acceptodds
Under review as a conference paper at ICLR 2027

VITAL: Visual and Task Advantage Learning for On Policy Distillation in Medical Diagnosis

Abstract

Medical vision-language models (VLMs) have shown strong diagnostic performance. On-policy distillation (OPD) transfers these capabilities to smaller VLMs for faster inference and easier deployment. However, existing OPD methods often overlook what the student actually lacks. We uncover two counter-intuitive forms of student deficit: (1) Visual Deficit: misaligned rather than simply weaker visual dependence reduces sensitivity to clinical findings, causing diagnostic errors. (2) Task Deficit, where student rollouts are not consistently inferior to the teacher, and indiscriminate imitation overwrites useful student behaviors and impair generalization. Motivated by these observations, we propose VITAL, a VIsual and Task-Advantage Learning framework for on-policy distillation in medical diagnosis. VITAL improves distillation from two complementary perspectives. Specifically, Visual Advantage Policy optimization (VAP) estimates the teacher’s token-level visual dependence through image perturbations and scales distillation accordingly to improve the student’s visual dependence. Task Advantage Distillation (TAD) estimates the teacher's task advantage over each student rollout and rejects rollouts without positive teacher advantage, avoiding redundant or potentially harmful imitation. Experiments across four adult and neonatal datasets show VITAL outperforms prior OPD methods in-domain, while boosting out-of-domain F1 by 4.60% (CheXpert Plus) and 2.79% (NeoCXR-EV) over the strongest baseline.

Primary area:
foundation or frontier models, including LLMs
Keywords:
On-policy distillation, Medical vision-language models, Visual Advantage Policy optimization, Task Advantage Distillation

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_RiAqOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies on-policy distillation (OPD) for medical vision-language models to generate chest X-ray radiology reports. The paper first identifies two empirical failure modes of vanilla OPD: (1) the student incorrectly allocates visual attention toward template tokens rather than clinically relevant tokens, and (2) a notable fraction of student rollouts can outperform the teacher. To address these two issues, the authors propose VITAL, a two-module framework. The VAP module uses image perturbation to compute token-level visual dependency and reweights policy gradients to prioritize clinical tokens. The TAD module selectively filters student rollouts for distillation, retaining only samples where the teacher achieves superior performance measured by ROUGE-L. The method is validated on four chest X-ray datasets including both in-domain and out-of-distribution benchmarks for adult and neonatal radiology. Experimental results show VITAL improves OOD clinical F1 over standard OPD baselines. Ablation studies verify that both VAP and TAD contribute to performance gains.

Critical strengths

1.The paper identifies two non-trivial, empirically validated failure modes for on-policy distillation (OPD) in medical VLM report generation. The authors quantitatively show student models suffer from visual misallocation deficit (visual sensitivity leaks into template tokens instead of clinical anatomy/severity tokens) and task deficit (over 20% student rollouts outperform the teacher). These two counterintuitive observations directly motivate the design of the two modules in VITAL, and the motivating figure provides solid empirical evidence. This insight is valuable for the medical multimodal distillation community, which often assumes the student only lacks visual capacity and the teacher is universally better. 2.The proposed two-module framework achieves strong out-of-distribution performance gains on chest X-ray datasets. VITAL combines token-level visual reweighting (VAP) and rollout-level selective distillation (TAD). Evaluations across 4 chest X-ray datasets, including both in-domain and challenging OOD splits for adult and neonatal radiology, demonstrate consistent OOD F1 improvements compared with strong OPD baselines. Ablation studies confirm both modules contribute to performance, which supports the core claim that the two designed components are complementary. The code is anonymously released for reproducibility.

Critical weaknesses

1.The technical novelty of VAP is limited, and the boundary with Med-OPD is not sufficiently established. Med-OPD already uses image perturbation to estimate token-wise visual evidence. VAP repurposes this perturbation signal as a scaling factor for policy gradient weights. The paper does not include the critical ablation of Med-OPD + TAD. Without this comparison, we cannot confirm whether the performance gain comes from the novel VAP design or merely combining Med-OPD’s visual module with TAD selective filtering. This blurs the core technical contribution of this work. 2. The TAD module’s reliance on ROUGE-L as the proxy metric introduces unvalidated systematic risk for clinical diagnosis. TAD uses ROUGE-L for rollout filtering. ROUGE-L measures text overlap and does not perfectly align with clinical diagnostic correctness. The reported correlation between ROUGE-L and clinical F1 is only provided for NeoCXR. The paper lacks correlation statistics for MIMIC-CXR and CheXpert Plus, and there is no control experiment replacing ROUGE-L with clinical F1 to test robustness. For safety-critical medical tasks, using an unvalidated proxy metric for sample selection is a major concern and weakens the reliability of the proposed TAD module.

Questions for authors

1.The paper states that over 20% of student rollouts outperform the teacher in Figure 1(b), but the metric used to compute this statistic is not specified. Please clarify whether this percentage is calculated using ROUGE-L or clinical F1. Please also update both the main text and figure caption to explicitly state the metric used. 2. The in-domain gains on MIMIC-CXR and NeoCXR are tiny and may fall within random training variance. Could you report results across multiple random seeds with mean ± std and statistical significance tests for all primary benchmarks? Especially for CheXpert Plus, which has only 202 test samples. Please answer: Are the reported ID and OOD improvements statistically significant? 3. Could you add two baselines: (1) simple top-k rollout filtering baseline; (2) VITAL with vanilla SFT teacher (without GSPO). Please answer: How much performance difference exists between TAD and simple top-k filtering? Does VITAL maintain consistent gains when using a non-GSPO SFT teacher? 4. The authors swept σnoise\sigma_{noise} and α\alpha, but temperature parameter τ\tau is missing from hyperparameter ablation. Please add ablation for τ\tau. Please answer: How sensitive is final performance to different choices of τ\tau? 5. The paper only evaluates on chest X-ray. The current claim "for medical diagnosis" is overbroad. Please revise title and abstract to restrict scope to chest X-ray report generation. Also report training compute overhead (GPU hours, runtime) introduced by VAP’s extra perturbed-image forward passes. Please answer: What is the additional training cost of VITAL compared with Native OPD? 6. Could you add more qualitative failure cases in the appendix? Please discuss two risks: (1) VAP only models token-level visual dependency and cannot spatially localize pathology; (2) TAD may learn incorrect teacher outputs when the teacher produces medical hallucinations. Please answer: In what typical failure scenarios does VITAL still fail, and how severe is the teacher hallucination risk for TAD? 7. The two core motivating challenges in the introduction are conceptually reasonable, but their supporting empirical validation is insufficient. For Challenge 1, please quantitatively compare the visual dependency weights on template tokens versus clinically relevant tokens between the Med-OPD baseline and the full VITAL model, to verify that VAP effectively suppresses visual weights assigned to template tokens. For Challenge 2, please disentangle whether performance gains originate from the proposed teacher-relative filtering mechanism, or merely from discarding low-quality rollout samples. Please answer: Can you provide quantitative evidence showing VAP reduces visual attention on template tokens and strengthens visual dependency on clinical tokens? Can you confirm that TAD’s gain comes from the proposed teacher-relative rule rather than simple low-quality sample filtering?

