acceptodds
Under review as a conference paper at ICLR 2027

EM-OPD: Evidence-Conditioned Multi-Expert On-Policy Distillation for Pathology Report Generation

Abstract

Multimodal large language models often fail to capture less salient visual evidence required for fine-grained reporting. In contrast, a comprehensive pathology report must integrate heterogeneous evidence from diagnostic patterns, abnormalities, glandular architecture, and cellular morphology. To address this challenge, we propose Evidence-Conditioned Multi-Expert On-Policy Distillation (EM-OPD), a framework that transforms heterogeneous pathological evidence into a unified report generation capability. EM-OPD constructs multiple evidence-conditioned experts equipped with spatial anchors and morphological descriptors. These experts provide complementary perspectives, enabling token-level supervision of student-generated trajectories. By adaptively integrating expert knowledge through balanced joint consolidation, EM-OPD learns to generate pathology reports with improved evidence coverage and diagnostic reliability.

Primary area:
reinforcement learning
Keywords:
On-policy Distillation, Evidence-conditioned Reasoning, Computational Pathology

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_KLunOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EM-OPD, an evidence-conditioned multi-expert on-policy distillation framework for pathology report generation. The method constructs multiple teacher experts from a shared frozen backbone, where each expert is conditioned on a different type of privileged pathological evidence (diagnosis, abnormality, glandular structure, or cellular morphology). The student generates on-policy trajectories, while the corresponding expert provides token-level supervision through Jensen-Shannon divergence. The expert objectives are then equally averaged through a balanced joint-consolidation objective. At inference time, the student uses only the pathology image and reporting prompt.

The paper evaluates EM-OPD on five pathology datasets and compares against general/pathology VLMs, SFT, and GRPO. The authors report substantial improvements, particularly for fine-grained cellular and abnormality-level evaluation, as well as additional out-of-domain results.

Critical strengths
  1. Practical use of heterogeneous pathology supervision.
    The paper provides a reasonable way to unify heterogeneous annotations such as regions, segmentation masks, diagnosis labels, and morphological information into a single report-generation student. The spatial-anchor + morphological-descriptor formulation is intuitive and practically relevant.

  2. Comprehensive empirical study.
    The paper evaluates multiple pathology evidence types, several student/teacher sizes, SFT and GRPO baselines, cue ablations, multi-expert ablations, divergence choices, and out-of-domain performance.

  3. No additional inference-time expert cost.
    Privileged evidence and expert teachers are used only during training; the final student retains a standard image-to-report inference interface.

Critical weaknesses
  1. Limited methodological novelty relative to prior OPD work.
    The core training formulation closely follows existing on-policy distillation: the student generates its own trajectories and a teacher provides token-level distribution supervision on those prefixes [1]. Privileged-information OPD has also already been extended to multimodal models, e.g., Vision-OPD [2], and on-policy distillation has recently been explored for medical VLM reasoning in OPVD [3]. The main novelty here therefore appears to be the pathology-specific construction of multiple evidence-conditioned experts and equal-weight consolidation, rather than a fundamentally new OPD framework. The Methods section should also more clearly distinguish inherited OPD components from newly proposed components.

  2. Missing comparison against prior medical/multimodal OPD methods.
    The main training baselines are SFT and GRPO, but the paper does not compare against existing on-policy distillation methods for multimodal or medical reasoning [2,3]. Since the claimed contribution is specifically an extension of OPD to pathology with multiple experts, an OPD baseline is important to determine whether the gains come from on-policy distillation itself or from the proposed multi-expert design.

  3. Report-generation quality is not directly evaluated.
    The authors state that no ground-truth pathology reports are available. Generated reports are therefore passed through Qwen3.5-27B to extract structured information, after which classification/F1/Mention Rate metrics are computed. This evaluates whether certain structured findings can be recovered from the report, but does not directly assess clinical correctness, completeness, coherence, hallucination, or overall report quality. If the target task is pathology report generation, I would expect either blinded pathologist evaluation or a validated clinical report-quality metric (e.g., a pathology-adapted GREEN-style evaluation). Otherwise, it is unclear why the task is formulated as open-ended report generation when both supervision and evaluation ultimately reduce to structured pathology annotations.

