arxiv: 2604.17570 · v1 · submitted 2026-04-19 · 💻 cs.CV · cs.AI

Recognition: unknown

PBSBench: A Multi-Level Vision-Language Framework and Benchmark for Hematopathology Whole Slide Image Interpretation

Yuanlong Wang , Weichi Chen , Adrian Rajab , Wenfang Liu , Yulan Jin , Andrew Srisuwananukorn , Ping Zhang

Authors on Pith no claims yet

Pith reviewed 2026-05-10 05:49 UTC · model grok-4.3

classification 💻 cs.CV cs.AI

keywords peripheral blood smearvision-language modelhematopathologywhole slide imagevisual question answeringcell morphology analysispathology benchmark

0 comments

The pith

A specialized vision-language model trained on blood smear data outperforms general pathology AI on hematopathology tasks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper builds the first large vision-language dataset for peripheral blood smears, which are microscope slides used to diagnose blood disorders by looking at individual cell shapes rather than tissue patterns. Existing AI models trained mostly on solid tissue samples do not transfer well to this cell-focused work, so the authors create PBSInstr with hundreds of slides, tens of thousands of cell images, and question-answer pairs, then fine-tune their PBS-VL model on it. They also release PBSBench, a set of visual questions across four categories, to test how well models understand PBS at both cell and whole-slide levels. A reader would care because better AI tools here could support faster and more consistent decisions by hematopathologists reviewing blood samples.

Core claim

We construct PBSInstr, the first vision-language dataset for PBS interpretation comprising 353 WSIs with microscopic impressions, 29k cell-level crops with type and morphology labels, and over 28k QA pairs. Building on this, we develop PBS-VL, a hematopathology-tailored vision-language model for multi-level interpretation at cell and slide levels. On the new PBSBench benchmark with four question categories and six tasks, PBS-VL outperforms existing general-purpose and pathology MLLMs, showing the value of PBS-specific training data.

What carries the argument

PBS-VL, a vision-language model instruction-tuned on PBS-specific cell crops, slide images, and QA pairs to perform both cell morphology classification and higher-level slide interpretation.

If this is right

PBS-specific instruction data produces measurable gains over models trained only on tissue pathology images.
A single model can address both cell-level details and full-slide context in the same framework.
PBSBench supplies a standardized testbed with defined tasks for comparing future PBS interpretation systems.
Public release of the dataset, benchmark, and model weights enables other researchers to build on this starting point.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same data-construction approach could be applied to create resources for related microscopic exams such as bone marrow aspirates or urine cytology.
Clinical deployment might allow AI to flag unusual cell morphologies in real time during routine smear review, potentially shortening turnaround for common blood disorder cases.
Combining this model with electronic health record data on patient history could support more integrated diagnostic suggestions beyond image analysis alone.

Load-bearing premise

That the PBSInstr dataset and the four question categories in PBSBench are representative enough of real-world blood smear variability and diagnostic reasoning to support general claims about model improvement.

What would settle it

Measuring whether PBS-VL still outperforms general models when tested on a fresh collection of PBS slides from different hospitals, staining protocols, or patient demographics not seen in the original dataset.

Figures

Figures reproduced from arXiv: 2604.17570 by Adrian Rajab, Andrew Srisuwananukorn, Ping Zhang, Weichi Chen, Wenfang Liu, Yuanlong Wang, Yulan Jin.

**Figure 2.** Figure 2: Illustration of the performance of existing models on PBSBench for both cell-level and slide-level QA. The metrics are normalized [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Figure 3: The Overview of our pipeline. pairs over 4,998 images sourced from two widely used pathology textbooks and the PEIR digital library, offering the first broad benchmark for pathology-specific VQA beyond generic medical visual QA. Educational video extraction. To scale image-text pairs, Quilt-1M [18] exploits publicly available YouTube histopathology teaching videos (1,087 hours) and additional web sources… view at source ↗

**Figure 4.** Figure 4: Case study of model performance of GPT-4o, PathGen-LLaVA, and our proposed model on each question type. We highlight the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

**Figure 5.** Figure 5: Quality control examples. Poor-quality patches are shown in the red box, and a good-quality patch is shown in the green box. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

