Abstract

Pathological images are inherently multi-scale, requiring pathologists to integrate evidence from global tissue architecture at low magnification to cellular morphology at higher magnification for accurate diagnosis. While existing pathological datasets for vision-language models (VLMs) include various scales, they often lack explicit cross-scale reasoning objectives. This limitation prevents VLMs from capturing essential cross-scale representations and learning evidence-based reasoning. To bridge this gap, we introduce the first cross-scale training and evaluation paradigm that formulates pathology interpretation as multi-magnification reasoning. However, creating such a task reveals a critical challenge: multi-image visual question answering (VQA) is prone to text-only shortcuts, which allow models to guess answers using magnification-dependent artifacts rather than visual evidence. To address this, we propose a leakage-aware curation pipeline that combines adversarial text-only screening with constraint-guided question design. Using this pipeline, we construct Scale-VQA, a high-quality benchmark with 4,685 multiple-choice questions grounded in 2,537 pathology images across multiple magnification levels. Finally, we present ScaleReasoner-R1, a model trained via reinforcement learning to optimize performance on cross-scale VQA tasks. ScaleReasoner-R1 achieves state-of-the-art performance on our cross-scale reasoning benchmark and generalizes to SOTA performance on established single-scale benchmarks. Findings suggest that even the limited cross-scale supervision can significantly improve pathological understanding. Code is available at https://github.com/iMVR-PL/ScaleReasoner-R1.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3574_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/iMVR-PL/ScaleReasoner-R1

Link to the Dataset(s)

N/A

BibTex

@InProceedings{PhaChi_Enhancing_MICCAI2026,
        author = { Phan, Chi AND Zhang, Tianyi AND Xue, Qiaochu AND Wu, Yufeng AND Hu, Dan AND Liu, Zeyu AND Wang, Sudong AND Jin, Yueming},
        title = { { Enhancing Pathological VLMs with Cross-scale Reasoning } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16879},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This research introduces a cross-scale training and evaluation paradigm for pathological vision language models to better replicate the way human pathologists integrate global tissue patterns with fine cellular details. The authors developed SCALE-VQA which is a high quality benchmark featuring thousands of multiple choice questions grounded in images across different magnifications and created using a curation pipeline designed to eliminate text only shortcuts. They also presented SCALEREASONER-R1 which is a model trained with reinforcement learning that achieves state of the art results on both the new cross-scale tasks and existing single scale pathology benchmarks.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.The study introduces a novel cross scale reasoning paradigm that mirrors clinical pathology by requiring models to connect global tissue patterns with fine cellular details across different zoom levels. This approach is significant because it moves beyond simple image classification and forces the AI to learn the hierarchical logic pathologists use to verify architectural suspicions with high power visual evidence. By framing diagnosis as a coordinated process across scales the work addresses a major limitation in existing models that often struggle with tasks requiring multi view verification. 2.The researchers developed a leakage aware curation pipeline that uses adversarial screening to stop models from cheating through text only shortcuts. By employing multiple language models as adversaries to identify questions solvable without images the authors ensured that the SCALE-VQA benchmark reflects true visual reasoning. This rigorous data strategy is interesting because it addresses a widespread problem in vision language modeling where high accuracy often stems from linguistic biases rather than actual image understanding. 3.The model utilizes reinforcement learning through group relative policy optimization to achieve superior generalization and state of the art performance on both multi scale and single scale benchmarks. Unlike standard supervised fine tuning which can lead to catastrophic forgetting of original representations this reinforcement learning approach allows the model to discover optimal decision rules from outcome based rewards. This demonstrates a highly efficient way to build medical experts that remain robust across different diagnostic settings even when trained on a relatively small amount of high quality data.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The training dataset relies on a very small scale of 177 whole slide images. While the authors curated 4685 questions from this set the limited number of underlying source images restricts the diversity of pathological patterns the model encounters. This small scale limits the generalizability of the model to rare diseases or varying tissue preparations found in broader clinical practice compared to models trained on tens of thousands of slides. 2.The methodological approach lacks algorithmic novelty because the reinforcement learning optimization relies entirely on existing techniques. The authors use Group Relative Policy Optimization to train their model. This specific reinforcement learning method is an established algorithm previously introduced in other foundation model works. The primary technical contribution is applying this existing optimization method to a newly curated medical dataset rather than proposing a fundamentally new model architecture. 3.The evaluation is restricted to multiple choice visual question answering benchmarks which does not fully demonstrate true clinical feasibility. While the model is evaluated on the SCALE VQA and PathMMU datasets answering structured questions is far removed from real world diagnostic workflows. There is no evaluation showing how this model integrates into an actual clinical setting or whether it improves the diagnostic speed and accuracy of human pathologists working on open ended complex cases.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The recommendation for a weak accept stems from the paper’s effective strategy of aligning model behavior with the multi magnification reasoning used by human pathologists. By introducing the SCALE-VQA benchmark and a curation pipeline that explicitly blocks text only shortcuts the authors have created a reliable way to measure genuine visual integration. While the training data is limited to a small number of slides and the reinforcement learning method is not fundamentally new the model achieves impressive results on both the proposed tasks and standard single image benchmarks. The combination of a strong practical motivation with the release of high quality open source resources provides a solid contribution to the field of computational pathology.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    Generally, the paper presents a novel cross-scale reasoning paradigm for pathological vision-language models. Post-rebuttal, major concerns regarding dataset leakage and evaluation bias have been partially resolved. Therefore, the paper is recommended for acceptance.