  4. The GRPO baseline is insufficiently specified.
    GRPO is introduced as the RL baseline, but the paper does not clearly define the reward function used for pathology report generation or provide the corresponding training objective. Since GRPO performance is highly sensitive to reward design, this makes the comparison difficult to assess or reproduce. If the reward is adopted from prior work, the source should be cited; if it is newly designed, the reward formulation should be explicitly defined.

  5. Several design choices need stronger justification.
    EM-OPD fixes the Jensen-Shannon mixture coefficient to 1/2. However, the original OPD/GKD work shows that divergence choice and JSD interpolation can be task-dependent and explicitly studies different values of the coefficient [1]. The appendix compares JSD with forward/reverse KL, but does not study sensitivity to this coefficient. In addition, the balanced consolidation strategy is simply an equal average of expert objectives; stronger comparison with adaptive multi-task/multi-expert weighting would help establish whether equal weighting is important.

References

[1] R. Agarwal et al., “On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes,” ICLR 2024.
https://proceedings.iclr.cc/paper_files/paper/2024/hash/5be69a584901a26c521c2b51e40a4c20-Abstract-Conference.html

[2] Q. Yuan et al., “Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation,” 2026.
https://arxiv.org/abs/2605.18740

[3] N. Zhang and Q. Yang, “OPVD: On-Policy Chain-of-Visual-Thought Distillation for Medical Reasoning,” MICCAI 2026.
https://papers.miccai.org/miccai-2026/0740-Paper3650.html

Questions for authors
  1. What is the methodological novelty beyond standard OPD/GKD, privileged multimodal OPD, and recent medical VLM on-policy distillation?
  2. Why are prior multimodal/medical OPD methods not included as baselines?
  3. What exact reward and optimization objective are used for the GRPO baseline?
  4. Given that no ground-truth reports exist, why is open-ended report generation chosen as the target task? Could the authors provide blinded pathologist evaluation or a validated clinical report-quality metric?
  5. How sensitive is performance to the JSD interpolation coefficient rather than fixing it to 1/2?
  6. How does balanced equal weighting compare with adaptive expert weighting or other multi-objective optimization strategies?
2 comments

We sincerely appreciate the time and effort you dedicated to reviewing our work. Your comments have helped us improve the clarity of our work.


W1&Q1: Limited methodological novelty relative to prior OPD work.

We agree that token-level supervision is an established OPD component. However, the central contribution of EM-OPD is an evidence-conditioned multi-expert framework that turns fragmented pathology annotations into supervision for reports covering multiple levels of pathological evidence. Our novelty therefore lies in the combination of multi-expert supervision, the use of otherwise fragmented annotations, and their consolidation toward multi-level report coverage.

W2&Q2: Missing comparison against prior OPD methods.

We add a stronger OPD baseline, Vision-OPD, to further demonstrate the effectiveness of our method. For OPSD, the official code is not open-sourced.

MethodBreakHisGleasonXAIGlaSLizardTCGA
Qwen3.5-4B17.7223.4768.7536.9320.74
Vision-OPD-4B22.9538.1070.0037.7840.53
Ours-4B36.3748.9091.2579.3343.07
Qwen3.5-9B29.1642.1367.5040.1640.97
Vision-OPD-9B44.9755.0190.0042.9749.89
Ours-9B57.9861.2190.0085.1365.56

As the table shows, EM-OPD outperforms Vision-OPD, which further demonstrates the effectiveness of our method.

W3&Q4: Report-generation quality is not directly evaluated.

