PathView-Bench: Can Multimodal Large Language Models Achieve Fine-grained Multiscale Understanding of Pathology Images?

PathView-Bench: Can Multimodal Large Language Models Achieve Fine-grained Multiscale Understanding of Pathology Images?

arXiv:2607.28318v1 [cs.AI] 30 Jul 2026

Zongyi Chen National Institute for Data Science in Health and Medicine, Xiamen University Xiamen, China Jie Lin Department of Computer Science School of Informatics Xiamen University, Xiamen, China

Yu Liang Institute of Artificial Intelligence Xiamen University, Xiamen, China Liansheng Wang Department of Computer Science School of Informatics Xiamen University, Xiamen, China

Abstract Multimodal large language models (MLLMs) are increasingly used to analyze pathology images. However, dominant multimodal benchmarks in pathology mainly score final diagnostic answers, captions, or reports. These evaluations provide limited insight into whether a model understands the multiscale visual content needed for pathology reasoning and decision-making. We introduce PathViewBench, a vision-anchored benchmark for fine-grained and multiscale visual understanding in computational pathology. Built from 23 public pathology imaging datasets with human-supervised labels and spatial annotations, PathView-Bench evaluates MLLM understanding in two fields of view: Region-FOV for highresolution local regions and Slide-FOV for macro whole-slide views. By converting raw annotations into deterministic task targets, PathView-Bench enables programmatic scoring of region localization, visual recognition, quantity estimation, spatial reasoning, and insufficient-context judgment. The benchmark contains 14 VQA-style tasks, 61, 673 images, and 308, 070 samples across 28 organs and 7, 253, 526 annotations. Evaluating 18 representative general-purpose, medicaldomain, and pathology-oriented MLLMs, we observe substantial limitations even in advanced models on fine-grained visual tasks across multiscale pathology images. PathView-Bench provides a reproducible basis for developing and evaluating pathology MLLMs with explicit multiscale visual understanding.

1

Introduction

Advances in digital pathology have enabled the digitization and computational analysis of histopathological slides, creating promising opportunities for AI-assisted diagnostics [1–10]. Recently, multimodal large language models (MLLMs) are increasingly being used for whole-slide analysis. By combining visual understanding with reasoning, these approaches are moving beyond single-task classifiers toward more general and fine-grained pathology assistants [11], including visual retrieval [12– 14], report-oriented interaction [15–18], multiscale exploration [19–21], navigation [22–24], and cross-FOV visual reasoning [25–28]. Existing pathology multimodal datasets and benchmarks mainly assess whether MLLMs can achieve diagnostic correctness [29–34], generate high-quality captions [35, 36], or produce plausible reports across pathology tasks [20, 19, 22]. This text-driven question-answering evaluation paradigm leaves a key question largely unexamined: whether a correct diagnostic answer is supported by finegrained visual understanding. More concerningly, models might score highly on medical multimodal benchmarks even without image input [37, 30]. This uncertainty makes it difficult to assess whether

Pathology Images Corpus Slide FOV

Region FOV

Region Classification GT malignant lesion

Object Counting GT 102

Existing Benchmark Q What was the grade of the invasive ductal carcinoma ?

Grade 3 Q What is the histological classification of the primary neoplasm observed in the provided whole slide image? Oligoastrocytoma,astrocytoma predominant Q What is the size of the carcinoma in the breast? 1.6 x 1.5 x 1.3 cm Q Draft a formal pathology report for the findings: The pathology slide describes a left adrenal gland biopsy taken for Conn's syndrome and genetic studies. The specimen showed a primarily trabecular adrenal tumor with a mitotic rate of less than 6 per 50 hpf and cytological atypia. The tumor is consistent with an adrenal cortical neoplasm of indeterminate malignant potential...

Text-driven Question-Answering

PathView-Bench

non-cancer tissue

Region Description lymphocyte,

GT fpibroblast, lasma cell

24

Lesion Grounding

Density Localization GT

GT Tumor non-cancer

Spatial Relation GT

Density Localization GT

Organ Classification GT Lung

Spatial Relation GT

Area Estimation GT 25%-50%

1

75%-100%

4

tumor cell, stroma

Colon

Object Classification GT epithelial cell inflammatory cell

Object Grounding GT

FOV Trap GT NA Stomach

Region FOV

FOV Trap GT NA 50 cells

Slide FOV

Performance Rank Better in Slide-FOV

Better in both

8 12

General MLLM Medical MLLM Pathology MLLM

16

Better in Region-FOVr

16

12

8

4

1

A vision-anchored benchmark for fine-grained and multiscale visual understanding

Figure 1: Motivating examples for PathView-Bench. The benchmark evaluates visual understanding across local-region and whole-slide fields of view through vision-anchored pathology tasks.