Review #2

  • Please describe the contribution of the paper

    The paper proposes a data creation and training pipeline with extensive experimenation in pathology domain.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    the paper is presented well. the proposed pipeline is novel and reasons the steps in the methodology in a good manner. the experimentation is extensive both on the models and the datasets. the formulation of multiple tasks (i.e. deature decompistion, mcqs, RL) is novel and has proven to be effective in strengthening the reasoning of models. the problem at hand is addressed with domain knowledge integrated in the pipeline.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    as the proposed work includes multiple steps and models, the ablation is on the weaker side. a detailed ablation can help strengthen the arguement the paper is trying to make for each step.

    the comparisons with models can be made more trustworthy by including the large medical models. currently, the second best results are mostly from commerical models (GPT, Gemini).

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper formulates the problem of data and training in a novel way addressing domain challenges. However, to strengthen the paper, the comparison needs be made more fair and ablations should include each step of the pipeline.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    I liked the work, but as pointed out earlier, I’m not satisfied with the evaluation methods and models being used for the study.



Review #3

  • Please describe the contribution of the paper

    This paper tries to solve a real problem that I think has been ignored for too long. Most pathology VLMs only look at one magnification level at a time, but real pathologists dont work that way. They zoom in and out and connect what they see at different scales. So the authors built Scale-VQA, a benchmark with 4685 questions from 177 TCGA whole slide images, annotated at three magnification levels by experienced pathologists. They also made a pipeline to prevent models from guessing the right answer just from the text without looking at the images. And finally they trained ScaleReasoner-R1 using reinforcement learning which beats other models on both their new benchmark and the existing PathMMU benchmark.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    I think the motivation of this paper is one of the strongest parts. It is clinically very logical. Pathologists do not just stare at one zoom level, they move between low and high magnification to build their diagnosis. I have not seen many papers that actually try to teach a model to do the same thing in a structured way, so this feels like a genuine gap being filled.The leakage-aware curation pipeline is what I found most interesting technically. The idea of using other strong models like Gemini and Qwen as adversaries to test whether a question can be answered without the images is clever. It is a simple but effective quality control step and I think other benchmark papers should adopt something similar.I also appreciated that pathologists were involved from the very beginning of annotation, not just at the end for validation. This is not always done and it makes the dataset more trustworthy from a clinical standpoint.The result showing that training on cross-scale data also improves single-scale performance on PathMMU is interesting and a bit surprising. It suggests the model is learning something more general about pathology reasoning, not just memorizing cross-scale patterns.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    My biggest concern is the size of the dataset. 177 whole slide images across 15 organ types is quite small. Some organs have around 20 slides which is really not enough to draw strong conclusions. I understand it is hard to get expert annotations at scale but the authors should at least discuss this limitation more honestly instead of making broad generalization claims. I have noticed that ScaleReasoner-R1 actually performs worse than Patho-R1 on the PathCLS subset, both on tiny and full test sets. This is a clear regression and the authors never mention it anywhere in the paper. This bothers me because it suggests the model may be forgetting something specific to that subset and the reader deserves an explanation. Another thing that I think is a problem is that GPT-5.2 was used to generate the benchmark questions and it is also one of the compared models in the evaluation table. This creates a circular dependency. It is hard to know if GPT-5.2 is being evaluated fairly here or if it has some kind of advantage or disadvantage because of its role in creating the data. The authors should at least acknowledge this openly. The reward weighting lambda equals 0.2 is just stated without any ablation or explanation. I would like to know if the results change meaningfully with different values. It seems like an important choice that was made without justification. There are also no statistical significance tests anywhere. With 380 samples per reasoning dimension some of the differences between models might not be meaningful and we cannot tell from the paper as written.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    I think separating the benchmark contribution from the model contribution more clearly in the writing would help readers evaluate each part independently. Right now they are mixed together which makes it a bit harder to follow. Also I would find per-organ accuracy breakdowns useful because averaging across 15 very different organ types can hide a lot of variation. One small thing I noticed is that the GPT-5.2 reference seems unusual and the arXiv identifier should be double checked for accuracy.