We add a blinded pathologist review of 20 sampled cases each from BreakHis and Lizard and three-judge LLM evaluation across all five datasets (Sec. 4.3, Table 2). GPT-6.1 Sol, Gemini 3.8 Flash, and Claude Sonnet 5.5 independently score image–report pairs from 1–5. Both evaluations consider correctness, completeness, coherence, hallucinations, contradictions, and omissions.

MethodExpert (BreakHis)Expert (Lizard)LLM (BreakHis)LLM (GleasonXAI)LLM (GlaS)LLM (Lizard)LLM (TCGA)
Qwen3.5-9B2.402.502.562.662.432.422.54
Qwen3.5-27B2.702.702.712.862.782.672.62
PathGen-LLaVA2.202.252.362.372.242.432.49
Patho-R1-7B1.852.151.992.111.921.862.50
Gemini 3.1 Pro3.102.953.053.473.043.133.23
GPT-5.53.553.453.363.823.523.313.39
Qwen3.5-9B (SFT)2.552.452.622.632.452.382.55
Qwen3.5-9B (GRPO)2.702.552.732.682.632.582.61
Vision-OPD-9B2.202.052.382.191.822.252.21
Ours-9B (OPSD)2.853.003.082.992.802.942.90
Ours-9B (Teacher 27B)3.153.103.303.263.023.133.59

As the table shows, our method outperforms most baselines in human evaluation. LLM evaluation likewise favors EM-OPD over the same-scale baselines. We further compute Spearman's rank correlation coefficient (ρ\rho) using LLM scores on the same sampled image–report pairs evaluated by the experts. Table 7 in Appendix A.5 shows strong model-level rank correlation between expert and LLM scores on BreakHis (ρ=0.939\rho=0.939) and Lizard (ρ=0.926\rho=0.926). LLM scores also correlate positively with automatic metrics.

W4&Q3: The GRPO baseline is insufficiently specified.

During GRPO training, we use the same extractor to convert the generated reports into structured findings, which are compared with the corresponding annotations to compute the GRPO reward. Eqs. (16)–(19) now specify the GRPO reward pipeline and objective.

W5&Q5&Q6: Several design choices need stronger justification, such as Jensen-Shannon mixture coefficient or adaptive multi-task/multi-expert weighting.

We vary the mixture weight β\beta under matched training settings (Appendix A.4, Table 6) and extend the original KL ablation.

SettingBreakHisGleasonXAIGlaSLizardTCGA
Forward KL42.9565.4481.2574.4062.15
β=0.25\beta=0.2548.9961.1285.0074.5864.18
β=0.50\beta=0.5057.9861.2190.0085.1365.56
β=0.75\beta=0.7551.0164.6781.2579.5466.01
Reverse KL53.5664.5176.2575.5959.24

Symmetric JSD achieves relatively better results, supporting it as a default optimal choice.

We also add GradNorm adaptive weighting (Sec. 4.4.2, Fig. 3).

Method / WeightingBreakHisGleasonXAIGlaSLizardTCGA
Qwen3.5-9B29.1642.1367.5040.1640.97
Original source distribution55.7163.6176.2542.5860.42
Adaptive (GradNorm)54.4950.3383.7567.2248.08
Balanced (Ours)57.9861.2190.0085.1365.56

Our method exceeds GradNorm on all five datasets. Compared with source-proportional weighting, it improves four datasets with a trade-off on GleasonXAI. This supports balancing reporting emphasis independently of annotation volume.

Thank you for the authors' rebuttal. The feel like the experiment of β\beta values shows a mixed result on how the method generalizes. However, the biggest concern that I still have is the generalization of the EM-OPD algorithm that the authors provide. Authors frame their story as a novel and generalized algorithm but only limited to the pathology domain. If the algorithm is tested on more models and different domains (e.g, document report generation also requires multi-level hops and reasoning layers), it will be more convincing. If this work is submitted to MICCAI or any specialized conference that accepts novel pathology method, I would give a clear accept. However, at ICLR conference level, I think that the method should be tested on more domains and models to see the generalizability of the algorithm. Therefore, I still keep my score.