the diagnostic output is trustworthy. Moreover, pathology images are extremely large, and local details can differ substantially from global semantic context, yet existing benchmarks often remain limited to a single field of view and therefore provide limited evidence about whether models can understand pathology images across scales. As illustrated in Figure 1, these limitations leave a critical gap for advancing pathology MLLMs: How can we evaluate their ability to understand fine-grained visual content in multiscale pathology images during reasoning and decision-making? To address this gap, a benchmark is needed to make the visual basis of pathology reasoning measurable. We therefore introduce PathView-Bench, a vision-anchored benchmark for fine-grained and multiscale visual understanding in computational pathology. PathView-Bench is organized around two fields of view: Region-FOV for high-resolution local regions and Slide-FOV for macro whole-slide views, reflecting the fact that pathology reasoning depends on both cellular morphology and tissue organization. Across these views, PathView-Bench probes evidence-grounded visual understanding through tasks such as region localization, visual recognition, quantity estimation, spatial reasoning, and judging whether the current view contains sufficient evidence to answer. This shifts evaluation away from final-response plausibility and toward evidence grounded in the multiscale visual context. To construct the benchmark, we map 23 publicly available pathology imaging datasets with humansupervised labels and spatial annotations into a shared, auditable benchmark space, where deterministic rules convert the annotations into task targets for programmatic scoring. The resulting benchmark contains 14 VQA-style tasks, 61, 673 images, and 308, 070 samples across 28 organs and 7, 253, 526 annotations. Across 18 representative general-purpose, medical-domain, and pathology-oriented MLLMs, PathView-Bench reveals a recurring gap: models are stronger at naming visual categories than at grounding, counting, spatially organizing, or judging the sufficiency of the underlying evidence. Matched-slide analyses further show that correct slide-level diagnostic answers can mask failures in these basic visual operations, and neither larger scale nor pathology-specific training consistently closes the gap. In summary, the contributions of this work are as follows: 1. We introduce a vision-anchored benchmark for fine-grained and multiscale visual understanding in pathology, organized around Region-FOV/Slide-FOV and designed to evaluate whether model responses are grounded in observable image content. 2

Table 1: Comparison of pathology VQA-style multimodal benchmarks. All-WSI Source indicates whether all images are derived from WSIs. Annotation: LLM = language-model-assisted generation, Exp. = expert annotation or validation, Rule = fixed-rule conversion. Benchmark PathVQA [29] PathMMU [30] WSI-VQA [31] Micro-Bench [32] WSI-Bench [20] PathBench [33] SlideBench [19] PathVG [34] MultiPathQA [22] PathView (ours)

#Samples

#Images

#Task Types

Annotation Method

All-WSI Source

Vision-anchored

Region-FOV

Slide-FOV

32,799 33,428 8,672 8,485 179,569 33,235 15,835 33,500 934 308,070

4,998 24,067 977 8,385 9,850 12,060 1,792 27,610 868 61,673

7 1 1 3 4 2 2 1 1 14

Exp.+Rule LLM+Rule+Exp. LLM+Rule LLM+Exp. LLM+Exp. LLM+Exp. LLM+Rule+Exp. LLM+Exp. Exp.+Rule Exp.+Rule

× × ✓ × ✓ × ✓ ✓ ✓ ✓

× × × ✓ × × × ✓ × ✓

✓ ✓ × ✓ × ✓ × ✓ × ✓

× × ✓ × ✓ ✓ ✓ × ✓ ✓

2. We define an auditable construction and evaluation protocol that standardizes human-supervised labels and spatial annotations into deterministic task targets with programmatic scoring, without LLM-as-a-judge evaluation. 3. We conduct comprehensive evaluations of 18 representative general-purpose, medical-domain, and pathology-oriented MLLMs on PathView-Bench, characterizing their capability boundaries in fine-grained multiscale visual understanding. We further include a PathView-tuned baseline to test whether supervision provides learnable signals for grounding-sensitive behavior.