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I think this paper has a genuinely good idea at its core and the leakage-aware pipeline is a real contribution that the community can benefit from. The transfer from cross-scale to single-scale performance is also a nice finding. But the dataset is small, the PathCLS regression is unexplained, the GPT-5.2 circularity is unaddressed, and there are no significance tests. These are not minor issues. I am willing to accept this paper if the authors respond clearly to these points in the rebuttal, especially the PathCLS drop and the GPT-5.2 overlap between data construction and evaluation.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We thank all reviewers for the insightful feedback and will address main concerns below [1] Dataset scale (R1, R3): We acknowledge the limited scale and will avoid broad generalization claims. We’re actively scaling to include more WSIs, organs, and pathological patterns. Beyond the dataset, our contribution includes a reproducible, leakage-aware curation pipeline for visually grounded cross-scale patho VQA, which was appreciated as a “rigorous” strategy (R1) that “other benchmark papers should adopt” (R3). We’ll open-source Scale-VQA and the curation pipeline to benefit the community to build larger datasets. Also, our results suggest that a small but high-quality, outcome-verifiable cross-scale data can provide effective RL signals for patho VLM, consistent with recent evidence [a]. For WSI-level separation, multi-scale views and QA derived from the same WSI stay in the same split to avoid leakage [2] PathCLS regression (R3): PathCLS is converted from classification datasets into MCQs, whereas other PathMMU subsets are built from image-text sources with morphology-focused descriptions. Thus, PathCLS is closer to single-image category recognition, while the others more align w/ descriptive, evidence-grounded reasoning. Since our RL objective emphasizes cross-scale evidence integration, fine-tuning may shift the model away from some label-centric priors learned by Patho-R1, causing a trade-off on this subset, while improving other evidence-oriented subsets and overall PathMMU [3] GPT5.2 circularity (R3): We acknowledge the concern. To clarify, GPT-5.2 was used in text-based construction stage to draft candidate MCQs from cross-scale captions; its outputs weren’t directly accepted as final QA. The benchmark was finalized through leakage-aware screening and pathologist verification of visual support, answer correctness, and distractor plausibility. While GPT-5.2 may have advantages in text-based analysis, eval on our final pathologist-verified questions requires cross-scale visual evidence integration and image-text alignment ability. As shown in Tab.1, GPT has 58.51% ACC on Scale-VQA, far below ScaleReasoner-R1(82.89), suggesting that the commercial VLMs may rely on textual priors in med-VQA, and hence struggle with the cross-scale grounded visual reasoning required by our bench. This is also consistent with existing literature [b] [4] Reward weight (R3): We apologize for a notation error: answer correctness weight is 0.8, format is 0.2.As Patho-R1 already exhibits instruction-following ability, the format reward regularizes malformed outputs, while the dominant acc reward drives cross-scale reasoning improvement. We’ll correct the formula and add justification [5] Ablation&Comparisons&Stats(R2): Ablations on key pipeline components: (i)after leakage screening, text-only adversary ACC (%) drops to 32.8 (Qwen3Max) and 29.5 (Gemini3Pro), close to chance (25%), confirming shortcut suppression; (ii) Avg across 4 models and 5 dimensions, dropping 1 view reduces ACC (%) by 3.8, dropping 2 by 8.5, text-only by 17.8, confirming multi-scale dependency. Our experiments already include representative 7B medical VLMs for a fair same-param comparison in Tab.1-2.For stats significance, ScaleReasoner-R1 achieves 82.89 ACC, CI [80.43, 85.09], vs. Gemini (2nd best) at 60.90, CI [57.84, 63.87]. McNemar’s test confirms significance (χ² = 141.11, p < 0.001) across all dimensions [6] Clinical feasibility & novelty (R1): We agree that MCQ eval doesn’t fully demonstrate real-world feasibility. Our clinical involvement lies in expert annotation/validation, and Scale-VQA is positioned as a controlled benchmark for cross-scale visual integration. We also clarify that our novelty is not GRPO, but the clinically grounded cross-scale reasoning formulation&curation pipeline that makes outcome-based RL meaningful for patho VLMs. [a] Wang et al.RL for Reasoning in LLMs with One Training Example [b] Asadi et al. Mirage the illusion of visual understanding




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The paper presents a clinically motivated cross-scale pathology VQA benchmark and training paradigm that addresses an important limitation of current pathology VLMs, which typically operate on isolated single-scale images despite pathologists routinely integrating evidence across magnifications. The Scale-VQA benchmark, pathologist-guided annotation process, and leakage-aware curation pipeline are meaningful contributions, and the reported performance of ScaleReasoner-R1 on the proposed benchmark is strong. The finding that cross-scale supervision transfers to overall PathMMU performance is also interesting.