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

This paper proposes evidence-conditioned on-policy distillation for pathology report generation. A shared frozen teacher is conditioned on source-specific spatial anchors and morphological descriptors and provides token-level supervision on student-generated trajectories. Training balances several annotation sources. The approach addresses a meaningful supervision problem and yields substantial gains on selected pathology benchmarks. However, the evaluation does not establish the reliability of complete reports, and the contribution of the proposed expert organization is not isolated from additional supervision and balanced multi-task training.

Critical strengths
  1. The method makes practical use of existing classification, segmentation, and local abnormality annotations when complete report labels are unavailable.
  2. The paper investigates spatial anchors, descriptors, sequential versus joint training, source balancing, and divergence choice.
  3. The student does not require privileged annotations or teacher calls at inference time.
Critical weaknesses
  1. Section 4.1 extracts labels from generated reports using an LLM and scores selected attributes. These metrics do not measure unsupported additional claims, contradictions, inappropriate certainty, or important omissions. Mention Rate can increase simply because a report mentions more categories. A blinded expert or claim-level assessment, with report length controlled, is needed to support the central report-quality claim.
  2. All experts share frozen parameters, and each instance is supervised by the expert associated with its source. Balanced consolidation uses fixed equal weights and stratified sampling. A unified conditioned teacher supplied with the same cues, examples, and compute is the most important missing baseline. Existing single-expert comparisons also change supervision coverage.
  3. Equation 8 samples trajectories from the parameter-dependent student policy. Differentiating its expectation includes a score-function term for the trajectory distribution. If sampled trajectories are treated as fixed during backpropagation, Equation 11 describes a surrogate or semi-gradient rather than an unbiased gradient of the stated objective. The authors should align the mathematical claim with the actual implementation.
  4. The SFT baseline uses successful Gemini-2.5-Pro generations, while the proposed method uses privileged Qwen teacher supervision. Supervision access, sample counts, generation budgets, and compute should be matched or disclosed. Qwen3.5-27B is both a teacher and a report-label extractor; independent validation of the extractor is needed, particularly for negation and uncertainty.
  5. In Table 1, the 4B model with a 27B teacher achieves 43.07 TCGA accuracy versus 57.51 for 4B GRPO. In Table 3, the proposed model obtains 46.67 ORCHID balanced accuracy versus 47.01 for the original 9B model and 54.86 for GRPO. Claims of consistent superiority should be qualified, and uncertainty across random seeds should be reported.
Questions for authors
  1. How does EM-OPD compare with a unified conditioned teacher under identical supervision, sampling, and compute budgets?
  2. Does independent expert assessment confirm fewer unsupported statements, rather than only better extraction of benchmark labels?
  3. Does training differentiate through the student trajectory distribution? If not, what precise surrogate objective is optimized?
  4. How do four expert categories map to five datasets, and how do the reported resampling repetitions implement equal expert weighting?
  5. Are SFT, GRPO, and OPD matched in their access to privileged annotations and training resources?
3 comments

We sincerely appreciate the time and effort you dedicated to reviewing our work. We have taken your comments seriously and below are our point-by-point response.


W1&Q2: The metrics do not measure unsupported additional claims, contradictions, inappropriate certainty, or important omissions. Mention Rate can increase simply because a report mentions more categories. A blinded expert or claim-level assessment, with report length controlled, is needed to support the central report-quality claim.

We examine category enumeration on TCGA. As the table shows, our method mentions fewer categories, while achieving higher accuracy and mention rate. Thus, its advantage in accuracy and mention rate over other baselines is less possible to be interpreted as solely broader category enumeration.