**Figure 6.** Figure 6: Additional case study on GPT-4o, the model shows performance degradation when no cell types are available as hints in questions. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗

**Figure 7.** Figure 7: Prompts for cell crop captioning [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗

**Figure 8.** Figure 8: Prompts for whole slide captioning [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗

**Figure 10.** Figure 10: Prompts for whole slide QA [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗

read the original abstract

Peripheral Blood Smear (PBS) is a critical microscopic examination in hematopathology that yields whole-slide imaging (WSI). Unlike solid tissue pathology, PBS interpretation focuses on individual cell morphologies rather than tissue architecture, making it distinct in both visual characteristics and diagnostic reasoning. However, current multimodal large language models (MLLMs) for pathology are primarily developed on solid-tissue WSIs and struggle to generalize to PBS. To bridge this gap, we construct PBSInstr, the first vision-language dataset for PBS interpretation, comprising 353 PBS WSIs paired with microscopic impression paragraphs and 29k cell-level image crops annotated with cell type labels and morphological descriptions. To facilitate instruction tuning, PBSInstr further includes 27k question-answer (QA) pairs for cell crops and 1,286 QA pairs for PBS slides. Building upon PBSInstr, we develop PBS-VL, a hematopathology-tailored vision-language model for multi-level PBS interpretation at both cell and slide levels. To comprehensively evaluate PBS understanding, we construct PBSBench, a visual question answering (VQA) benchmark featuring four question categories and six PBS interpretation tasks. Experiments show that PBS-VL outperforms existing general-purpose and pathology MLLMs, underscoring the value of PBS-specific data. We release our code, datasets, and model weights to facilitate future research. Our proposed framework lays the foundation for developing practical AI assistants supporting decision-making in hematopathology.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper's real contribution is releasing the first PBS-specific vision-language dataset and benchmark, though the modest scale of 353 WSIs makes the outperformance claims hard to generalize without more checks.

read the letter

The main thing to know is that the authors have built PBSInstr, a vision-language dataset with 353 PBS whole-slide images, 29k cell crops, morphological annotations, and over 27k QA pairs, then used it to train PBS-VL and created PBSBench with four question categories and six tasks. They report that their model beats both general-purpose and existing pathology MLLMs, and they release the data, code, and weights. That is the concrete new piece: prior pathology MLLM work has stayed with solid-tissue slides, where architecture matters more than individual cell morphology, so PBS has been underserved. The multi-level setup (cell crops plus full slides) and the decision to target hematopathology directly are reasonable responses to that gap. Releasing everything openly is the part that actually helps other groups who want to experiment with PBS data rather than starting from scratch. The paper does a clean job of spelling out why PBS interpretation differs from the usual pathology tasks and of structuring the benchmark around practical question types. The soft spots sit mostly around scale and coverage. Three hundred fifty-three WSIs is a modest collection when real-world PBS varies by staining protocol, lab, patient demographics, and rare cell populations. It is not obvious yet whether the QA pairs were validated for clinical accuracy or whether the four categories and six tasks probe the full sequence of steps a hematopathologist actually uses. The outperformance numbers are stated in the abstract, but without seeing the splits, statistical tests, or ablation details it is difficult to tell how much is domain-specific signal versus dataset-specific fitting. This work is aimed at researchers building or evaluating medical vision-language models who need a PBS starting point or a new testbed. Anyone working on clinical decision support in hematopathology could get practical value from the released assets once they are checked. It is worth sending to peer review so the data construction and evaluation choices get proper scrutiny; the resources themselves are the part that could matter more than the particular model.

Referee Report

2 major / 2 minor

Summary. The paper introduces PBSInstr, the first vision-language dataset for peripheral blood smear (PBS) interpretation comprising 353 WSIs with impression paragraphs, 29k cell-level crops with type labels and morphological descriptions, and over 27k QA pairs (plus 1,286 slide-level QA pairs). It develops PBS-VL, a hematopathology-specific MLLM for multi-level (cell and slide) interpretation, and PBSBench, a VQA benchmark with four question categories and six tasks. Experiments show PBS-VL outperforming general-purpose and pathology MLLMs, and the authors release code, datasets, and model weights.