8 comments

Thank you for your thorough and expert feedback. We have done our best to address your concerns below.

1. Task Quality Proxy (ROUGE-L) . We use ROUGE-L, which is consistent with the definition of task advantage in Equation (9). Thank you for pointing this out. We have changed the sub-captions of Figure 1 to "(a) Visual deficit (Visual Sensitivity by Token Category)" and "(b) Task Deficit (ROUGE-L-Based Rollout Win Rates)". In addition, we have highlighted the phrase "task quality (proxied by ROUGE-L scores)," in red on Line 075 in the main text.

2. Mean and Standard Deviation.

As shown in the table below, VITAL achieves consistent results across runs. The standard deviation of the average score is 0.01 on MIMIC-CXR and rounds to 0.00 on NeoCXR-EV. We acknowledge that VITAL’s gains on the in-domain MIMIC-CXR benchmark are modest. However, our method generalizes substantially better than the state-of-the-art method on out-of-domain test sets, improving F1 by 3.66 percentage points on CheXpert Plus and 2.79 percentage points on NeoCXR-EV. We have added this table to session Additional Analysis of VITAL in the appendix.

\hline\hline \textbf{Metric} & \textbf{MIMIC-CXR} & \textbf{CheXpert Plus} & \textbf{NeoCXR} & \textbf{NeoCXR-EV} \\\\ \hline \text{R-L } (\uparrow) & 27.94 \pm 0.02 & 25.03 \pm 0.02 & 46.32 \pm 0.00 & 32.91 \pm 0.08 \\\\ \text{R-1 } (\uparrow) & 29.33 \pm 0.02 & 25.88 \pm 0.01 & 46.80 \pm 0.00 & 33.84 \pm 0.08 \\\\ \text{MET } (\uparrow) & 20.54 \pm 0.00 & 17.02 \pm 0.22 & 45.76 \pm 0.00 & 32.43 \pm 0.17 \\\\ \text{RaTE } (\uparrow) & 51.14 \pm 0.15 & 43.26 \pm 0.24 & 59.91 \pm 0.01 & 50.33 \pm 0.01 \\\\ \text{P } (\uparrow) & 42.68 \pm 0.20 & 38.27 \pm 1.08 & 55.19 \pm 0.03 & 55.14 \pm 0.00 \\\\ \text{R } (\uparrow) & 40.14 \pm 0.03 & 42.93 \pm 0.06 & 55.19 \pm 0.00 & 45.68 \pm 0.00 \\\\ \text{F1 } (\uparrow) & 41.37 \pm 0.11 & 40.47 \pm 0.52 & 55.19 \pm 0.00 & 49.97 \pm 0.00 \\\\ \hline \textbf{Avg. } (\uparrow) & \textbf{36.16} \pm 0.01 & \textbf{33.27} \pm 0.31 & \textbf{52.05} \pm 0.00 & \textbf{42.90} \pm 0.02 \\\\ \hline\hline \end{array}$$ **3. Two Baseline Settings** **1). Simple top-$k$ filtering.** We compared our method with a Top-K baseline that selects the top-2 rollouts based on student task performance, see table below. Although Top-K filters out poor samples, it ignores the hard samples the student truly needs guidance on. Moreover, it imposes constraints on samples where the student already outperforms the teacher, which can degrade the student's capabilities. As a result, this approach yields inferior performance compared to our TAD. We have added this ablation study to session Additional Analysis of VITAL in the appendix. $$\begin{array}{l|cc|cc} \hline\hline \textbf{Method} & \textbf{MIMIC R-L } (\uparrow) & \textbf{MIMIC F1 } (\uparrow) & \textbf{CheXpert R-L } (\uparrow) & \textbf{CheXpert F1 } (\uparrow) \\\\ \hline \text{VAP + Top-K (k=2)} & \text{27.31} & \text{40.01} & \text{24.69} & \text{38.85} \\\\ \textbf{VAP + TAD (Ours)} & \textbf{27.94 } \textcolor{teal}{(+0.63)} & \textbf{41.37 } \textcolor{teal}{(+1.36)} & \textbf{25.03 } \textcolor{teal}{(+0.34)} & \textbf{40.47 } \textcolor{teal}{(+1.62)} \\\\ \hline\hline \end{array}$$ **2). VITAL using only a standard SFT teacher (without GSPO).** Distillation from an SFT teacher proves suboptimal due to the limited performance gap between the teacher and student. As a result, neither VITAL nor native OPD achieves meaningful performance gains in this setting. $$\begin{array}{llccccccccc} \hline \textbf{Dataset} & \textbf{Model} & \textbf{R-L} & \textbf{R-1} & \textbf{MET} & \textbf{RaTE} & \textbf{P} & \textbf{R} & \textbf{F1} & \textbf{Avg} & \textbf{$\Delta$Avg} \\\\ \hline \text{MIMIC} & \text{8B SFT teacher} & 26.58 & 27.95 & 19.85 & 49.96 & 32.14 & 20.35 & 22.79 & 28.52 & \text{---} \\\\ \text{MIMIC} & \text{2B student} & 17.73 & 19.25 & 20.17 & 49.08 & 35.27 & 24.87 & 29.17 & 27.93 & \text{---} \\\\ \text{MIMIC} & \text{Native OPD} & 22.74 & 24.63 & 16.27 & 45.59 & 33.19 & 23.72 & 27.67 & 27.69 & -0.25 \\\\ \text{MIMIC} & \text{VITAL} & 21.02 & 22.51 & 14.90 & 44.12 & 27.20 & 17.28 & 21.14 & 24.03 & -3.91 \\\\ \hline \text{CXP} & \text{8B SFT teacher} & 21.78 & 22.82 & 16.56 & 41.42 & 31.11 & 28.04 & 28.18 & 27.13 & \text{---} \\\\ \text{CXP} & \text{2B student} & 18.92 & 19.98 & 17.28 & 42.76 & 36.45 & 38.71 & 37.55 & 30.23 & \text{---} \\\\ \text{CXP} & \text{Native OPD} & 17.42 & 18.74 & 8.59 & 35.83 & 31.73 & 24.57 & 27.69 & 23.51 & -6.73 \\\\ \text{CXP} & \text{VITAL} & 17.32 & 18.51 & 8.44 & 35.44 & 31.93 & 26.50 & 28.96 & 23.87 & -6.36 \\\\ \hline \end{array}$$

4. Sensitivity analysis for the temperature parameter (τ\tau)

The results of our sensitivity analysis for the temperature parameter (τ\tau) on MIMIC-CXR are shown below. Performance varies slightly across the tested temperatures, with the best result at τ=0.5\tau = 0.5.