MethodMean countAccMR
Qwen3.5-9B3.3840.9770.19
Qwen3.5-27B3.1952.6479.11
PathGen-LLaVA1.8332.4969.33
Patho-R1-7B1.3645.3988.83
Gemini-3.1-Pro4.1162.8789.70
GPT-5.54.0056.2796.23
Qwen3.5-9B (SFT)3.6842.8684.05
Qwen3.5-9B (GRPO)4.6561.2795.72
Vision-OPD-9B3.3149.8965.26
Ours-9B (OPSD)3.1748.8874.55
Ours-9B (Teacher 27B)3.4365.5691.01

We also add a blinded pathologist review of a subset from BreakHis and Lizard (each source 20), and a three-judge LLM evaluation across all five datasets. The evaluation scores image–report pairs from 1–5, regarding correctness, completeness, coherence, hallucinations, contradictions, and omissions. Considering the influence of the report length, we explicit a rule in the criteria which instructs the LLM to prioritize factual reliability over meaningless lengthy.

MethodExpert (BreakHis)Expert (Lizard)LLM (BreakHis)LLM (GleasonXAI)LLM (GlaS)LLM (Lizard)LLM (TCGA)
Qwen3.5-9B2.402.502.562.662.432.422.54
Qwen3.5-27B2.702.702.712.862.782.672.62
PathGen-LLaVA2.202.252.362.372.242.432.49
Patho-R1-7B1.852.151.992.111.921.862.50
Gemini 3.1 Pro3.102.953.053.473.043.133.23
GPT-5.53.553.453.363.823.523.313.39
Qwen3.5-9B (SFT)2.552.452.622.632.452.382.55
Qwen3.5-9B (GRPO)2.702.552.732.682.632.582.61
Vision-OPD-9B2.202.052.382.191.822.252.21
Ours-9B (OPSD)2.853.003.082.992.802.942.90
Ours-9B (Teacher 27B)3.153.103.303.263.023.133.59

As shown in the table, our method outperforms most baselines in human evaluation. LLM evaluation likewise favors EM-OPD over the same-scale baselines. We further compute Spearman's rank correlation coefficient (ρ\rho) using LLM scores on the same sampled image–report pairs evaluated by the experts. Table 7 in Appendix A.5 shows strong model-level rank correlation between expert and LLM scores on BreakHis (ρ=0.939\rho=0.939) and Lizard (ρ=0.926\rho=0.926). LLM scores also correlate positively with automatic metrics, particularly BreakHis accuracy (ρ=0.900\rho=0.900).

W2&Q1: A unified conditioned teacher supplied with the same cues, examples, and compute is the most important missing baseline.

We add a unified conditioned teacher (Fig. 3) using a standardized multi-source dataset. To provide a unified spatial interface across datasets, we represent heterogeneous spatial anchors as bounding boxes (this is commonly available). This retains region-level localization while removing source-specific differences in anchor format to achieve a unified teacher with the full dataset.

MethodBreakHisGleasonXAIGlaSLizardTCGA
Qwen3.5-9B29.1642.1367.5040.1640.97
Qwen3.5-9B (SFT)32.0655.6288.7549.0342.86
Qwen3.5-9B (GRPO)47.2557.9490.0072.7661.27
Unified Expert50.3351.4768.7550.9158.37
EM-OPD57.9861.2190.0085.1365.56

As the table shows, the unified expert shows limited gains on glandular and cellular tasks, suggesting that standardizing spatial anchors might not fully exploit the task-specific spatial information available in their original representations. In contrast, EM-OPD performs better on all five datasets, which demonstrates the effectiveness of multi-expert.

AuthorsOpenReview ↗

Response Continue

W3&Q3: Equation 8 samples trajectories from the parameter-dependent student policy. Differentiating its expectation includes a score-function term for the trajectory distribution. If sampled trajectories are treated as fixed during backpropagation, Equation 11 describes a surrogate or semi-gradient rather than an unbiased gradient of the stated objective.