    The main concerns are about interpretation and scope. The benchmark is built from a relatively small number of WSIs, so the generalization claims should be framed carefully. The authors should clarify in rebuttal how WSI-level separation was maintained, how GPT-5.2 was used during data/question construction relative to its role as an evaluated baseline, and how to interpret the observed regression on the PathCLS subset compared with Patho-R1.The rebuttal should also clarify the choice of the RL reward weighting and the extent to which leakage-aware screening was independent of the final evaluated models.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    I recommend acceptance. The rebuttal addressed the main concerns by clarifying that multi-scale views and QA from the same WSI are kept in the same split to avoid leakage, that GPT-5.2 was used only to draft candidate MCQs from text captions before leakage screening and pathologist verification with its final Scale-VQA accuracy of 58.51% well below ScaleReasoner-R1 at 82.89%, that the PathCLS regression reflects a trade-off between cross-scale evidence integration and label-centric priors, and that the reward weighting is 0.8 for correctness and 0.2 for format with a notation error acknowledged. Additional clarifications on leakage-screening independence (text-only adversary accuracy near chance), multi-scale dependency, and McNemar significance testing (p<0.001) addressed R2 and R3’s points, leading R1 and R3 to maintain their positive recommendations. R2 lowered the score to Reject, citing evaluation and model-comparison concerns, and the small dataset size of 177 WSIs remains a limitation. Still, these are framing-level issues that cannot be expanded under the MICCAI rebuttal policy. The leakage-aware curation pipeline and cross-scale reasoning formulation are recognized as meaningful contributions to the community.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The paper introduces a cross-scale training and evaluation paradigm for pathology vision-language models. The main contributions are SCALE-VQA, a curated benchmark of multiple-choice questions grounded in pathology images across different magnifications, and SCALEREASONER-R1, a reinforcement-learning-trained model intended to improve cross-scale pathological reasoning by linking global tissue patterns with fine cellular details.

    The initial reviews were uniformly but weakly positive, with scores of 4 / 4 / 4.Reviewers appreciated the importance of cross-scale reasoning in pathology, the attempt to mimic pathologists’ diagnostic workflow, the leakage-aware curation strategy, and the potential value of SCALE-VQA as a community benchmark. The main concerns were the limited dataset size, possible data leakage or circularity in the data construction and evaluation pipeline, the small number of WSIs, insufficient validation of whether the benchmark truly requires visual cross-scale reasoning rather than exploiting textual or linguistic shortcuts, unclear PathCLS regression, limited statistical testing, and limited comparison to broader VLM baselines.

    The rebuttal addressed several key concerns. The authors acknowledged the limited dataset scale and agreed to avoid broad generalization claims. They clarified the leakage-aware curation pipeline, stating that QA pairs from the same WSI remain in the same split to avoid leakage. They also explained the PathCLS regression as a trade-off caused by shifting the model toward cross-scale evidence integration rather than single-image category recognition. In response to concerns about GPT-based circularity, they stated that GPT outputs were used as draft candidate MCQs but not directly accepted as final QA, and that the benchmark was finalized through leakage-aware screening and pathologist verification. The rebuttal also clarified the reward-weight notation and summarized ablation and comparison evidence supporting multi-scale dependency.

    Some concerns remain. The dataset is still relatively small, and the evaluation remains limited compared with the breadth of claims about pathology VLM reasoning. Some reviewers remained concerned about the benchmark construction process, possible residual leakage/circularity, and whether the task fully distinguishes visual reasoning from language priors. The final version should clearly describe the dataset construction pipeline, leakage-control steps, GPT-assisted generation and human verification process, statistical testing, and the limitations of the benchmark scale.

    After rebuttal, the reviewer opinions are A / R / A. Reviewer 1 recommended accept, stating that the major concerns regarding dataset leakage and evaluation bias were partially resolved. Reviewer 2 moved to reject, remaining dissatisfied with the evaluation methods and model comparisons. Reviewer 3 recommended accept and viewed the leakage-aware curation and cross-scale reasoning framing as valuable enough to create useful discussion.

    I recommend Accept. The paper is not without limitations, especially regarding dataset scale and the need for more rigorous evaluation of visual reasoning versus language priors. However, the problem is timely and important, the cross-scale pathology VLM framing is relevant to clinical diagnostic workflows, and the rebuttal addressed the most serious concerns about leakage and benchmark construction sufficiently for acceptance. The work should be accepted with clear limitations, and the camera-ready version should carefully document the curation pipeline, leakage prevention, model-comparison setting, and residual benchmark limitations, but the authors should avoid adding new experiments or new baseline comparisons beyond what is permitted by the MICCAI guidelines.



back to top