acceptodds
Withdrawn from ICLR 2027

ClinOPD: On-Policy Distillation with Clinical Hints for Medical Report Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Automatic medical report generation aims to generate clinically accurate reports from medical images. Existing approaches learn the mapping from images to reference reports, but generated reports may still miss clinically important findings or introduce unsupported abnormalities. Inspired by the clinical reporting workflow, we introduce Clinical Hints, a compact signal that retains only the organs associated with positive findings. We find that providing such hints to the same model substantially improves report generation, suggesting that relevant clinical knowledge is often already encoded in the model but is not reliably activated from medical images alone. Based on this observation, we propose ClinOPD, an on-policy self-distillation framework that uses clinical hints as privileged supervision. A frozen hint-conditioned teacher provides token-level guidance along student-generated trajectories, while the student takes only the medical image as input and is regularized toward the original policy for stable optimization. No clinical hints are required at inference time. We evaluate clinical correctness using a factualness-driven F1 metric based on matched, missed, and unsupported findings, which shows stronger agreement with physician judgments than conventional report-generation metrics. Across the 2D IU-Xray and 3D CT-RATE datasets, ClinOPD improves F1 from 46.06 to 51.12 and from 23.65 to 30.06, corresponding to relative gains of 11.0% and 27.1% over SFT, respectively, while outperforming competing reinforcement-learning and on-policy distillation methods. Code will be made available.

Primary area:
reinforcement learning
Keywords:
report generation, multimodal large language models, on-policy distillation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_YZXgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors propose ClinOPD, an on-policy self-distillation framework for radiology report generation using medical multimodal LLMs. It introduces clinical hints identifying organs with positive findings to guide a teacher model, whose behavior is distilled into a hint-free student using KL-based objectives. The method does not require hints at inference. Authors evaluated ClinOPD on IU-Xray and CT-RATE against SFT, GRPO, DAPO, and OPSD baselines. ClinOPD improves clinical F1 scores over the baselines, with additional experiments validating the proposed metric and analyzing the effects of regularization and clinical hints.

Critical strengths
  1. Well-motivated approach: The paper introduces a simple and effective self-distillation framework that does not require clinical hints at inference.

  2. Strong empirical evidence: Clinical hints improve F1 by 26.4 and 28.1 points on IU-Xray and CT-RATE, respectively, using the same checkpoint.

  3. Detailed analysis: Additional experiments demonstrate the importance of organ-specific hints, and ablations highlight the role of regularization.

Critical weaknesses
  1. Tables 1–3 report results from single runs without standard deviations or confidence intervals. The small improvement over GRPO on IU-Xray (+0.62 F1) may not be statistically significant.

  2. The evaluation relies entirely on Clin-F1, an LLM-based metric. Established clinical metrics such as RadGraph-F1, CheXbert-F1, and GREEN are not reported.

  3. ClinOPD recovers only approximately 19% and 23% of the hint-conditioned improvements on IU-Xray and CT-RATE, respectively. This limits the strength of the claimed distillation benefits.

  4. The paper does not sufficiently evaluate against existing trained radiology report generation systems. A standard image-only SFT baseline is also missing, making it difficult to assess the benefits of mixed-SFT training.

Questions for authors
  1. What are the mean ± standard deviation results over multiple seeds for all methods, and what are the test-set sizes and checkpoint selection criteria?

  2. What proportion of test cases are normal, and how does a constant normal-report baseline perform under Clin-F1?

  3. How does ClinOPD perform under established clinical metrics such as RadGraph-F1, CheXbert-F1, or GREEN? Are the method rankings consistent across these metrics?

  4. How does mixed-SFT compare against standard image-only SFT, and were the competing methods tuned under comparable settings?

Reviewer_hBAbOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposed a medical report generation method ClinOPD which is an on-policy self-distlllation framework that uses clinical hints as supervision signal. As the clinical hints are not available during inference, so they design a teacher student architecture where the teacher use ground truth clinical hints and provide token level guidance along student generated trajectories and the student takes only the medical image as input and is regularized toward the original policy for stable optimization.