We agree that the original objective requires a score-function term. Eqs. (8)–(11) now define a fixed-rollout surrogate: a frozen student snapshot θˉ\bar\theta generates trajectories, and backpropagation holds sampled tokens and teacher outputs fixed. Gradients flow through student token distributions, with the rollout policy refreshed during training. Eq. (11) claims unbiasedness only for the surrogate gradient conditional on θˉ\bar\theta.

W4&Q5: The SFT baseline uses successful Gemini-2.5-Pro generations, while the proposed method uses privileged Qwen teacher supervision. Independent validation of the extractor is needed, particularly for negation and uncertainty.

As clarified in Section 4.1, original SFT ensures reports covering annotated key findings from multiple Gemini-2.5-Pro generations, which provides a stronger privilege. To isolate the effect of OPD from potential confounding factors, we add matched SFT with the same teacher model and privileged cues as EM-OPD (Appendix A.6, Table 8).

MethodTeacherTeacher receives cuesBreakHisGleasonXAIGlaSLizardTCGA
Qwen3.5-9B——29.1642.1367.5040.1640.97
Original SFTGemini-2.5-ProNo32.0655.6288.7549.0342.86
Matched SFTQwen3.5-27BYes30.2050.2981.2542.3843.51
EM-OPDQwen3.5-27BYes57.9861.2190.0085.1365.56

As shown in the table, matched SFT performs generally worse than the original SFT baseline. EM-OPD outperforms both SFT variants, which supports the effectiveness of our evidence-conditioned on-policy distillation.

For the extractor, we would like to clarify that report-label extraction is a relatively constrained task based on a predefined label schema. We additionally use Baichuan-M2-32B-Instruct and Gemma-4-31B-it as extractors and observe over 99% label agreement across these models and Qwen3.5-27B.

W5: Claims of consistent superiority should be qualified.

Thank you for pointing out this problem. We have toned down the performance claims, rather than uniform superiority. To avoid a lengthy response, we kindly recommend you check these changes in the revised paper.

Q4: How do four expert categories map to five datasets, and how do the reported resampling repetitions implement equal expert weighting?

The four experts correspond to evidence groups, rather than individual datasets. Diagnosis combines BreakHis and GleasonXAI; Gland uses GlaS; Cell uses Lizard; and Abnormality uses TCGA. For equal expert weighting, as described in Section 3.4, it is implemented through a fixed number of mini-batches sampled from each source at every resampling repetition.

I appreciate the authors’ efforts in adding extensive new results and addressing some of my concerns. However, according to the newly provided results, GPT-5.5 achieves better performance than the proposed method in expert evaluation, which is inconsistent with metric results such as ACC, i.e., the findings in Table 1 and Table 2 conflict with each other. While comparing the 27B model against GPT-5.5 may not be a fair setup, the inclusion of this comparison weakens the strength of the work: despite elaborate methodological designs, the proposed approach still fails to surpass the general large model. Overall, as a report-generation work, its evaluation protocol at the report level remains questionable. I choose to maintain the original score.

Reviewer_UybXOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EM-OPD trains a report-generating student using a shared frozen teacher conditioned on spatial annotations and morphological descriptors. It combines token-level supervision on student-generated trajectories with balanced training across diagnostic, abnormality, glandular, and cellular evidence. Experiments on five datasets report improvements in extracted diagnostic and descriptive attributes, including an increase in Lizard micro-F1 from 40.16 to 85.13 for the 9B student.

Critical strengths

Useful training-only supervision. The method reuses region annotations, segmentation masks, and morphological descriptions from separate sources. The student generates reports using only the original image and prompt at inference, without requiring these annotations or the teacher.

Critical weaknesses

Unmatched SFT comparison. The SFT baseline learns from Gemini-2.5-Pro generations, and EM-OPD uses Qwen teachers with privileged cues. The paper does not clarify whether the SFT teacher receives the same cues. The comparison cannot isolate the benefit of on-policy training from differences in teacher quality and supervision.

