acceptodds
Withdrawn from ICLR 2027

Variational Finding-Anchored Direct Preference Optimization for Radiology Report Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Radiology Report Generation (RRG) aims to generate diagnostic reports from medical radiology images and has achieved substantial progress on benchmark datasets. Nonetheless, existing RRG methods may produce clinically inconsistent reports, with omitted findings and imperfect entity-relation representations. Direct Preference Optimization (DPO) and its variants offer a way to improve report generation through preference alignment. However, adapting DPO to finding‑specific preference alignment remains challenging. In this work, we propose a Variational Finding-Anchored Direct Preference Optimization (VF-DPO) approach that introduces finding-aware Question Answer pairs as latent variables to anchor the generation process. Specifically, we approximate the conditional report generation probability with finding anchors as conditions via its Evidence Lower BOund (ELBO), with which we derive a novel DPO formulation in a unified preference optimization objective that jointly models preference in latent finding-state alignment and high-fidelity report generation. To support this bi-level alignment, we construct a corresponding bi-level preference dataset comprising both finding-anchored latent variables and report sequences for training VF-DPO. At test time, we use the trained model to directly generate the radiology report, where the finding-anchored latent variables are implicitly accounted for in training, without being explicitly generated in the testing phase. Extensive experiments on MIMIC-CXR (X-ray) and CT-RATE (CT) demonstrate that VF-DPO achieves the highest F1 and RadGraph scores across all backbone comparisons, while obtaining the best RadCliQ and RaTEScore results on most backbones. These results indicate that VF-DPO enhances overall clinical consistency and outperforms state-of-the-art preference optimization methods for RRG.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Preference Optimization, Radiology Report Generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Co2iOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Variational Finding-Anchored Direct Preference Optimization (VF-DPO) as an improvement over DPO with a focus on improving clinical consistency of radiology reports by introducing finding-specific question-answer pairs as latent anchors in the training process. The method is evaluated on MIMIC-CXR and CT-RATE across eight model backbones and compared against five (MIMIC-CXR) or four (CT-RATE) other preference optimization methods. The additional computations required for the finding-state alignment and conditional report alignment increase the training time by 2.6x compared to standard DPO.

Critical strengths
  • The method addresses an important limitation of DPO for radiology report generation (finding awareness)
  • Compared against several preference optimization methods
Critical weaknesses
  • Small improvements at high costs: The choice of backbone appears to have a substantially larger impact than the alignment method. On REVTAF, VF-DPO improves RadGraph by only 0.0078 over DPO and 0.001 over the original model, while performing worse than the original model on RaTEScore (Table 1). Overall there are slight performance improvements, but they come with substantially higher costs (2.6x runtime and 2.8x memory compared to DPO). A compute-matched comparison would be insightful.
  • Unknown robustness of improvements and theoretical justification: The reported significance tests appear to be based on individual checkpoints and not repeated training runs. Given the small improvement, run-to-run variance could exceed the reported differences. There are no information on checkpoint selection and hyperparameter tuning for baselines (nor a reproducibility section or provided code). Furthermore, replacing the log probabilities in the DPO reward margin with individual ELBOs does not guarantee an accurate approximation of that margin, since the approximation gaps may differ. This relationship is not sufficiently justified.
Questions for authors
  • How do the baselines perform with a matched computational budget, e.g. when trained for 2.6x longer?
  • Are the results based on single training runs? Can the authors report variance across seeds and clarify checkpoint selection and baseline hyperparameter tuning?
  • Equation 4 replaces the log probabilities in the DPO reward margin with their respective ELBOs. Since the four approximation gaps may differ, the resulting margin is not neccessarily a lower bound on the original margin and could even have the opposite sign. Have the authors analyzed this approximation error or established conditions, under which preference ordering is preserved?
Reviewer_22C1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes VF-DPO for radiology report generation. It extracts finding-specific QA anchors from preferred and retrieved dispreferred reports, and combines preference margins for finding answers, anchor-conditioned reports, and direct report generation. Experiments cover five CXR and three CT backbones, with generally favorable clinical-metric results.