Critical strengths
  1. the paper is well organized, and the writing is clear
  2. consistent empirical improvements across different medical imaging modalities.
Critical weaknesses
  1. questionable validity of clinical hints the central assumption behind clinical hints is insufficiently justified, radiology reports are selective descriptions influenced by clinical indications, patient history and radiologist's reporting decisions, extracting hints from positive finds in reference reports may encode dataset specific reporting habits rather than true anatomical attention.
  2. the evaluation protocol is problematic. Clin-F1 labels any abnormality absent from the reference report as hallucination, even though a minor or unrelated abnormality may still be visible in the image but omitted by the radiologist. Similarly, the physician evaluation only compares generated text against reference text without the clinicians inspecting the actual medical images. Thus both evaluations measure text to text alignment with reference reports rather than image grounded clinical accuracy.
  3. As hints are extracted entirely from ground-truth reports, the teacher model provides no new clinical information beyond existing training supervision, the sft baseline in table 3 is a mixed-sft trained on 50% image-only and 50% hint-conditioned examples, rather than 100% image only sft. Training with gt hints on half the data risks inducing shortcut relianceon text prompts and weakening the visual part. Without a standard 100% image-only SFT baseline trained with matched compute, it is unclear whether ClinOPD truly outperforms direct SFT or merely recovers performance lost to the mixed-training setup. And becaue the hints only list organs with positive abnormalities, the gains in table1 may from telling the model which organs are abnormal rather than improved visual recognition.
Questions for authors
  1. for table 3, the sft baseline is trained on 50% image only and 50% hint conditioned data, can you report results for a standart sft model trained on 100% image-only inputs?
  2. for the evaluation problem, how does your evaluation distinguish true hallucinations from valid, visually present abnormalities that the radiologist chose not to mention? Can the clinicians re-evaluate a subset of cases while inspecting the actual images? 3.figure 2 only measures the teacher-forced probability of a single positive finding, what happens to overall precision and Clin-F1 in full report generation (Table 1) if the hint lists all major organs (both normal and abnormal) or uses the unrelated organs from P3, to verify whether the model hallucinates false abnormalities in normal hinted organs?
Reviewer_7B3vOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether providing just a list of affected organs instead of a full medical report as privileged information improves the performance of OPD in radiology report generation. The authors show that ClinOPD improves the performance on two datasets compared to SFT, GRPO, DAPO, and OPSD.

Critical strengths

Simple method with evaluation on two datasets and compared against several baseline methods.

Critical weaknesses

Minimal methodological novelty and failure to isolate the proposed design choice

  • Very limited methodological novelty and barely any new insights in privileged information design or other interesting questions. The only change over existing work is to use a list of affected organs instead of a full report as privileged information in a OPD setup. No other constructions of privileged information are examined or compared (eg organ + finding, condensed report, ...), nor is the selected choice sufficiently investigated. Additionally, the ablations focus on rather established components (length and KL-regularization) rather than changes introduced by ClinOPD (privileged information design). This raises the question of what contributions the paper provides.
  • The "motivating" experiment (Table 1) fails to provide a suitable motivation, and does not answer the question "Why do clinical hints help?" (line 242). Providing the list of affected organs to a model unsurprisingly raises the F1 score, but eg providing the full report would likely raise the performance even more. However, this is a worse privileged information in a OPD setup as the authors show, raising the question of the supposed information gain provided by this experiment.
  • The claim "the model already posses medical knowledge, it only needs to be directed toward relevant information" is not demonstrated sufficiently. E.g., what happens when incorrect organs are mentioned as the ones to "focus on"? Perhaps it would "report" findings in those as well, making it more a filter for guessing, instead of leveraging some kind of existing knowledge.

Experimental evidence insufficient to substantiate the contribution

  • All methods compared in Table 3 start from the same "mixed-SFT" checkpoint. The "mixed" nature of that stage is likely to benefit ClinOPD (though not ablated!), but could potentially hurt the other methods (GRPO, DAPO) since the training with organ-list is wasted / confusing. This becomes especially impactful when comparing to OPSD, which receives full reports as privileged information (clear mismatch in teacher capabilities between OPSD and ClinOPD). E.g. starting from a "pure-image SFT" checkpoint would be appropriate. Without it, it is unclear whether the experimental setup is biased in favour of ClinOPD.
  • The experiments seem to be single-seed runs, and the method is, as showed in Figure 6, subject to rather noisy performance improvements. Therefore, multi-seed runs, ideally with confidence intervals or std would be beneficial to evaluate the method's robustness.
  • Figure 2 is hard to understand (P1, P2, ..., differences of probabilities instead of absolute ones), and not compatible to Appendix G: Figure 2 reports differences in probabilities, while Appendix G shows differences in log probabilities. Additionally, Appendix G reports results on three samples (why not more?) with high variability, and Figure 2 shows only the most favorable one. This raises questions of how the three samples where selected and what the aggregated results would look like on a larger set.
Questions for authors
  • How are the results in Table 3 generated? Is each method evaluated after 100 steps? Or every X steps on a validation set?
  • How are the other methods performing over the number of steps (like for ClinOPD in Figure 6 b)? Are they constantly improving or as noisy as ClinOPD?

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

Related papers

Discussion (0)

Sign in to comment.