2

Related Work

Multimodal Large Language Models. Recent general-purpose MLLMs have broadly advanced multimodal perception, interaction, and reasoning [38–40]. This progress has been adapted to medicine through medical image-text pretraining [41, 42], broad biomedical supervision [43], and domain-specific instruction tuning or medical reasoning alignment [44, 45]. In pathology, the same trajectory has moved from patch-level dialogue and visual instruction tuning [11–13], to report generation [15–17], and then toward whole-slide interaction, WSI-level modeling, navigation, and evidence-seeking reasoning [19, 20, 22, 25–27]. These models motivate a stricter evaluation regime: once systems claim richer whole-slide behaviors, they should be judged not only by linguistic fluency but also by whether their outputs remain grounded in visible pathology content. Pathology MLLM Benchmarks. Pathology multimodal evaluation has progressed from patch image question answering and image-text resources to WSI-level interaction. PathVQA [29] established pathology VQA on patch images, while Quilt-1M [35] and PathGen-1.6M [36] provided large-scale histopathology image-text pairs for representation learning. Later resources broadened the evaluation scope: PathMMU [30] and Micro-Bench [32] emphasized expert-level or microscopy-oriented local understanding, WSI-VQA [31], WSI-Bench [20], SlideBench [19], and MultiPathQA [22] moved toward whole-slide QA and navigation, and PathBench [33] covered both patch and WSI settings. PathVG [34] is closer to explicit visual grounding, but remains focused on region-level localization. Overall as shown in Table 1, existing resources either emphasize local morphology without global slide context, or evaluate WSI-level answers without requiring grounded intermediate visual analysis. PathView-Bench is designed to cover both Region-FOV and Slide-FOV while making visual grounding and fine-grained visual-understanding operations central to the evaluation.

3

PathView-Bench

In this section, we describe the design and construction of PathView-Bench, including its dual-scale task organization, annotation-grounded protocol, and deterministic split and evaluation procedure. 3.1

Task Design

WSIs are gigapixel-scale images whose diagnostically relevant information appears at different visual scales [46]. In routine pathology review, pathologists assess whole-slide tissue architecture and lesion distribution, while also examining selected high-resolution regions that contain essential morphological details [23]. PathView-Bench follows this dual-scale reading process by organizing its 3

Table 2: Overview of the diverse task types, query and answer examples of PathView-Bench Task name

Query example

Answer example Region FOV

Object Grounding Object Classification Region Classification Object Counting Region Description Density Localization Spatial Relation FOV Trap Organ Classification Lesion Grounding Object Classification Area Estimation Density Localization FOV Trap

Locate all <object>. Which category best describes the predominant finding in this region <bbox>? Which category best describes the predominant finding in this image? How many <object> are present in this image? Describe what is in the image. Which quadrant contains the densest distribution of <object>? Local coordinate <bbox> marks a core region. Which category is nearest to this region? Which organ is this image most likely from? Slide FOV Which organ is this image most likely from? Locate all regions of <object>. Which category best describes the predominant finding in this region <bbox>? What percentage of the tissue area in the image is occupied by <object>? Which quadrant contains the densest distribution of <object>? How many <cell> are present in this image?

[0, 0, 8, 12], ... A. plasma cell A. tumor cell 35 eosinophil, ... Bottom-right quadrant B. tumor cell Insufficient information B. spleen [50, 138, 87, 227], ... A. malignant lesion A. 0%-25% A. Top-left quadrant Insufficient information

VQA-style tasks into two fields of view: Slide-FOV tasks for macro whole-slide understanding and Region-FOV tasks for fine-grained local-region understanding, as shown in Table 2. 3.1.1

Slide-FOV Whole-Slide Understanding

Slide-FOV tasks evaluate visual understanding from macro whole-slide views. They focus on information that can be assessed at low magnification, including tissue context, lesion location, abnormal burden, and large-scale spatial distribution. These tasks test whether a model can interpret slide-level visual patterns and recognize when the current view is sufficient for the requested inference. Organ Classification. Identifying the organ or tissue source is a basic prerequisite for interpreting a macro view with representative tissue morphology. In this task, the model is asked to infer the most likely organ or tissue source from the selected whole-slide view. Object Classification. Object Classification evaluates whether the model can interpret a specified macro-scale object or ROI. The model is provided with a region on the macro image and asked to classify the predominant pathological or tissue category inside the marked area. Candidate objects are constructed from projected standard annotations. Lesion Grounding. Whole-slide analysis often starts by locating suspicious regions before local inspection. Lesion Grounding asks the model to identify all macro-view regions matching a target lesion or abnormal structure. Multiple independent lesions on the same slide are kept separate to preserve multifocal pathology. Area Estimation. Lesion burden is routinely assessed when estimating tumour extent, necrosis, or other abnormal tissue components. Area Estimation evaluates this ability by asking the model to select the interval that best matches the fraction of visible effective tissue occupied by a target lesion or structure. We compute the fraction over the effective tissue area in the macro image and use four discretized intervals as answer options. Density Localization. Estimating the density of a lesion helps distinguish focal, clustered, and regionally concentrated abnormalities. Density Localization divides the macro image into four quadrants and asks the model to identify the quadrant with the densest target distribution. The target quadrant is determined from both the number of target regions and their occupied area. This group further includes a Slide-FOV trap, where the model receives a macro WSI view but is asked to count cell-level objects that are not visible at that scale. The correct response is the insufficiency option, not an unsupported count. 3.1.2

