.55780.24950.52620.21560.23800.15320.01000.14690.69030.3661

Due to space limitations, we will continue our response in the next window.

Reviewer_DoyBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the hallucination and omission problems in chest X-ray report generation by medical MLLMs, and proposes HIDE, a training-free correction framework. The core idea is that greedy decoding produces only a single output, while correct findings for the same image may be hidden in the model's other stochastic samples. HIDE therefore first extracts candidate findings from one initial draft and several stochastic samples, then verifies each candidate using two types of evidence — cross-sample support and localized prototype-based visual comparison — and finally applies minimal edits to add supported findings and remove unsupported claims. Across six frozen medical MLLMs and three datasets, clinical metrics improve substantially, and the correction signal transfers across datasets.

Critical strengths
  1. The proposal has some novelty. It treats the frozen report generator as a "proposal distribution" rather than the final answer, turning the correction problem from "retraining the model" into "recovering and verifying findings from existing outputs," so report quality can be improved without retraining.

  2. The module design is careful and practical. For example, correction actions are divided into branches such as ADD (adding omitted findings) and DEL (removing hallucinated claims), with a high threshold for deletion; edits are minimal and must reuse source sentences from the stochastic samples, which reduces the risk of introducing new errors.

  3. The experiments are fairly comprehensive. They cover six report generators, three datasets, and eleven metrics, and also include ablation studies, experiments with limited fitting data, and cross-dataset transfer evaluation. Data leakage is also avoided when constructing the prototype bank.

Critical weaknesses
  1. The "training-free" claim is somewhat of an overclaim. Although the generator remains frozen, the corrector still needs to be fitted on the target dataset, with thresholds tuned per model and per branch. It is essentially only "training-free with respect to the report generator."

  2. RadSCR[1] is a highly related work, and the two share a similar paradigm: first propose candidate findings, then verify them against localized visual evidence and "alternatives," and finally assemble the verified findings into the final report. This weakens the novelty of this paper to some extent (though I do appreciate the idea of "recovering information from the sampling distribution of a frozen generator").

  3. Since a consensus predictor has already been trained, why not directly examine all possible abnormalities (at least as a baseline)? Would this improve recall? Having such experimental results for comparison would better demonstrate that "sampling from the frozen generator" is necessary.

References:

[1] Yan, Sixing, et al. "Learning self-critiquing mechanisms for region-guided chest x-ray report generation." International Conference on Learning Representations. Vol. 2026. 2026.

Questions for authors

Please see weaknesses.

4 comments

We sincerely thank the reviewer for the careful and constructive evaluation. We are especially pleased that the reviewer recognized HIDE’s central idea of treating a frozen report generator as a proposal distribution. We also appreciate the recognition of our practical module design and the competitive performance of HIDE across multiple medical multimodal language models.We have uploaded the revised paper.


The reviewer’s concerns focus on terminology, related work, and a missing controlled baseline. We address each point below.

Weaknesses

[W1] Potentially overstated “training-free” claim

Response: We agree that “training-free” should be qualified. The term refers to the report generator, as reflected in the title and abstract and explained in Appendix F. All generator and vision–language parameters remain frozen, and HIDE samples candidates from the generator’s unchanged distribution. Only the ADD/DEL predictor requires fitting.

We have clarified this scope in the main text(Lines 18, 88, etc.).

[W2] A paradigm similar to RadSCR’s

Response: Although both methods propose and verify findings, their underlying mechanisms differ. RadSCR focus on grounding proposed abnormalities in in a specific image region. In contrast, HIDE goes beyond candidate verification by exploiting the complementary clinical findings encoded in stochastic outputs of a frozen MLLM. A direct comparison is as follows.

RadSCRHIDE
Candidate sourceRegional abnormality proposals and a whole-chest check for missed findings.Findings the same generator expresses in other samples but omits from its greedy draft.
VerificationComparisons with alternative abnormalities and patient images refine sentence retrieval scores.Cross-sample recurrence and localized image evidence guide each edit.
Report revisionAssembles retrieved sentences into a new report.Combines verified findings from different samples through local edits to the greedy draft.

We have revised Sec. 2 to clarify HIDE’s contribution relative to RadSCR (Lines 141-144).

[W3] A baseline enumerating all potential abnormalities is not included.

Response: We added two exhaustive baselines. For each image, both test all unreported findings from the 12 categories without sampling, using the same 0.5% fitting cohort.