Significance. If the empirical results hold under scrutiny, the work fills a clear gap by providing the first dedicated PBS resources, recognizing that PBS analysis centers on cell morphology rather than tissue architecture and that existing pathology MLLMs do not transfer well. The explicit release of datasets, code, and weights is a concrete strength that supports reproducibility and future extensions in this specialized domain.

major comments (2)

[§3] §3 (PBSInstr construction): The 353-WSI collection is described only by aggregate counts (29k crops, 27k+ QA pairs) with no details on acquisition sources, staining variability, patient demographics, or inclusion of rare cell types/morphologies. This omission directly undermines the central inference that measured gains demonstrate the value of PBS-specific data rather than an artifact of a narrow collection.
[§5] §5 (Experiments): The headline outperformance tables report raw metrics for PBS-VL versus baselines but contain no statistical significance tests, confidence intervals, or variance estimates across runs or data splits. Without these, the claim that PBS-VL 'outperforms' cannot be distinguished from training or evaluation noise and is therefore not yet load-bearing evidence for the paper's conclusion.

minor comments (2)

[PBSBench] The mapping from the four question categories to the six PBS interpretation tasks is stated at a high level; an explicit table or diagram would clarify coverage of diagnostic reasoning steps.
[Abstract and §3] Minor notation inconsistency: '27k' is used for cell-crop QA pairs while '1,286' is given exactly for slide-level pairs; uniform precision or a breakdown by task would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback, which highlights important aspects for improving the clarity and rigor of our work. We address each major comment below and will incorporate revisions to strengthen the manuscript.

read point-by-point responses

Referee: [§3] §3 (PBSInstr construction): The 353-WSI collection is described only by aggregate counts (29k crops, 27k+ QA pairs) with no details on acquisition sources, staining variability, patient demographics, or inclusion of rare cell types/morphologies. This omission directly undermines the central inference that measured gains demonstrate the value of PBS-specific data rather than an artifact of a narrow collection.

Authors: We agree that greater transparency on dataset construction is necessary to substantiate our claims. In the revised manuscript, we will expand Section 3 with details on WSI acquisition sources (specific clinical sites and laboratories), staining protocols and observed variability, patient demographics (age and sex distributions, subject to ethical de-identification constraints), and the full distribution of cell types including rare morphologies (e.g., blasts, schistocytes, and atypical forms). These additions will demonstrate the collection's breadth and support that performance gains arise from PBS-specific characteristics rather than narrow sampling. revision: yes
Referee: [§5] §5 (Experiments): The headline outperformance tables report raw metrics for PBS-VL versus baselines but contain no statistical significance tests, confidence intervals, or variance estimates across runs or data splits. Without these, the claim that PBS-VL 'outperforms' cannot be distinguished from training or evaluation noise and is therefore not yet load-bearing evidence for the paper's conclusion.

Authors: We concur that statistical validation is required for robust conclusions. In the revised version, we will augment the experimental results in Section 5 with appropriate statistical tests (e.g., paired t-tests or Wilcoxon signed-rank tests on per-task metrics), 95% confidence intervals, and variance estimates derived from multiple independent training runs with different random seeds or bootstrap resampling of the evaluation sets. This will provide quantitative support for the observed improvements. revision: yes

Circularity Check

0 steps flagged

No significant circularity: empirical benchmark paper with held-out evaluation

full rationale

The paper constructs PBSInstr (353 WSIs, cell crops, QA pairs) and PBSBench (four categories, six tasks), trains PBS-VL via instruction tuning, and reports empirical outperformance on the benchmark. No mathematical derivation, equations, or first-principles results exist that could reduce to inputs by construction. Performance claims rest on experimental comparisons against general and pathology MLLMs using held-out evaluation rather than fitted parameters or self-referential predictions. No self-citation load-bearing, uniqueness theorems, or ansatz smuggling appear in the derivation chain. The central claim (value of PBS-specific data) is supported by external benchmark results, not tautological redefinition of the training data itself.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities beyond standard ML training assumptions.

axioms (1)

domain assumption Standard machine learning assumptions hold, including that training and test splits are representative and that VQA performance correlates with clinical utility.
Implicit in any empirical VQA benchmark paper for medical imaging.