\hline \textbf{$\tau$} & \textbf{R-L } (\uparrow) & \textbf{R-1 } (\uparrow) & \textbf{MET } (\uparrow) & \textbf{RaTE } (\uparrow) & \textbf{P } (\uparrow) & \textbf{R } (\uparrow) & \textbf{F1 } (\uparrow) & \textbf{Avg. } (\uparrow) \\\\ \hline 0.1 & 27.85 & 29.24 & 20.45 & 51.05 & 42.59 & 40.05 & 41.28 & 36.07 \\\\ 0.5 & \textbf{27.94} & \textbf{29.33} & \textbf{20.54} & \textbf{51.14} & \textbf{42.68} & \textbf{40.14} & \textbf{41.37} & \textbf{36.16} \\\\ 1.0 & 27.89 & 29.28 & 20.49 & 51.09 & 42.63 & 40.09 & 41.32 & 36.11 \\\\ \hline \end{array}$$ **5. Scope and Computational Cost** 1)Clarification of Scope We agree with your assessment. We will change the title from "VITAL: Visual and Task Advantage Learning for On-Policy Distillation in Medical Diagnosis" to "VITAL: Visual and Task Advantage Learning for On-Policy Distillation in Medical Report Generation",and synchronously update abstract. 2)Computational Cost We quantified the computational resources required for training VITAL and Native OPD, see table below. Both runs used four NVIDIA A800 GPUs. According to the training logs, VITAL took 504 seconds (8.4 minutes) longer than Native OPD, a 14.6% increase in end-to-end wall-clock time and an additional 0.56 GPU-hours. The additional cost comes primarily from VAP. In Native OPD, the teacher uses the original image to compute log probabilities for the tokens generated by the student. VAP adds one teacher forward pass on the same GPU using a Gaussian-perturbed image. It scores the same tokens and computes the visual weights from the differences in log probabilities between the two passes. We have added this table to session Additional Analysis of VITAL in the appendix. $$\begin{array}{l|ccc} \hline\hline \textbf{Training cost} & \textbf{Native OPD} & \textbf{VITAL} & \textbf{Additional cost} \\\\ \hline \text{Training Time} & \text{3,463 s} & \text{3,967 s} & \text{+504 s (+14.6\\%)} \\\\ \text{GPU-hours} & \text{3.85} & \text{4.41} & \text{+0.56 (+14.6\\%)} \\\\ \hline\hline \end{array}$$ **6.Failure Cases** Thank you for this suggestion. VAP models visual dependence at the token level. Because VITAL optimizes report generation at the text level, it cannot spatially localize lesions. VAP models dependence on visual information at the token level. Because VITAL optimizes report generation at the text level, it cannot spatially localize lesions. TAD uses the teacher’s score advantage over the student to decide when to distill. However, because the rejection decision is made at the rollout level, the student may imitate the teacher’s entire output, including incorrect tokens. In future work, we will explore finer-grained rejection distillation to selectively learn from the teacher’s correct tokens. We have discussed both limitations in the Limitations section and provide qualitative examples of TAD failure cases in the appendix. In future work, we will explore region- and pixel-level lesion localization and finer-grained distillation that selectively learns from correct teacher tokens.

7. Supplementary Evidence Addressing Two Core Challenges

1)Quantitative Comparison with MedOPD

As shown below, VAP shifts sensitivity from non-clinical to clinical tokens: Other drops from 32.01% to 23.64%, while Anatomy and Severity rise from 7.60% to 15.34% and 2.90% to 9.81%, respectively. In contrast, MedOPD increases overall sensitivity but yields lower clinical attention (Anatomy: 11.61%, Severity: 6.52%) and maintains high redundant sensitivity (Other: 24.62%, Visual Sensitivity Rate: 38.13%). VITAL achieves better calibration—driving stronger clinical attention while minimizing redundant visual dependencies (Visual Sensitivity Rate: 34.42%).

\hline\hline \textbf{Metric} & \textbf{Direction} & \textbf{2B Student} & \textbf{MedOPD} & \textbf{VITAL (Ours)} & \textbf{8B teacher} \\\\ \hline \text{other \\%} & \downarrow & 32.01\\% & 24.62\\% & \textbf{23.64\\%} & 21.78\\% \\\\ \text{anatomy \\%} & \uparrow & 7.60\\% & 11.61\\% & \textbf{15.34\\%} & 12.05\\% \\\\ \text{severity \\%} & \uparrow & 2.90\\% & 6.52\\% & \textbf{9.81\\%} & 6.94\\% \\\\ \text{Visual Sensitivity Rate} & \downarrow & 44.73\\% & 38.13\\% & \textbf{34.42\\%} & 35.65\\% \\\\ \hline\hline \end{array}$$ 2)Source of TAD's Advantage We compared our method with the Top-K distillation approach, which distills only the top-2 rollouts ranked by student task performance. The comparison shows that the improvement of our method stems from bridging the teacher-student gap rather than merely filtering out low-quality samples. $$\begin{array}{l|cc|cc} \hline\hline \textbf{Method} & \textbf{MIMIC R-L } (\uparrow) & \textbf{MIMIC F1 } (\uparrow) & \textbf{CheXpert R-L } (\uparrow) & \textbf{CheXpert F1 } (\uparrow) \\\\ \hline \text{VAP + Top-K (k=2)} & \text{27.31} & \text{40.01} & \text{24.69} & \text{38.85} \\\\ \textbf{VAP + TAD (Ours)} & \textbf{27.94 } (+0.63) & \textbf{41.37 } (+1.36) & \textbf{25.03 } (+0.34) & \textbf{40.47 } (+1.62) \\\\ \hline\hline \end{array}$$
Reviewer_bGs1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This work proposes an on-policy distillation framework for medical VLMs that combines visual-sensitivity-based token weighting with additional distillation on rollouts where the teacher achieves a higher task score. Experiments on adult and neonatal chest X-ray datasets report improved out-of-domain performance over existing distillation baselines.