Questions for authors
  1. Do the metric improvements translate into more accurate and complete reports? Please add a blinded expert evaluation on a subset, measuring unsupported claims, omissions, and contradictions, and check the accuracy of the LLM-extracted labels against human annotations.
  2. Does EM-OPD still outperform SFT with the same teacher and privileged cues? Please generate SFT targets using the Qwen teacher and match the training data, source weighting, and training budget.
1 comment

We would like to first express our sincere gratitude for the time and effort in reviewing our work. We have taken your comments seriously and hope that our responses address your concerns effectively.


W1&Q2: Unmatched SFT comparison. The SFT baseline learns from Gemini-2.5-Pro generations, and EM-OPD uses Qwen teachers with privileged cues. The comparison cannot isolate the benefit of on-policy training from differences in teacher quality and supervision.

As clarified in Section 4.1, original SFT ensures reports covering annotated key findings from multiple Gemini-2.5-Pro generations, which provides a stronger privilege. To isolate the effect of OPD from potential confounding factors, we add matched SFT with the same teacher model and privileged cues as EM-OPD (Appendix A.6, Table 8).

MethodTeacherTeacher receives cuesBreakHisGleasonXAIGlaSLizardTCGA
Qwen3.5-9B——29.1642.1367.5040.1640.97
Original SFTGemini-2.5-ProNo32.0655.6288.7549.0342.86
Matched SFTQwen3.5-27BYes30.2050.2981.2542.3843.51
EM-OPDQwen3.5-27BYes57.9861.2190.0085.1365.56

As shown in the table, matched SFT performs generally worse than the original SFT baseline. EM-OPD outperforms both SFT variants, which supports the effectiveness of our evidence-conditioned on-policy distillation.

Q1: Do the metric improvements translate into more accurate and complete reports? Please add a blinded expert evaluation on a subset, measuring unsupported claims, omissions, and contradictions, and check the accuracy of the LLM-extracted labels against human annotations.

We add a blinded pathologist review of a subset from BreakHis and Lizard (each source 20). We also add a three-judge LLM evaluation across all five datasets (Sec. 4.3, Table 2), using GPT-6.1 Sol, Gemini 3.8 Flash, and Claude Sonnet 5.5. The evaluation scores image–report pairs from 1–5, considering correctness, completeness, coherence, hallucinations, contradictions, and omissions.

MethodExpert (BreakHis)Expert (Lizard)LLM (BreakHis)LLM (GleasonXAI)LLM (GlaS)LLM (Lizard)LLM (TCGA)
Qwen3.5-9B2.402.502.562.662.432.422.54
Qwen3.5-27B2.702.702.712.862.782.672.62
PathGen-LLaVA2.202.252.362.372.242.432.49
Patho-R1-7B1.852.151.992.111.921.862.50
Gemini 3.1 Pro3.102.953.053.473.043.133.23
GPT-5.53.553.453.363.823.523.313.39
Qwen3.5-9B (SFT)2.552.452.622.632.452.382.55
Qwen3.5-9B (GRPO)2.702.552.732.682.632.582.61
Vision-OPD-9B2.202.052.382.191.822.252.21
Ours-9B (OPSD)2.853.003.082.992.802.942.90
Ours-9B (Teacher 27B)3.153.103.303.263.023.133.59

As shown in the table, our method outperforms most baselines in human evaluation. LLM evaluation likewise favors EM-OPD over the same-scale baselines. We further compute Spearman's rank correlation coefficient (ρ\rho) using LLM scores on the same sampled image–report pairs evaluated by the experts. Table 7 in Appendix A.5 shows strong model-level rank correlation between expert and LLM scores on BreakHis (ρ=0.939\rho=0.939) and Lizard (ρ=0.926\rho=0.926). LLM scores also correlate positively with automatic metrics, particularly BreakHis accuracy (ρ=0.900\rho=0.900).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.