pith-pipeline@v0.9.0 · 5576 in / 1112 out tokens · 33377 ms · 2026-05-10T05:49:13.902495+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

62 extracted references · 10 canonical work pages · 5 internal anchors

[1]

Claude sonnet 4.5.https : / / www

Anthropic. Claude sonnet 4.5.https : / / www . anthropic.com/claude/sonnet, 2025. Model page; accessed 2025-11-13. 7

2025
[2]

Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923, 2025. 7, 1

work page internal anchor Pith review Pith/arXiv arXiv 2025
[3]

H-optimus-0: the world’s largest open-source ai foundation model for pathology, 2024

Bioptimus. H-optimus-0: the world’s largest open-source ai foundation model for pathology, 2024. 3

2024
[4]

H-optimus-1: Foundation models for histology,

Bioptimus. H-optimus-1: Foundation models for histology,
[5]

Company/model announcement. 3
[6]

WsiCaption: Multiple instance generation of pathology reports for gigapixel whole-slide im- ages.arXiv:2311.16480, 2023

Pingyi Chen, Honglin Li, Chenglu Zhu, Sunyi Zheng, Zhongyi Shui, and Lin Yang. WsiCaption: Multiple instance generation of pathology reports for gigapixel whole-slide im- ages.arXiv:2311.16480, 2023. 3

work page arXiv 2023
[7]

WSI-VQA: Interpreting whole slide images by generative visual question answering

Pingyi Chen, Chenglu Zhu, Sunyi Zheng, Honglin Li, and Lin Yang. WSI-VQA: Interpreting whole slide images by generative visual question answering. InECCV, 2024. 3, 4

2024
[8]

Chen et al

Richard J. Chen et al. Towards a general-purpose foundation model for computational pathology.Nature Medicine, 2024. 3

2024
[9]

Slidechat: A large vision-language assistant for whole- slide pathology image understanding

Ying Chen, Guoan Wang, Yuanfeng Ji, Yanjun Li, Jin Ye, et al. Slidechat: A large vision-language assistant for whole- slide pathology image understanding. InCVPR, 2025. 2, 3, 4, 7

2025
[10]

Peripheral blood smear.https://my

Cleveland Clinic. Peripheral blood smear.https://my. clevelandclinic . org / health / diagnostics / 22742 - peripheral - blood - smear - test, 2022. Accessed: 2025-10-14. 1

2022
[11]

Pa-LLaV A: A large language-vision assis- tant for human pathology image understanding, 2024

Dawei Dai et al. Pa-LLaV A: A large language-vision assis- tant for human pathology image understanding, 2024. 4, 7

2024
[12]

Computa- tional analysis of peripheral blood smears detects disease- associated cytomorphologies.Nature Communications, 14 (1):4378, 2023

Jos ´e Guilherme de Almeida, Emma Gudgin, Martin Besser, William G Dunn, Jonathan Cooper, Torsten Haferlach, George S Vassiliou, and Moritz Gerstung. Computa- tional analysis of peripheral blood smears detects disease- associated cytomorphologies.Nature Communications, 14 (1):4378, 2023. 4, 1

2023
[13]

Gemini 2.5 pro.https : / / deepmind.google/models/gemini/pro/, 2025

Google DeepMind. Gemini 2.5 pro.https : / / deepmind.google/models/gemini/pro/, 2025. Model page; accessed 2025-11-13. 7

2025
[14]

Pathvqa: 30000+ questions for medical visual question answering

Xuehai He, Yichen Zhang, Luntian Mou, Eric Xing, and Pengtao Xie. Pathvqa: 30000+ questions for medical visual question answering. InAAAI Workshop on Health Intelli- gence, 2020. 2

2020
[15]

CHIEF: Clinical histopathology imaging eval- uation foundation model, 2024

HMS DBMI. CHIEF: Clinical histopathology imaging eval- uation foundation model, 2024. Code repository and model card. 4

2024
[16]

Densely connected convolutional net- works

Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kil- ian Q Weinberger. Densely connected convolutional net- works. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4700–4708, 2017. 1

2017
[17]

A visual–language foundation model for pathology image analysis.Nature Medicine, 2023

Ziyi Huang et al. A visual–language foundation model for pathology image analysis.Nature Medicine, 2023. 3

2023
[18]

