Enhancing Pathological VLMs  with Cross-scale Reasoning

Phan, Chi; Zhang, Tianyi; Xue, Qiaochu; Wu, Yufeng; Hu, Dan; Liu, Zeyu; Wang, Sudong; Jin, Yueming

ScaleReasoner-R1:

Enhancing Pathological VLMs
with Cross-scale Reasoning

Chi Phan^1,*Tianyi Zhang^1,*Qiaochu Xue¹Yufeng Wu²Dan Hu³Zeyu Liu²Sudong Wang^2,†Yueming Jin^1,†

¹ National University of Singapore, Singapore² PuzzleLogic Pte Ltd, Singapore³ Department of Pathology, Fujian Medical University Cancer Hospital & Fujian Cancer Hospital, Fuzhou, China

* Equal contribution · † Corresponding authors

Read the paper Code Model Dataset

01 / Overview

Pathological reasoning across scales.

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:

Scale-VQA: a high-quality benchmark with 4,685 multiple-choice questions grounded in 2,537 pathology images across multiple magnification levels.
ScaleReasoner-R1: a model trained via reinforcement learning to optimize performance on cross-scale VQA tasks, achieves best performance on our cross-scale reasoning benchmark and also generalizes to have SOTA performance on established single-scale benchmarks.

Cross-scale questions: 4,685
Organs represented: 15
Reasoning dimensions: 5
Cross-scale accuracy: 82.89%

Method overview: scale-specific feature decomposition, text-only adversarial screening and clinical validation create Scale-VQA; GRPO trains ScaleReasoner-R1 on the resulting cross-scale questions. — Figure 1. (a) Leakage-aware curation pipeline. (b) Dataset coverage across organs and magnifications. (c) GRPO-based reinforcement learning. Select the figure to view it at full resolution.

02 / Scale-VQA

Visual evidence comes first.

A benchmark designed to require images, with 4,685 multiple-choice questions across 15 organs and five clinically aligned reasoning dimensions.

Correspondence

Do findings at different scales describe the same underlying phenomenon?

Confirmation

Does evidence at one scale confirm or contradict the interpretation at another?

Localization

Which tissue compartment does a finding belong to, based on cross-scale evidence?

Explanation

What process explains the relationship between findings across scales?

Diagnosis

What diagnosis is supported by integrating evidence across scales?

Leakage-aware curation

Beyond text-only shortcuts.

Questions should depend on the images. Our curation pipeline screens out questions that can be answered using linguistic or biomedical priors alone.

Decompose scale-specific features
Expert annotations become per-scale evidence sets, with visual-grounding and scale-dependency constraints.
Challenge with text-only adversaries
Gemini 3 Pro and Qwen3-Max attempt questions without images. If either answers correctly, constraints are tightened and questions are regenerated.
Validate clinically
Senior pathologists review the final questions for visually grounded answers and clinically plausible distractors.

03 / ScaleReasoner-R1

Learning to integrate the evidence.

Initialized from Patho-R1-7B, ScaleReasoner-R1 is fine-tuned on Scale-VQA using GRPO. The model takes multi-scale images, a question, and answer options, then produces a structured reasoning trace and a final answer.

Patho-R1-7BScale-VQAGRPOScaleReasoner-R1

Training reward curve for ScaleReasoner-R1 during reinforcement learning — Training reward. Reward evolution during GRPO training.

Validation accuracy curve for ScaleReasoner-R1 during reinforcement learning — Validation accuracy. Performance on the validation split during training.

04 / Evaluation

Stronger cross-scale reasoning.
Transfer to single-image tasks.

ScaleReasoner-R1 leads across all five reasoning dimensions among the compared models and improves overall accuracy on PathMMU.

Evaluation figure comparing ScaleReasoner-R1 with baseline models — Evaluation overview. Results from the accompanying research repository.