Region-FOV Fine-Grained Histological Understanding

Region-FOV tasks evaluate fine-grained visual understanding at the microscopic scale, where relevant visual content appears as cellular morphology, tissue microstructures, annotated object instances, and 4

local spatial relationships. These tasks test whether a model can interpret the histological entities visible in a high-resolution field, rather than relying on slide-level recognition alone. Region Classification. At high magnification, local interpretation often begins with identifying the dominant cellular or tissue component in the field. Region Classification evaluates this ability by asking the model to assign the image to its primary morphology category. Region Description. Pathology image understanding also requires concise descriptions of the components visible in a local field. Region Description asks the model to describe the main objects or structures in the image using standardized terminology. This task does not require a full diagnostic report; instead, it tests whether the model can summarize the key histological components in a local view in a reproducible form. Object Classification. A local histology field can contain multiple annotated objects, including different cell types, tissue structures, or pathological components. Object Classification evaluates whether the model can assign the correct category to a specified instance, rather than only describe the image at the field level. Object Grounding. High-magnification interpretation often depends on accurately localizing key cells, glandular structures, necrotic regions, and other morphology-defined entities. Object Grounding evaluates whether the model can localize these fine-grained pathological entities in a high-resolution local image. Object Counting. Counting cells or small structures is central to many microscopic pathology assessments, including mitotic counts, inflammatory-cell evaluation, positive-cell estimation, and grading or scoring workflows. Object Counting asks the model to report the number of visible target-category instances in a local image, testing whether it can make precise quantitative judgments in dense fields with visually similar or partially overlapping objects. Density Localization. The spatial arrangement of cells or structures in a local field can indicate clustering or heterogeneity. Density Localization asks the model to identify the quadrant with the highest density of the target category. Spatial Relation. Local interpretation can also depend on proximity relationships between histological entities at a comparable scale. Spatial Relation asks the model to identify which candidate category is closest to a specified target instance, using annotated local geometry as supervision. This group also includes a Region-FOV trap, where the model receives a high-magnification crop with organ-agnostic morphology, such as malignant cells that can appear similar across organs, but is asked to infer the source organ. Because this information is not reliably available from the local crop alone, the correct response is the insufficiency option rather than an unsupported slide-level inference. 3.2

Standardized Protocol

To keep benchmark supervision tied to reliable visual content, we treat benchmark construction as an auditable conversion problem rather than automatic data repackaging or model-generated relabeling. Our standardized protocol maps public pathology resources with human-supervised labels and annotations into task targets only when the resulting target is label-valid, observable, and answerable. The protocol consists of four stages: data collection, data standardization and quality control, task construction, and dataset splitting and evaluation. Figure 2 summarizes the full protocol, and Figure 3 reports the resulting dataset statistics. 3.2.1

Data collection

We screened public sources, including Zenodo, Kaggle, GitHub, and dataset challenge websites, and retained pathology datasets containing real images with human-supervised labels or annotations. This yielded a source corpus of 23 datasets, comprising 63, 452 image records and 6, 019, 032 raw annotation entries. Patch-level and annotated-region datasets mainly support Region-FOV tasks, whereas whole-slide datasets support Slide-FOV tasks. 5

1 Data Collection Public Pathology

Resources

2 Standardization & QC Source

corpus

Source

Corpus

Collection resources

3 Task Construction Standardized

corpus

Standardized

corpus

Normalize labels, annotations

Kaggle

Github

Manual Review question templates, answer spaces

119 raw labels -> 44 standard labels

Others

Multi-scale

Annotation conversion

Classification

26 formats -> bounding boxes

Region tasks

Multi-organ Orgin WSI-source Clinical labels

Region-FOV

Rule-based filter

Slide-FOV

remove unresolved, background labels

Spatial reasoning FOV-trap

14 VQA-style tasks

308,070 samples

61673 images | 44 labels

7,253,526 bboxes

23 datasets | 63,452 images

28 organs | 6,019,032 annotations

Grounding Counting

Quality control

Human annotations

files

Construct VQA-style tasks

Label mapping Zenodo

Task

4 Splitting & Evaluation Task

Released

Benchmark

files

Release splits and metrics Group-wise leakage control dataset split by tasks Train 70%