GPT-4o System Card

Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perel- man, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli- hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 7

work page internal anchor Pith review Pith/arXiv arXiv 2024
[19]

Ikezogwo, Mehmet S

Wisdom O. Ikezogwo, Mehmet S. Seyfioglu, Fatemeh Ghezloo, Dylan S. C. Geva, Fatwir S. Mohammed, Pa- van K. Anand, Ranjay Krishna, and Linda Shapiro. Quilt- 1M: One million image-text pairs for histopathology. arXiv:2306.11207, 2023. 3

work page arXiv 2023
[20]

Perceiver: General perception with iterative attention

Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. InInternational confer- ence on machine learning, pages 4651–4664. PMLR, 2021. 6

2021
[21]

TCGA-Reports: A machine-readable pathol- ogy reports dataset for the cancer genome atlas.Scientific Data, 2024

J Kefeli et al. TCGA-Reports: A machine-readable pathol- ogy reports dataset for the cancer genome atlas.Scientific Data, 2024. 3

2024
[22]

Dinobloom: A foundation model for generalizable cell embeddings in hematology

Vincent Koch et al. Dinobloom: A foundation model for generalizable cell embeddings in hematology. InMICCAI,
[23]

Building and better understanding vision-language models: insights and future directions

Hugo Laurenc ¸on, Andr´es Marafioti, Victor Sanh, and L ´eo Tronchon. Building and better understanding vision- language models: insights and future directions.arXiv preprint arXiv:2408.12637, 2024. 7

work page arXiv 2024
[24]

Llava-med: Training a large language- and-vision assistant for biomedicine in one day.Advances in Neural Information Processing Systems, 36:28541–28564,

Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language- and-vision assistant for biomedicine in one day.Advances in Neural Information Processing Systems, 36:28541–28564,
[25]

Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InIn- ternational conference on machine learning, pages 19730– 19742. PMLR, 2023. 6, 7

2023
[26]

WSI-LLaV A: A multimodal large lan- guage model for whole slide image.arXiv:2412.02141,

Yuci Liang, Xinheng Lyu, Meidan Ding, Wenting Chen, Jipeng Zhang, et al. WSI-LLaV A: A multimodal large lan- guage model for whole slide image.arXiv:2412.02141,

work page arXiv
[27]

Rouge: A package for automatic evaluation of summaries

Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. InText summarization branches out, pages 74–81, 2004. 6

2004
[28]

Llavanext: Im- proved reasoning, ocr, and world knowledge.https: / / llava - vl

Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llavanext: Im- proved reasoning, ocr, and world knowledge.https: / / llava - vl . github . io / blog / 2024 - 01 - 30 - llava-next/, 2024. 7

2024
[29]

DeepSeek-VL: Towards Real-World Vision-Language Understanding

Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision- language understanding.arXiv preprint arXiv:2403.05525,

work page internal anchor Pith review arXiv
[30]

Lu et al

Ming Y . Lu et al. A visual-language foundation model for computational pathology.Nature Medicine, 2024. 4

2024
[31]

Lu et al

Ming Y . Lu et al. TITAN: A multimodal whole-slide foun- dation model for pathology.Nature Medicine, 2025. 4, 7

2025
[32]

A single-cell morphological dataset of leukocytes from aml patients and non-malignant controls,

Christian Matek, Simone Schwarz, Carsten Marr, and Karsten Spiekermann. A single-cell morphological dataset of leukocytes from aml patients and non-malignant controls,
[33]

Med-flamingo: a multimodal medical few-shot learner

Michael Moor, Qian Huang, Shirley Wu, Michihiro Ya- sunaga, Yash Dalmia, Jure Leskovec, Cyril Zakka, Ed- uardo Pontes Reis, and Pranav Rajpurkar. Med-flamingo: a multimodal medical few-shot learner. InMachine Learning for Health (ML4H), pages 353–367. PMLR, 2023. 7

2023
[34]

Cellpose-sam: Superhuman generalization for cellular seg- mentation.bioRxiv, 2025

Marius Pachitariu, Michael Rariden, and Carsen Stringer. Cellpose-sam: Superhuman generalization for cellular seg- mentation.bioRxiv, 2025. 4

2025
[35]