Critical strengths
  • Finding-state errors are an important limitation of report-level preference alignment.
  • Results across eight architectures and two modalities provide useful evidence of practical applicability.
  • The paper includes multitask-SFT, auxiliary-classification, alignment-target, additional-finding, and computational-cost analyses.
Critical weaknesses
  1. The variational surrogate lacks a sufficient connection to the original DPO objective. Proposition 1 correctly expands the defined ELBO-based margin, but does not establish that it approximates the marginal DPO margin faithfully. The deterministic anchor construction provides no demonstrated control over them. A valid ELBO alone therefore does not justify the resulting preference surrogate.
  2. Equation (1) defines report generation by marginalizing an anchor-conditioned model. However, sharing parameters between QA, anchor-conditioned generation, and direct generation does not ensure that the directly prompted distribution equals this marginal. The added standard DPO term trains direct generation, but does not establish probabilistic consistency. The implemented method currently appears better supported as a joint multitask preference objective.
  3. Table C-2 maps No Mention to statements including “The image does not contain {finding}.” This expresses absence, rather than non-mention in the source report. Other templates similarly convert reporting status into claims about image assessability. This contradicts the stated distinction between No Mention and Negative and can corrupt image-conditioned finding supervision.
  4. Some gains on strong backbones are small, for example, PromptMRG F1 increases from 0.4774 with RRG-DPO to 0.4795, and multiple-seed variability is not reported. The substantially higher training cost also motivates compute-matched comparisons.
Questions for authors

See the weaknesses.

Reviewer_pPq2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents an approach to improving the quality of radiology report generation using DPO method based on findings. As needed for DPO, preferred and dis-preferred lists need to be formed. Unlike prior DPO approaches to report generation that use the full report, this approach goes fine-grained and extracts findings from the reports and forms preferred and dispreferred lists based on the findings that are mentioned versus unmentioned findings in a report. QA pairs are formed for training the VLM from these finding lists. The text generation is now in two stages, estimating the probability of finding conditioned on image and then generated text conditioned on both the finding and image. The loss combines image-conditioned finding-answer preference, anchor-conditioned report preference, and standard image-to-report DPO in one logistic objective. An ELBO motivates the anchor terms. During inference time the model does not need the finding generation as the intermediate step. The method has been evaluated on both CT and Chest X-ray datasets using various lexical and clinical metrics, and comparisons are made to other DPO based training time improvement methods

Critical strengths

1.The idea of using findings to tailor the DPO approach is spot on for the radiology reports. 2.The two step formulation for VLM text generation is also interesting. I wonder if this can be used for VLMs outside the DPO formulation as well. This used to be the early approaches to report generation before autoregressive decoding came along. 3. If the paper can extend in that direction, that may be an interesting contribution. 4. The ablations are more substantial than a comparison with standard DPO alone

Critical weaknesses

The paper as described here in terms of extracting the findings and using them in the DPO supervision is similar in spirit to other papers that have appeared in the past. First, the idea of generating finding first followed by report generation is a long standing approach to radiology report generation (See Ref1 which has earlier references). The context of forming preferred and dispreferred lists from findings is interesting but not as advanced as some of the other methods out there. For example, there are methods that extract fine-grained findings capturing not only the finding name but also presence/absence, anatomy, laterality, severity etc. and have formed real/fake synthetic datasets that are used to train fact-checking models post-inference for RRG models (See Ref.2), Further, they have also bee used to train RRG models using the DPO (called automatic preference optimization or APO) (See Ref. 3). In the light of these existing methods, the paper needs to discuss and perform experiments to do comparison. These methods also attempt spatial grounding of the findings in addition to their textual descriptions.

  1. Chest X-ray Report Generation through Fine-Grained Label Learning, MICCAI 2020, https://arxiv.org/abs/2007.13831
  2. Phrase-grounded Fact-checking for Automatically Generated Chest X-ray Reports, MICCAI 2025, https://arxiv.org/abs/2509.21356
  3. https://huggingface.co/datasets/razi-mahmood/RadCheck
  4. GREEN: Generative Radiology Report Evaluation and Error Notation, https://aclanthology.org/2024.findings-emnlp.21/
  5. CRIMSON: A Clinically-Grounded LLM-Based Metric for Generative Radiology Report Evaluation, https://arxiv.org/abs/2603.06183
Questions for authors

The paper treats unmentioned findings as negatives which is not always the case. Sometimes they are just not relevant to mention or may have been omitted in error.

The method of extraction of findings using ChexBert has limitations in terms of the number of findings that can be extracted from the reports. No details on this has been provided.

The methods for assessing the performance of the RRG model needs to take into account more modern scoring methods such as GREEN, Crimson, and RQ. (see references).

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

Related papers

Discussion (0)

Sign in to comment.