Critical strengths
  1. The paper addresses an important problem in medical VLM distillation: how to allocate teacher supervision according to visual relevance and relative response quality. The observation that teacher-generated responses do not consistently outperform student rollouts provides a useful motivation for selective distillation.

  2. The main comparisons use the same teacher–student backbones and student initialization, reducing confounding from model differences. Evaluation covers both adult and neonatal chest X-ray datasets, with reported out-of-domain F1 gains of 4.60 and 2.79 percentage points over the strongest competing distillation baselines. Substaintial experiments results and OOD experiments also give more understanding on the performance of proposed method.

Critical weaknesses
  1. Limited differentiation from closely related work. VITAL shares substantial similarities with Visual-Advantage On-Policy Distillation for Vision-Language Models (Liu et al., 2026), which already uses changes in teacher token probabilities under different visual conditions to identify visually dependent tokens and redistribute distillation supervision. VAP largely follows this principle, with modifications to the perturbation and weighting schemes. This make the technique contribution seems increment and limited.

  2. The standalone effect of VAP is insufficiently understood. On NeoCXR-EV, VAP alone achieves an F1 of 39.34, substantially below Native OPD’s 46.59—a decrease of 7.25 percentage points. This result challenges the claim that the proposed visual weighting reliably improves distillation. Although the complete framework performs better, the paper should explain why VAP alone causes such degradation and how its interaction with TAD reverses it. Comparisons against the undistilled student do not adequately isolate the benefits of the proposed components over standard OPD.

  3. ROUGE-L is a questionable criterion for deciding when teacher supervision is preferable. ROUGE-L measures textual overlap and may not reliably capture clinically important differences in findings, severity, or negation. The reported correlation with clinical F1 does not establish that ROUGE-L correctly ranks individual teacher–student response pairs, which is precisely what TAD requires. The paper should validate these pairwise decisions and compare ROUGE-L-based selection with clinically grounded alternatives, such as clinical F1 or a combined factuality metric.

Questions for authors

See Weakness above

4 comments

Thank you for your thoughtful and valuable feedback. We hope our responses address your concerns and would be happy to clarify any remaining points.

1.Differences Between VA-OPD and Our Method. We sincerely thank the reviewer for pointing out this highly relevant work. However, we are saddened that you did not recognize the novelty of our work. Our method differs significantly from VA-OPD in the following aspects:

1) Our VAP differs fundamentally from VA-OPD in both its design motivation and effects:

Although our method and VA-OPD use the same approach to estimate visual sensitivity, as acknowledged in the “Visual Advantage Policy Optimization” section (line 211), we respectfully argue that their design philosophies and mechanisms for correcting visual dependencies are fundamentally different, rather than mere modifications to perturbation and weighting schemes:

Motivation and mechanism: VA-OPD is "Group-based": It partitions tokens into high and low visual-advantage groups and performs normalized distillation separately. Its primary goal is to prevent the dilution of distillation signals for visually sensitive tokens, essentially "protecting" them.

Our VAP is "Continuous and Bidirectional": Instead of coarse-grained grouping, VAP applies a continuous, token-wise scaling of the distillation intensity based on visual sensitivity. Crucially, VAP operates bidirectionally: it not only amplifies the distillation weights for highly sensitive tokens (reinforcing reasonable visual reliance) but also actively suppresses the weights for low-sensitivity tokens (correcting inappropriate or hallucinated visual reliance).

Effectiveness of visual sensitivity: The visual sensitivity analysis (see table below) demonstrates that VA-OPD does not improve the sensitivity distribution, whereas our method increases sensitivity to clinically relevant tokens and reduces sensitivity to irrelevant tokens.

\hline\hline \textbf{Metric} & \textbf{Direction} & \textbf{2B Student} & \textbf{VA-OPD} & \textbf{VITAL (Ours)} & \textbf{8B teacher} \\\\ \hline \text{Other \\%} & \downarrow & 32.01\\% & 30.15\\% & \textbf{23.64\\%} & 21.78\\% \\\\ \text{Anatomy \\%} & \uparrow & 7.60\\% & 9.27\\% & \textbf{15.34\\%} & 12.05\\% \\\\ \text{Severity \\%} & \uparrow & 2.90\\% & 4.15\\% & \textbf{9.81\\%} & 6.94\\% \\\\ \hline \text{Visual Sensitivity Rate} & \downarrow & 44.73\\% & 40.85\\% & \textbf{34.42\\%} & 35.65\\% \\\\ \hline\hline \end{array}$$ **2) Regarding the overall methodology:** **Motivation and Methodology:** We are the first to identify misplaced visual reliance and variation in task advantage across rollouts in student models for medical report generation, and we specifically address both problems. Beyond VAP, we also proposed the novel TAD to address the student’s task-specific weaknesses. Through a novel task-based rejection sampling strategy, this module substantially improves distillation performance on out-of-domain datasets and enhances model generalization. We believe this is a valuable contribution. **Experimental results:** The overall experimental results also demonstrate that our method outperforms VA-OPD, as shown in the table below. This is especially evident on the out-of-domain dataset, CheXpert Plus, where VITAL exceeds VA-OPD by 2.11% in Precision (P) and 1.85% in F1 score. $$\begin{array}{llcccccccc} \hline \text{Dataset} & \text{Method} & \text{R-L} \uparrow & \text{R-1} \uparrow & \text{MET} \uparrow & \text{RaTE} \uparrow & \text{P} \uparrow & \text{R} \uparrow & \text{F1} \uparrow & \text{Avg.} \uparrow \\\\ \hline \text{MIMIC-CXR} & \text{VA-OPD} & 0.2867 & 0.3010 & 0.2181 & 0.5051 & 0.4116 & 0.4340 & 0.4225 & 0.3684 \\\\ & \text{VITAL} & 0.2853 & 0.2999 & 0.2223 & 0.5142 & 0.4155 & 0.4330 & 0.4241 & 0.3706 \\\\ \text{CheXpert Plus} & \text{VA-OPD} & 0.2494 & 0.2566 & 0.1832 & 0.4247 & 0.3091 & 0.3963 & 0.3473 & 0.3095 \\\\ & \text{VITAL} & 0.2411 & 0.2504 & 0.1610 & 0.4131 & 0.3255 & 0.4174 & 0.3658 & 0.3106 \\\\ \hline \end{array}$$ **2. Performance of VAP alone.** We attribute the lower F1 relative to Native OPD primarily to VAP’s token-level reweighting. VAP covers all rollouts and attenuates each token's distillation weight according to its visual sensitivity. We acknowledge that VAP alone underperforms Native OPD. Although this strengthens the distillation of visually relevant tokens, it weakens the overall distillation signal. TAD, in contrast, selects rollouts and provides dense additional supervision across all tokens within each selected rollout, compensating for the token-level distillation signal on these rollouts. This explains why combining VAP with TAD reverses the F1 drop observed with VAP alone.