Bleu: a method for automatic evaluation of machine translation

Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. InProceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318,
[36]

Sentence-bert: Sentence embeddings using siamese bert-networks

Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. InProceedings of the 2019 Conference on Empirical Methods in Natural Lan- guage Processing. Association for Computational Linguis- tics, 2019. 6

2019
[37]

Au- tomatic recognition of five types of white blood cells in pe- ripheral blood.Computerized Medical Imaging and Graph- ics, 35(4):333–343, 2011

Seyed Hamid Rezatofighi and Hamid Soltanian-Zadeh. Au- tomatic recognition of five types of white blood cells in pe- ripheral blood.Computerized Medical Imaging and Graph- ics, 35(4):333–343, 2011. 5

2011
[38]

Quilt-LLaV A: Visual in- struction tuning by extracting localized narratives from open- source histopathology videos, 2023

Mehmet Saygin Seyfioglu et al. Quilt-LLaV A: Visual in- struction tuning by extracting localized narratives from open- source histopathology videos, 2023. 2, 4, 7

2023
[39]

PRISM: A multi-modal generative foundation model for slide-level histopathology, 2024

George Shaikovski et al. PRISM: A multi-modal generative foundation model for slide-level histopathology, 2024. 4

2024
[40]

Acute promyelocytic leukemia (apl) peripheral blood smear images.https : / / www.kaggle.com/datasets/eugeneshenderov/ acute - promyelocytic - leukemia - apl, 2020

Eugene Shenderov. Acute promyelocytic leukemia (apl) peripheral blood smear images.https : / / www.kaggle.com/datasets/eugeneshenderov/ acute - promyelocytic - leukemia - apl, 2020. Dataset on Kaggle; accessed 2025-11-10. 5, 1, 2

2020
[41]

PathMMU: A massive multimodal expert-level benchmark for understanding and reasoning in pathology

Yuxuan Sun, Hao Wu, Chenglu Zhu, Sunyi Zheng, et al. PathMMU: A massive multimodal expert-level benchmark for understanding and reasoning in pathology. arXiv:2401.16355, 2024. 3

work page arXiv 2024
[42]

Yuxuan Sun, Chenglu Zhu, Sunyi Zheng, Kai Zhang, Lin Sun, Zhongyi Shui, Yunlong Zhang, Honglin Li, and Lin Yang. PathAsst: A Generative Foundation AI Assistant to- wards Artificial General Intelligence of Pathology.Proceed- ings of the AAAI Conference on Artificial Intelligence, 38(5): 5034–5042, 2024. 4

2024
[43]

Cpath-omni: A unified multimodal foundation model for patch and whole slide image analysis in computa- tional pathology

Yuxuan Sun, Yixuan Si, Chenglu Zhu, Xuan Gong, Kai Zhang, Pingyi Chen, Ye Zhang, Zhongyi Shui, Tao Lin, and Lin Yang. Cpath-omni: A unified multimodal foundation model for patch and whole slide image analysis in computa- tional pathology. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 10360–10371, 2025. 4

2025
[44]

Pathgen-1.6m: 1.6 million pathology image-text pairs generation through multi-agent collabora- tion, 2024

Yuxuan Sun et al. Pathgen-1.6m: 1.6 million pathology image-text pairs generation through multi-agent collabora- tion, 2024. 4

2024
[45]

The cancer genome atlas (tcga) program.https : / / www

The Cancer Genome Atlas (TCGA) Research Network. The cancer genome atlas (tcga) program.https : / / www . cancer . gov / ccg / research / genome - sequencing/tcga, 2025. National Cancer Institute; ac- cessed 2025-11-13. 2

2025
[46]

Generating dermatopathology reports from gigapixel whole-slide images.Nature Communica- tions, 2025

Minh Tran et al. Generating dermatopathology reports from gigapixel whole-slide images.Nature Communica- tions, 2025. 4, 7

2025
[47]

A foundation model for clinical- grade computational pathology.Nature Medicine, 2024

Eugene V orontsov et al. A foundation model for clinical- grade computational pathology.Nature Medicine, 2024. 3

2024
[48]