Cross-scale multi-image VQA Accuracy (%) · higher is better
Model	Correspondence	Confirmation	Localization	Explanation	Diagnosis	Average
Qwen2.5-VL-7B	41.79	47.26	71.14	44.28	63.68	53.63
Gemini 3 Flash	48.26	58.71	71.64	53.73	72.14	60.90
GPT-5.2	47.76	59.70	74.13	45.77	65.17	58.51
LLaVA-Med-7B	25.87	16.92	28.36	20.90	24.88	23.38
Quilt-LLaVA	32.34	14.43	45.27	30.35	29.35	30.35
CLOVER	37.31	61.69	73.13	46.27	65.77	56.82
Patho-R1	31.84	40.30	59.70	56.22	68.66	51.34
ScaleReasoner-R1	80.60	89.05	84.58	76.12	84.08	82.89

Single-image VQA · PathMMU Accuracy (%) · higher is better · scroll horizontally for all subsets
Model	Val overall	Test overall		PubMed		SocialPath		EduContent		Atlas		PathCLS
Model	Val overall	Tiny	All	Tiny	All	Tiny	All	Tiny	All	Tiny	All	Tiny	All
Qwen2.5-VL-7B	44.3	48.2	44.4	53.4	50.5	50.0	47.7	59.6	48.9	48.6	46.8	29.4	28.0
Gemini Pro Vision†	41.9	42.8	42.7	43.8	44.9	42.4	42.0	43.5	43.7	49.5	49.4	32.8	34.7
GPT-4V†	49.3	53.9	49.8	59.4	53.5	58.7	53.9	60.4	53.6	48.1	52.8	36.2	33.8
LLaVA-Med-7B	17.5	22.2	22.7	24.6	25.1	23.4	21.1	25.1	23.6	17.8	24.5	20.3	19.2
HuatuoGPT-V-7B	43.2	43.8	41.2	49.1	47.7	52.3	46.0	52.9	46.1	44.7	46.9	19.8	19.2
Lingshu-7B	51.9	56.0	53.7	60.5	57.6	61.6	58.1	69.4	58.7	58.1	62.5	30.4	31.6
Quilt-LLaVA	33.8	32.3	31.9	30.3	34.2	26.8	34.6	35.7	33.3	41.8	33.3	27.1	24.0
CLOVER	53.1	60.6	56.6	68.0	59.8	67.6	60.5	68.6	62.2	62.0	64.1	36.7	36.6
Patho-R1	62.3	64.8	62.6	68.7	64.9	63.9	65.7	71.0	66.1	78.4	73.5	41.8	42.7
ScaleReasoner-R1	62.4	66.2	63.8	71.2	67.3	67.6	67.6	74.9	69.2	79.3	76.2	37.9	38.5

Tiny and All denote Test-Tiny and the full test set; subset scores are shown for both splits. Bold marks the best model score in each column. † Results taken directly from the original PathMMU paper.

Qualitative evaluation

Case studies across five reasoning dimensions

Explore multi-magnification evidence and model responses for correspondence, confirmation, localization, explanation, and diagnosis.

Cross-scale correspondence case study: multi-magnification images, question, ground truth, and model responses — Correspondence. Multi-scale evidence and response comparisons for cross-scale correspondence.

Enable JavaScript to browse the gallery, or open each figure: Correspondence, Confirmation, Localization, Explanation, Diagnosis.

05 / Citation

Build on this work.

If you find ScaleReasoner-R1 or Scale-VQA useful in your research, please cite our paper.

@article{phan2026enhancing,
  title={Enhancing Pathological VLMs with Cross-scale Reasoning},
  author={Phan, Chi and Zhang, Tianyi and Xue, Qiaochu and Wu, Yufeng and Hu, Dan and Liu, Zeyu and Wang, Sudong and Jin, Yueming},
  journal={arXiv preprint arXiv:2606.17412},
  year={2026}
}

Acknowledgements

This work was supported by the Ministry of Education, Singapore, under the Tier 1 grant (24-1250-P0001) and Tier 2 grant (T2EP20224-0028), and by PuzzleLogic Pte Ltd, Singapore.

We thank the contributors to verl, LLaMA-Factory, vLLM, and Patho-R1, along with the developers of the open-source models used in our experiments.

Enhancing Pathological VLMs with Cross-scale Reasoning