3. Why use ROUGE-L as task proxy?

1). Ranking Agreement. We measured the agreement between ROUGE-L and clinical F1 in ranking teacher–student reports for the same case on MIMIC-CXR and found an agreement rate of 80.3%. This provides stronger evidence supporting the use of the ROUGE-L gap for TAD sample selection. Thank you for your suggestion; we will incorporate this result into the main text at line 250.

2). Comparisons with Clinical F1–Based Proxies. We compared ROUGE-L with clinically grounded alternatives, including clinical F1 and a combined ROUGE-L–F1 metric. As shown in Tables 1 and 2, using ROUGE-L as the task proxy provides denser and more stable distillation signals, yielding better downstream performance.

Table 1. Ablation Study of Task Proxy Settings on MIMIC-CXR and CheXpert Plus

\hline \text{Configuration} & \text{MIMIC ROUGE-L} & \text{MIMIC F1} & \text{CXP ROUGE-L} & \text{CXP F1} \\\\ \hline \text{VAP only} & 0.2774 & 0.3503 & 0.2423 & 0.3616 \\\\ \text{VAP + TAD (CE-F1 gap)} & 0.2564 & 0.3384 & 0.2340 & 0.3265 \\\\ \text{VAP + TAD (ROUGE-L gap)} & 0.2794 & 0.4137 & 0.2503 & 0.4047 \\\\ \text{VAP + TAD (Hybrid 0.5R + 0.5F1)} & 0.2621 & 0.3924 & 0.2487 & 0.3802 \\\\ \hline \end{array}$$ We have added Table 1 to session Additional Analysis of VITAL in the appendix. **3). Further analysis.** Our analysis suggests that the poorer performance of alternative proxies is primarily due to differences in the quality of the distillation signals they provide. On MIMIC-CXR training samples, the teacher's ROUGE-L scores are non-zero in 99.9% of cases, whereas the CE-F1 is only 64.2%, indicating sparse signals. Furthermore, the high variance of CE-F1 leads to unstable distillation. The combination of 0.5 × ROUGE-L + 0.5 × F1 also exhibits higher variance than ROUGE-L alone because individual label flips can cause large changes in F1 scores, resulting in poorer distillation performance (see table 2). Prior studies [1] [2] have demonstrated that the density and stability of distillation signals are closely correlated with distillation effectiveness. **Table 2. Analysis of the Numerical Properties of different task proxy** $$\begin{array}{cccc} \hline \text{Teacher score statistic} & \text{ROUGE-L} & \text{CE-F1} & \text{0.5 ROUGE-L + 0.5 F1} \\\\ \hline \text{Nonzero scores} & 99.9\\% & 64.2\\% & 96.7\\% \\\\ \text{Standard deviation} & 0.11 & 0.35 & 0.24 \\\\ \hline \end{array}$$ [1]Y. Gu, L. Dong, F. Wei, and M. Huang, "MiniLLM: On-policy distillation of large language models," in Proc. Int. Conf. Learn. Represent. (ICLR), 2024. [2]Y. Wen, Z. Li, W. Du, and L. Mou, "f-Divergence minimization for sequence-level knowledge distillation," in Proc. 61st Annu. Meet. Assoc. Comput. Linguistics (ACL), 2023
AuthorsOpenReview ↗

Looking forward to your feedback.

Dear Reviewer,

Thank you for your thoughtful and valuable feedback. In our response, we have clarified the differences between VITAL and VA-OPD, discussed the performance of VAP alone and its complementarity with TAD, and explained our choice of ROUGE-L as the task proxy. We have provided additional comparisons and ablation experiments to support these explanations.

We would greatly appreciate your feedback on whether our responses address your concerns and would be happy to clarify any remaining points.

Best regards,

The Authors

AuthorsOpenReview ↗

Looking forward to your reply.

Dear Reviewer,

We sincerely appreciate your careful evaluation and constructive feedback. As the discussion period is closing in just a few hours, we are writing to follow up on our previous response.

We completely understand that you may be very busy. We want to ensure that our detailed clarifications and additional experiments (regarding the differences between VITAL and VA-OPD, the complementarity of VAP+TAD, and the ROUGE-L proxy) are visible to you. We would greatly appreciate it if you could share your feedback on our response before the final decisions are made. We hope our responses address your concerns, and we would be more than happy to clarify any remaining points.

Thank you again for your valuable time and responsible review. We wish you all the best.

Sincerely,

The Authors

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

This paper studies distilling a large medical vision-language model (VLM) into a smaller one for chest X-ray diagnostic report generation. The authors argue that conventional on-policy distillation often relies on an oversimplified assumption: the student is uniformly weaker than the teacher and therefore should simply imitate it. In practice, however, the student’s deficiencies are more nuanced. First, the student does not necessarily lack visual sensitivity; rather, its visual dependence may be misaligned. For example, it may be highly sensitive to unimportant template words while insufficiently sensitive to clinically meaningful findings. Second, some reports generated by the student are not inferior to those produced by the teacher, so indiscriminate imitation may overwrite useful student behaviors. Based on these observations, the paper proposes VITAL, which consists of two components: Visual Advantage Policy Optimization (VAP), which estimates the teacher’s token-level visual dependence through image perturbation and reweights distillation accordingly; and Task Advantage Distillation (TAD), which selectively distills only those student rollouts where the teacher exhibits a positive task advantage. Experiments on adult and neonatal chest X-ray datasets demonstrate that VITAL improves out-of-domain generalization, particularly in clinical F1 scores.