1. Exh_OPR₀ retains the original predictor, sets kb(g)=0k_b(g)=0, and recalibrates branch-specific thresholds.

2. Exh_VPR removes kb(g)k_b(g) and refits the predictor using only finding identity and visual margin.

As shown in the table below, HIDE achieved the best overall clinical performance. Exh_OPR₀ sometimes improved recall but introduced more unsupported findings. Refitting the predictor in Exh_VPR did not close the gap.

Exhaustive enumeration can recover findings missed by sampling, but the larger candidate pool also creates more opportunities for unsupported additions. It can also make predictor fitting harder. With many examples of findings that should be skipped, the predictor may also skip findings that were genuinely missed. This could explain why Chexbert14-F1 fell below the baseline in some exhaustive settings. Sampling narrows the candidate pool to findings the frozen generator expresses for that image, provides cross-sample evidence for verification, and supplies source text for authorized additions.

We have added both experiments to Appendix G (Lines 1169-1241).

Stochastic sampling and exhaustive enumeration on MIMIC-CXR (first three base models).

MethodCheXbert₁₄ ↑CheXbert₅ ↑RadGraph-F1 ↑RaTEScore ↑RadEval-BERT ↑Precision ↑Recall ↑
Lingshu-7B0.34390.37080.20210.51220.20750.44620.2798
+Exh_OPR₀0.24470.30130.19690.51820.19390.17530.4051
+Exh_VPR0.19760.22280.18390.50470.19710.17350.2296
+HIDE0.45910.48630.21960.52610.21310.42130.5044
MedGemma-4B-IT0.47830.45840.20080.52520.20900.54920.4237
+Exh_OPR₀0.26440.30160.17910.53980.18990.18170.4853
+Exh_VPR0.23430.20990.16130.50560.19590.20920.2662
+HIDE0.51430.51730.22690.55130.20970.50760.5212
MedVLM-R1-2B0.21700.13860.10440.44190.19300.30090.1697
+Exh_OPR₀0.30170.31710.11240.52460.19270.21100.5291
+Exh_VPR0.23850.31290.11180.47990.19690.20790.2796
+HIDE0.31650.31650.11930.46180.19740.34420.2930

Due to space limitations, we will continue our response in the next window.

Stochastic sampling and exhaustive enumeration on MIMIC-CXR (remaining three base models).

MethodCheXbert₁₄ ↑CheXbert₅ ↑RadGraph-F1 ↑RaTEScore ↑RadEval-BERT ↑Precision ↑Recall ↑
LLaVA-RAD-7B0.56340.57410.27470.54940.25430.59880.5320
+Exh_OPR₀0.23050.20800.22060.53720.21060.16390.3879
+Exh_VPR0.21460.13840.21650.51460.22570.18730.2512
+HIDE0.58580.59770.27980.55450.25760.57280.5993
Med-R1-2B0.20980.13570.08830.43830.18990.27160.1709
+Exh_OPR₀0.29800.28340.10270.44700.18980.20700.5317
+Exh_VPR0.22810.30150.09840.47510.19250.20760.2531
+HIDE0.30530.29650.11000.45410.19310.34280.2751
CheXAgent-2-3B0.41040.44110.12890.42780.16390.53570.3326
+Exh_OPR₀0.30860.37380.13110.47400.17140.23440.4517
+Exh_VPR0.20030.29620.09190.44610.16560.17340.2372
+HIDE0.52200.57310.17220.46780.17090.48160.5696

Stochastic sampling and exhaustive enumeration on CheXpert Plus (all six base models).