Evaluating open-QA eval- uation

Cunxiang Wang, Sirui Cheng, Qipeng Guo, Yuanhao Yue, Bowen Ding, Zhikun Xu, Yidong Wang, Xiangkun Hu, Zheng Zhang, and Yue Zhang. Evaluating open-QA eval- uation. InThirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6

2023
[49]

Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024. 7

work page internal anchor Pith review Pith/arXiv arXiv 2024
[50]

InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, et al. Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265, 2025. 7

work page internal anchor Pith review arXiv 2025
[51]

Wbc lisc dataset.https : / / universe

WBCs. Wbc lisc dataset.https : / / universe . roboflow.com/wbcs/wbc- lisc, 2022. visited on 2025-11-10. 5, 1, 2

2022
[52]

A vision–language foundation model for precision oncology.Nature Cancer, 2025

Jinxi Xiang et al. A vision–language foundation model for precision oncology.Nature Cancer, 2025. 4

2025
[53]

A whole-slide foundation model for digital pathology from real-world data.Nature, 2024

Han Xu et al. A whole-slide foundation model for digital pathology from real-world data.Nature, 2024. 3

2024
[54]

virchow2: scaling self-supervised mixed magnification pretraining for computational pathol- ogy, 2024

Eric Zimmermann et al. virchow2: scaling self-supervised mixed magnification pretraining for computational pathol- ogy, 2024. 3 PBSBench: A Multi-Level Vision-Language Framework and Benchmark for Hematopathology Whole Slide Image Interpretation Supplementary Material

2024
[55]

4, we present more detailed statistics on the number of questions in our datasets, with a breakdown by question type

Dataset Statistics Here in Tab. 4, we present more detailed statistics on the number of questions in our datasets, with a breakdown by question type. Note that we remove questions with high similarity or trivial answers, leading to some small sub- groups
[56]

We also publish PBSInstr,PBSBench, and our evaluation toolkit together with the training code

Code and Data Availability We publish our code for trainingPBS-VL2. We also publish PBSInstr,PBSBench, and our evaluation toolkit together with the training code
[57]

Ethical Statement Our proposed datasetsPBSInstrandPBSBenchare pri- marily built on a publicly available PBS dataset [11] and several blood cell datasets [31, 39, 50]

Ethical, Limitation, and Hallucination Statements 8.1. Ethical Statement Our proposed datasetsPBSInstrandPBSBenchare pri- marily built on a publicly available PBS dataset [11] and several blood cell datasets [31, 39, 50]. They are thoroughly de-identified and reveal no personal information. PBSInstr,PBSBench, andPBS-VLare released solely for research and ...
[58]

Oth- ers

Additional Data Processing Details 9.1. Patch Quality Control We apply the tile quality control model from [11] to identify and remove patches with extremely low or high cell con- centrations, as well as patches dominated by artifacts. Fol- lowing [11], we extract patches of size512×512at40× magnification and reuse their released DenseNet121-based quality...
[59]

Prompts As we scale up our annotation using GPT-4o, we present the prompt used for various tasks in Fig. 7, Fig. 8, Fig. 9, and Fig. 10. Note that we remove some prompt sections on input/output formatting and examples for simplicity
[60]

Breakdown Performance on PBSBench In this section, we report a detailed breakdown of bench- mark performance by task and question types

Additional Experimental Results 11.1. Breakdown Performance on PBSBench In this section, we report a detailed breakdown of bench- mark performance by task and question types. They can be found in Tab. 8, Tab. 9, Tab. 10, Tab. 11, and Tab. 12. For cell-level questions, we report only one metric per ques- tion type for simplicity: accuracy for True-or-False...
[61]

this cell

Additional Case Study We provide an additional case study in Fig. 6 that illus- trates how model performance degrades when explicit cell type information is removed from the question. In this case study, we modify the original questions by replacing spe- cific cell types (e.g. monocyte) with a generic reference (“this cell”). Thus, we eliminate hints that...
[62]

Figure 7

Large Tables and Figures See below. Figure 7. Prompts for cell crop captioning Table 7. The mapping of cell type from out-of-distribution cell image datasets for normalization. AML-Cytomorphology LMU APL-kaggle fine-coarsed types normalized types fine-coarsed types normalized types BAS Basophil Artifact Others EBO Others Band neutrophils Neutrophil EOS Eo...