Critical strengths
  1. The paper has a clear problem formulation. Existing on-policy distillation methods usually assume that the teacher is uniformly stronger across samples, and that the student simply needs stronger supervision. However, the authors do not directly follow this assumption. And the paper further analyze what the student model actually lacks. The two issues identified in the paper are both intuitive and closely related to medical report generation. VAP and TAD correspond to two complementary questions: where should the student rely more on visual evidence, and when should it imitate the teacher? This decomposition is easy to understand and aligns naturally with the two problems observed in the paper.
  2. The ablation study is reasonably comprehensive. The authors remove VAP and TAD separately and jointly, provide mechanism analyses for both components, and examine key hyperparameters such as the TAD loss weight α and the Gaussian noise strength σ_noise.
Critical weaknesses
  1. The current experiments rely on a single teacher–student configuration, Qwen3-VL-8B and Qwen3-VL-2B. Although this setup may be sufficient to demonstrate the basic effectiveness of the proposed method, it limits the generalizability of the conclusions. In particular, it remains unclear whether VITAL can generalize to stronger general-purpose multimodal models or medically specialized MLLMs such as HuluMed and Lingshu. Given that the method targets medical VLMs, evaluating only one general-purpose MLLM backbone is not sufficient. I recommend that the authors validate VITAL on at least one different teacher–student configuration to strengthen the evidence for its generality.
  2. Recent work has proposed several multimodal on-policy distillation methods, such as VOLD, CVPD, and Vision-OPD. Given that this paper targets on-policy distillation for medical multimodal models, comparing with only three baselines is not sufficient to support the advantages. It would be better to add 2-3 baselines.
  3. VAP estimates visual dependence using global Gaussian noise, but such perturbations may only reflect the model’s sensitivity to image quality changes but not meaningful lesion evidence. The authors should give additional validation to demonstrate that the estimated visual dependence is clinically reliable.
Questions for authors
  1. In TAD, why did the authors choose ROUGE-L as the task advantage signal instead of clinical F1, RadGraph-style metrics, or a learned clinical reward model?

  2. It is possible that the teacher is not stronger than the student. However, another line of OPD methods provides the teacher with privileged information to make the teacher as strong as possible. The authors’ perspective is interesting. Could they analyze the relationship between these two paths, ideally with some empirical discussion?

3 comments

Thank you for your positive assessment of our work and for your constructive feedback. We address your two questions below and have added comparisons with the two baselines you suggested, along with additional distillation experiments using other model architectures.

1. Why did we choose ROUGE-L as the task advantage in TAD?

We chose ROUGE-L as the task proxy for radiology report generation based on 4 considerations.

1). Task coverage. Our task is radiology report generation. ROUGE-L measures textual overlap between generated and reference reports across both imaging findings and diagnostic impressions. Across the teacher and student models evaluated on NeoCXR, ROUGE-L and CE-F1 were positively correlated (Pearson r=0.751r = 0.751). This correlation and ROUGE-L’s coverage of the full report text support its use as a proxy for report-generation performance. CE-F1 measures agreement on extracted clinical labels. Two reports with identical CE-F1 scores can therefore differ in how closely their descriptions match the reference. This makes CE-F1 insufficient as a standalone proxy for overall report-generation performance.

2). Quality of the distillation signal. On 3,000 MIMIC-CXR training samples, the teacher's ROUGE-L scores are non-zero in 99.9% of cases, whereas the CE-F1 is only 64.2%, indicating sparse signals. Furthermore, the high variance of CE-F1 leads to unstable distillation. Prior studies [1] [2] have demonstrated that the density and stability of distillation signals are closely correlated with distillation effectiveness.

Table 1. Analysis of the Numerical Properties of different task proxy

\hline \text{Teacher score statistic} & \text{ROUGE-L} & \text{CE-F1} \\\\ \hline \text{Nonzero scores} & 99.9\\% & 64.2\\% \\\\ \text{Standard deviation} & 0.11 & 0.35 \\\\ \hline \end{array}$$ [1]Y. Gu, L. Dong, F. Wei, and M. Huang, "MiniLLM: On-policy distillation of large language models," in Proc. Int. Conf. Learn. Represent. (ICLR), 2024. [2]Y. Wen, Z. Li, W. Du, and L. Mou, "f-Divergence minimization for sequence-level knowledge distillation," in Proc. 61st Annu. Meet. Assoc. Comput. Linguistics (ACL), 2023 **3). Distillation performance.** We compared VAP alone with TAD using three task proxy. The ROUGE-L achieved the highest scores. **Table 2. Ablation Study of Task Proxy Settings on MIMIC-CXR and CheXpert Plus** $$\begin{array}{cccc} \hline \text{Configuration} & \text{MIMIC ROUGE-L} & \text{MIMIC F1} & \text{CXP ROUGE-L} & \text{CXP F1} \\\\ \hline \text{VAP only} & 0.2774 & 0.3503 & 0.2423 & 0.3616 \\\\ \text{VAP + TAD (CE-F1 gap)} & 0.2564 & 0.3384 & 0.2340 & 0.3265 \\\\ \text{\textbf{VAP + TAD (ROUGE-L gap)}} & \mathbf{0.2794} & \mathbf{0.4137} & \mathbf{0.2503} & \mathbf{0.4047} \\\\ \hline \end{array}$$ We have added Table 2 to session Additional Analysis of VITAL in the appendix. **4). Efficiency**. Compared with RadGraph-style metrics and learned clinical reward models, ROUGE-L requires no additional model inference for repeated rollout scoring. We chose this lightweight metric because TAD’s main contribution is to distill from the teacher only when it outperforms the student on a given rollout.

2.VITAL vs. Privileged-Teacher Distillation. Thank you for this insightful comment. We agree that there are two main paths in OPD: one focuses on information compensation (e.g., using privileged information to make the teacher as strong as possible), and the other focuses on aligning the distillation conditions. Our VITAL follows the latter path.

Why might a stronger teacher lead to a weaker student? As discussed in prior works [1, 2], the information gap between the teacher and the student is a critical factor. In MedOPD, the teacher utilizes privileged information (e.g., reference reports) during distillation, while the student only has access to images. Because the teacher is guided by text reports, it may assign high probabilities to tokens that are extremely difficult to infer from images alone. Consequently, the student struggles to learn the true mapping from visual evidence to text.

Empirical analysis: Interestingly, despite MedOPD's teacher achieving a 93% win rate against the student, the distilled student performs worse than our VITAL (see Table 1 in main text or Table below). We observe that on the CheXpert Plus dataset, the student model of MedOPD suffers from significantly more severe repetitive generation issues compared to VITAL. This strongly suggests that the student is struggling to ground its generation in visual evidence, further validating our perspective. VITAL avoids this mismatch by ensuring the distillation supervision perfectly matches the student's image-only inference condition, thereby achieving better results.