MethodCheXbert₁₄ ↑CheXbert₅ ↑RadGraph-F1 ↑RaTEScore ↑RadEval-BERT ↑Precision ↑Recall ↑
Lingshu-7B0.28070.29700.15060.46530.19790.40000.2162
+Exh_OPR₀0.17680.28110.16460.47480.21070.12840.2838
+Exh_VPR0.13390.15840.14210.46870.19750.16040.1149
+HIDE0.37690.42280.18150.47420.20000.43750.3311
MedGemma-4B-IT0.43720.33680.18450.50700.21360.54550.3649
+Exh_OPR₀0.23300.27590.16980.50510.21510.16350.4054
+Exh_VPR0.25440.19130.15080.48250.21860.26670.2432
+HIDE0.48670.38100.19980.51430.21930.55650.4324
MedVLM-R1-2B0.19730.13790.09900.43710.20200.29330.1486
+Exh_OPR₀0.23870.32200.12870.50870.21390.17910.3581
+Exh_VPR0.18320.20370.10830.45730.21040.20000.1689
+HIDE0.30830.31150.12170.45810.21430.34750.2770
LLaVA-RAD-7B0.51130.51610.22300.52120.22510.57630.4595
+Exh_OPR₀0.22140.20530.18800.51810.22530.15320.3986
+Exh_VPR0.28470.26320.17420.48090.21070.29290.2770
+HIDE0.58220.51850.23960.52400.23310.57640.5881
Med-R1-2B0.14680.02500.08390.42690.21040.22860.1081
+Exh_OPR₀0.28910.32930.10830.51860.23030.20450.4932
+Exh_VPR0.19620.25860.09740.45030.21600.22220.1757
+HIDE0.31750.33330.11550.45250.21770.38460.2703
CheXAgent-2-3B0.48510.45310.20890.48550.20940.54170.4392
+Exh_OPR₀0.16410.20970.15070.49600.20720.11440.2905
+Exh_VPR0.21680.22410.11270.44020.20380.22460.2095
+HIDE0.55970.55780.24950.52620.21560.56550.5541

We hope these clarifications and additional analyses address your concerns and better convey the scope and contribution of our work. We would be grateful if you would consider them in your final assessment.

I appreciate the authors for their detailed responses.

The clarification on the "training-free" claim (W1) and the positioning relative to RadSCR (W2) are satisfactory, and the newly added exhaustive-enumeration baselines (W3) convincingly demonstrate the necessity of sampling from the frozen generator. My concerns have been largely addressed, so I will maintain my score.

On behalf of all authors, we sincerely thank you for your careful evaluation of our rebuttal and manuscript.

We are pleased that our clarifications addressed your main concern, and we appreciate your constructive suggestion for further strengthening the work. We greatly appreciate your expertise and the time you devoted to reviewing our work. Your continued support will encourage further research on radiology report generation and faithful medical multimodal generation.

Comments on the paper

1 comment

We thank all reviewers for their detailed and constructive feedback and for the time and effort they devoted to evaluating our work.

In their reviews, they expressed strong appreciation for HIDE’s central idea of treating a frozen report generator as a source of complementary findings. A finding omitted from the greedy draft may still appear in another sample from the same generator. Reviewers also recognized the value of localized image evidence for verifying these findings, finding-level edits for combining them across reports, and the evaluation across multiple generators and datasets.

Alongside these strengths, the reviewers identified four areas where clearer explanations and additional experiments would strengthen the paper. We describe how we addressed each below.

  1. Additional experiments (Reviewer DoyB W3, Reviewer avjq W2.1, Q2.3).

    • To test whether exhaustive enumeration could replace sampling, we evaluated two exhaustive baselines (24 experiments). HIDE achieved a better overall balance between precision and recall. This supports our core idea of recovery through sampling.
    • To address gaps in baseline coverage, we added 45 comparisons with five competing methods, showing competitive performance across all six generators on MIMIC-CXR and CheXpert Plus.
    • To assess the predictor design, we evaluated voting-only and visual-only variants on the same candidates (24 experiments). The full predictor achieved higher CheXbert-14 F1 than either variant across all six generators on both datasets. These results support our predictor design.
  2. Scope and contribution (Reviewer DoyB W1 and W2, Reviewer avjq W1, Q1, W3.1).

    • We further clarified the scope of “training-free” by describing HIDE as a frozen-generator correction framework. The generator remains frozen, and only the predictor requires fitting.
    • We distinguished HIDE from prior work in the source of candidate findings, the role of multiple samples, and how verified findings enter the report.
  3. Implementation and resource details (Reviewer gAgZ W1, W2, Q2, Reviewer avjq W2.2, Q2.1, Q2.2).

    • We expanded the descriptions of HIDE’s editing module and predictor, documented baseline implementations and resource requirements, and verified that the values in Figure 4 are correct.
  4. Edit-level analysis (Reviewer gAgZ Q1, Reviewer avjq W3.2).

    • We assessed individual edits against CheXbert-14 labels extracted from physician-authored reference reports. Positive finding additions accounted for the largest share of reference-consistent corrections, with agreement rates of 81.2% on MIMIC-CXR and 85.4% on CheXpert Plus.

We have responded to every concern raised by the reviewers and uploaded the revised manuscript. We thank the reviewers again for their suggestions, which helped us strengthen the paper. We respectfully ask that the information provided in our rebuttal be considered in the final assessment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.