\hline \textbf{Dataset} & \textbf{Method} & \textbf{R-L} & \textbf{R-1} & \textbf{MET} & \textbf{RaTE} & \textbf{P} & \textbf{R} & \textbf{F1} & \textbf{Avg.} \\\\ \hline \text{MIMIC-CXR} & \text{MedOPD} & 24.39 & 25.67 & 18.34 & 45.91 & 37.16 & 31.39 & 34.03 & 30.98 \\\\ \text{(in-domain)} & \textbf{VITAL (Ours)} & \mathbf{27.94} & \mathbf{29.33} & \mathbf{20.54} & \mathbf{51.14} & \mathbf{42.68} & \mathbf{40.14} & \mathbf{41.37} & \mathbf{36.16} \\\\ \hline \text{CheXpert Plus} & \text{MedOPD} & 21.73 & 22.52 & 13.91 & 38.43 & 33.89 & 34.99 & 34.43 & 28.56 \\\\ \text{(out-of-domain)} & \textbf{VITAL (Ours)} & \mathbf{25.03} & \mathbf{25.88} & \mathbf{17.02} & \mathbf{43.26} & \mathbf{38.27} & \mathbf{42.93} & \mathbf{40.47} & \mathbf{33.27} \\\\ \hline \end{array}$$ [1] D. Lopez-Paz, L. Bottou, B. Schölkopf, and V. Vapnik, "Unifying distillation and privileged information," in International Conference on Learning Representations (ICLR), 2016. [2] J. H. Cho and B. Hariharan, "On the efficacy of knowledge distillation," in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019.

We have also supplemented the experimental results for the two relevant OPD methods you suggested: VOLD and CVPD. Regarding distillation across other model architectures, we are currently conducting experiments on distilling Lingshu to Qwen2.5-VL-3B. We will update these results here and include them in the final version of the paper.

\hline\hline \textbf{Dataset} & \textbf{Method} & \textbf{R-L} & \textbf{R-1} & \textbf{MET} & \textbf{RaTE} & \textbf{P} & \textbf{R} & \textbf{F1} & \textbf{Avg.} \\\\ \hline \text{MIMIC-CXR} & \text{VOLD} & 26.47 & 27.74 & 19.56 & 48.46 & 40.39 & 37.76 & 39.03 & 34.20 \\\\ & \text{CVPD} & 27.80 & 29.20 & 20.50 & 50.10 & 42.20 & 37.50 & 39.71 & 35.29 \\\\ \hline \text{CheXpert Plus} & \text{VOLD} & 23.42 & 24.29 & 16.32 & 41.29 & 35.46 & 38.28 & 36.81 & 30.84 \\\\ & \text{CVPD} & 24.60 & 25.50 & 17.10 & 42.90 & 34.70 & 38.80 & 36.64 & 31.46 \\\\ \hline\hline \text{NeoCXR} & \text{VOLD} & 45.10 & 46.50 & 45.20 & 59.80 & 53.00 & 55.20 & 54.08 & 51.34 \\\\ & \text{CVPD} & 45.60 & 46.20 & 45.20 & 60.10 & 54.00 & 55.00 & 54.50 & 51.51 \\\\ \hline \text{NeoCXR-EV} & \text{VOLD} & 30.50 & 31.40 & 30.40 & 49.90 & 43.50 & 40.50 & 41.95 & 38.31 \\\\ & \text{CVPD} & 31.20 & 32.30 & 31.00 & 50.20 & 48.20 & 45.30 & 46.71 & 40.70 \\\\ \hline\hline \end{array}$$
Reviewer_RibbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper targets on-policy distillation (OPD) of a large medical VLM into a compact student for chest X-ray report generation. It argues that prior OPD treats teacher supervision as uniformly beneficial, and identifies two student-side deficits: a visual deficit (the student’s sensitivity to image perturbation is misallocated toward templated tokens rather than clinical findings) and a task deficit (>20% of student rollouts already match or beat the teacher). VITAL addresses these with VAP, which reweights token-level updates by the teacher’s perturbation-measured visual dependence, and TAD, which filters out rollouts where the teacher has no ROUGE-L advantage. Reported gains are marginal in-domain and larger OOD (+4.60 F1 on CheXpert Plus, +2.79 on NeoCXR-EV).

Critical strengths

The decomposition is clean: Native OPD averages uniformly over both tokens and rollouts, and the two components break exactly these two assumptions.

The task-deficit evidence is convincing. Fig. 5(c) shows Native OPD degrading specifically on student-better samples while VITAL does not a direct test of the motivation. Table 1 reinforces it: all three baselines fall below the undistilled student on CheXpert Plus F1 (35.87 / 35.03 / 34.43 vs 37.55) while VITAL reaches 40.47, above even the teacher’s 38.51.

Critical weaknesses
  1. The visual-deficit comparison is confounded by capacity and post-training. Fig. 1(a) contrasts a GSPO-trained 8B teacher with a 2B student that received only SFT. The student’s higher variation under perturbation is more simply explained by it being a smaller, less stable model, one explanation covering all three categories: templated phrasing has not consolidated (higher “Other”), and findings are not grounded (lower Anatomy/Severity).

  2. VAP and TAD conflict on the behavior the paper claims to protect. TAD exists to avoid pulling student-better rollouts toward the teacher, but L_VAP sums over all n rollouts with no I⁺ condition (Eq. 8; Algorithm 1 lines 9–11). On a rollout TAD correctly rejects, VAP still applies a teacher-log-ratio advantage pushing toward the teacher. “Preserves useful student behaviors” holds for only one of the two terms. Please share the filtering or justify the exception, with an ablation applying I⁺ to VAP.

  3. The in-domain margins are +0.09 on MIMIC-CXR and +0.11 on NeoCXR, and on NeoCXR the baselines win on RaTE, precision, recall and F1, all four clinical metrics, with VITAL’s lead coming entirely from n-gram overlap. “Outperforms prior OPD methods in-domain” should be softened to parity. No seeds, standard deviations or significance tests are reported anywhere, and CheXpert Plus has only 202 examples, so the headline +4.60 F1 needs an interval. The α sweep on NeoCXR-EV (40.26 / 39.71 / 42.90 / 39.81 for α = 0.5/1/2/4) makes this concrete: a 3-point isolated spike landing exactly on the chosen value is as consistent with run variance as with a real optimum, and it exceeds several of the reported improvements.

Questions for authors
  1. In Fig. 1(a), the per-category bars sum to 40.7% (teacher) and 42.5% (student), but the annotated overall rates are 35.7% and 44.7%. What are the two denominators, and how is “visual sensitivity rate” defined exactly?

  2. What does L_VAP reduce to when w ≡ 1? If it is not exactly Native OPD, then the comparison in Table 1 confounds the visual reweighting with the change from a KL objective to a clipped policy-gradient one, and a w ≡ 1 row would be needed to separate them.

  3. TAD is judged by ROUGE-L, but the teacher was GSPO-trained on a ROUGE-L reward, so the gate’s criterion favors the teacher by construction. ROUGE-L is also negation-insensitive. Why not use clinical F1 or RaTE, both already computed? And what is the acceptance rate |I⁺|/n over training?

  4. What are the numerator and denominator of “visual sensitivity rate” in Fig. 1(a), and is it normalized per category? student: 32+7.6+2.9 < 44.7. Teacher 21.8+12+6.9 > 35.7

5 comments

Thank you for your careful review and constructive feedback. Your comments helped us substantially improve the manuscript. We address your four questions below. Because Questions 1 and 4 are closely related, we address them together before turning to Questions 2 and 3. We hope our responses address your concerns and would be happy to clarify any remaining points.

1. Definition and denominators of visual sensitivity in Figure 1

The dashed lines indicate the overall visual sensitivity rate, defined as the number of tokens edited following image perturbation divided by the total number of tokens. By this measure, the student model exhibited greater overall visual sensitivity than the teacher model. To further examine how the student model’s visual sensitivity is distributed across token categories, the bars show the percentage of edited tokens that fall into each of the three displayed categories. The denominator for the bars is therefore the total number of tokens edited following image perturbation.

4. Relationship between the dashed lines and bar values in Figure 1

Because the dashed lines and bars use different denominators, the sum of the three displayed percentages need not equal the overall visual sensitivity rate. The remaining visually sensitive edited tokens fell predominantly into the disease-name category, which accounted for 57.5% of edited tokens in the student model and 59.3% in the teacher model, a difference of 1.8 percentage points. This finding is consistent with prior work on knowledge distillation in medical vision–language models [Ref. 1]. The student and teacher models already show similar sensitivity to high-frequency entities such as disease names, while substantial gaps remain in sensitivity to subtle clinical cues and in complex reasoning.

We thus simplified Figure 1 to emphasize the category-level differences most relevant to our research question. We apologize for the confusion and thank you for pointing this out. We will revise the y-axis label of the bar chart to “Distribution of Sensitivity (%)” and explain the rationale for displaying only these three categories in the appendix.

[Ref. 1] H. Tsaniya, C. Fatichah, N. Suciati, T. Obi, and J.-S. Lee, "Medical report generation with knowledge distillation and multi-stage hierarchical attention in vision transformer encoder and GPT-2 decoder," IEEE Access, vol. 13, pp. 132973–132989, 2025

2. What does LVAP\mathcal{L}_{\mathrm{VAP}} reduce to when w≡1w \equiv 1?

Thank you for raising this question. When w≡1w \equiv 1, LVAP\mathcal{L}_{\mathrm{VAP}} reduces to the native OPD objective.

3. (1) Why use ROUGE-L to quantify the teacher–student gap instead of RaTE, F1, or a combination of metrics?

Why not RaTE? RaTE relies on medical entity extraction. Entity-level agreement alone is insufficient to establish agreement in imaging findings or disease diagnoses. Its applicability also varies across clinical domains, and it is unsuitable for the neonatal conditions considered in our study.

Why not F1 and its combination with ROUGE-L? We compared three metrics for estimating task advantage: ROUGE-L, CE-F1, and an equally weighted combination of ROUGE-L and F1 (0.5×ROUGE-L+0.5×F10.5 \times \mathrm{ROUGE\text{-}L} + 0.5 \times \mathrm{F1}). ROUGE-L yielded the best downstream performance (Table 2). CE-F1 showed high variance, yielded fewer samples eligible for distillation, and led to unstable training (Table 1) and poorer downstream performance (Table 2). Because F1 is computed from discrete labels, a change in a single label can shift the score by approximately 0.25–0.50. Such fluctuations make the decision to distill a sample sensitive to individual label changes and hinder convergence. The equally weighted combination also underperformed ROUGE-L in our additional experiments (Table 2).

Table 1. Analysis of the Numerical Properties of Teacher Scores Across Three Task Proxies

\hline \text{Teacher score statistic} & \text{ROUGE-L} & \text{CE-F1} & \text{0.5 ROUGE-L + 0.5 F1} \\\\ \hline \text{Nonzero scores} & 99.9\\% & 64.2\\% & 96.7\\% \\\\ \text{Standard deviation} & 0.11 & 0.35 & 0.24 \\\\ \hline \end{array}$$ **Table 2. Ablation Study of Task Proxy Settings on MIMIC-CXR and CheXpert Plus** $$\begin{array}{cccc} \hline \text{Configuration} & \text{MIMIC ROUGE-L} & \text{MIMIC F1} & \text{CXP ROUGE-L} & \text{CXP F1} \\\\ \hline \text{VAP only} & 0.2774 & 0.3503 & 0.2423 & 0.3616 \\\\ \text{VAP + TAD (CE-F1 gap)} & 0.2564 & 0.3384 & 0.2340 & 0.3265 \\\\ \text{VAP + TAD (ROUGE-L gap)} & 0.2794 & 0.4137 & 0.2503 & 0.4047 \\\\ \text{VAP + TAD (Hybrid 0.5R + 0.5F1)} & 0.2621 & 0.3924 & 0.2487 & 0.3802 \\\\ \hline \end{array}$$ We have added Table 2 to session Additional Analysis of VITAL in the appendix. **3. (2) Proportion of training samples selected by TAD** Consistent with Figure 1, TAD selected 76% of the training samples for distillation.
AuthorsOpenReview ↗

Looking forward to your feedback.

Dear Reviewer,

Thank you for your thoughtful comments. In our response, we have clarified the definition and denominators of visual sensitivity in Figure 1, the VAP objective when w=1w=1, the rationale for using ROUGE-L, and the proportion of training samples selected by TAD. We have also provided additional experiments and outlined revisions to improve the clarity of Figure 1.

As the discussion period draws to a close, we would like to follow up on our response and ask whether you have any remaining questions or concerns. We would be happy to provide further clarification.

Best regards,

The Authors

Dear Authors,

Thanks for your reply, some of my concerns have been addressed. The remaining ones in the weaknesses section have not been addressed yet. I will keep my previous score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.