Val

10%

Test

20%

Input harmonization Raw images, processed features

Programmatic scoring Dice | Accuracy | F1 | MAE

Entity recall | SRR Released benchmark

Evaluation protocol

Object tion Classifica [54,693]

i 22 on-F 0,1 O 06 V

O Groubject n [56,21ding 1]

t jec on Ob ificati] ss 10 Cla [26,8 Spatial Relation [11,565]

De Loca nsity liz [17,50ation 4]

t jec ing Ob unt 27] Co 41,9 [



Organ Classification [500]

%RXQGLQJER[HVSHULPDJH

Sli 87 de-F ,96 OV 4

PathView

Bench Re g 308,070

Region-FOV Trap [475]

1XPEHURILPDJHV

s [14 criptin ,51 on 7]



r s [16oundion ,88 ing 0]

V Trap Slide-FO [22,249]

Area Estim [14, ation 006] G Le

n n gio tio Re ifica 4] ss ,21 Cla [23 De Regio

Density Localizatio [7,519] n

Figure 2: Overview of the PathView-Bench standardized protocol pipeline. 5HJLRQ)29 6OLGH)29

N

 

PHDQ PHGLDQ

N



              !

4XHVWLRQVSHULPDJH

  

$OOLPDJHV

Figure 3: Dataset statistics for PathView-Bench. The four panels summarize the distribution of VQA-style tasks, the number of questions per image, the distribution of spatial annotations per image, and the standardized pathology label vocabulary used to construct the benchmark. 3.2.2

Data Standardization and Quality Control

The source corpus differed substantially in label vocabularies and annotation formats. For each dataset, experts reviewed the mapping from raw pathology labels to a shared benchmark vocabulary. We then implemented dataset-specific scripts to convert native spatial annotations into bounding boxes and normalized the resulting boxes to a common coordinate system. Overall, this process consolidated 119 raw annotation labels into 44 standard labels and converted 26 source-specific annotation formats into bounding boxes. Using predefined thresholds and decision rules, we filtered out records with unresolved label mappings, non-pathological background categories, or spatial annotations that were invalid, out of bounds, excessively overlapping, or zero-area. The resulting corpus contains task-ready labels and annotations for downstream task construction. 3.2.3

Task Construction

We first defined a standard template for each task, specifying the query format, answer format, candidate answer space, and required supervision. Using these templates, we generated task instances from the standardized corpus according to each dataset’s field-of-view scale and available annotations. During generation, we removed samples whose answers could not be uniquely determined, including borderline cases at answer-option thresholds, targets that were invalid or not observable at the current field of view, and cases for which the predefined area, count, or distance criteria could not be computed. We manually reviewed representative converted samples for label consistency, visual observability, and deterministic answerability. 3.2.4

Dataset Splitting and Evaluation

We merged the constructed task files by task type while retaining sample-level provenance, including the source dataset, visual input path, task metadata, and split assignment. PathView-Bench uses training, validation, and test splits in a 7:1:2 ratio. To reduce leakage, samples from each source dataset are split using the most specific grouping key available. We provide standard image inputs for image-based models and model-specific feature files for WSI models [19, 20] that use feature-based inference. Model outputs are scored with task-native deterministic metrics rather than LLM-as-judge 6

Table 3: Main benchmark results on PathView-Bench. Region-FOV columns: OGD = Object Grounding, RCL = Region Classification, OCL = Object Classification, OCN = Object Counting, RDS = Region Description, DLO = Density Localization, SPR = Spatial Relation. Slide-FOV columns: ORG = Organ Classification, LGD = Lesion Grounding, OCL = Object Classification, ARE = Area Estimation, DLO = Density Localization. OCN is lower better, all other columns are higher better. Best results are bolded and second-best results are underlined. RDS is not assigned to random/frequency baselines because entity recall is not meaningful for frequent-choice description. Region-FOV

Slide-FOV

Model

Params

OGD

RCL

OCL

OCN

RDS

DLO

SPR

ORG

LGD

OCL

ARE

DLO

Control baselines Random Choice Frequent Choice

– –

0.073 0.000

0.189 0.247

0.175 0.213

482.21 22.23

– –

0.179 0.102

0.201 0.260

0.280 0.340

0.151 0.005

0.311 0.465

0.181 0.103

0.179 0.105

General-purpose MLLMs Gemini-2.5-Flash – GPT-5-mini – InternVL3 8B InternVL3.5 8B Qwen3-VL 8B Qwen3-VL 30B GLM-4.6V-Flash 10B

0.170 0.185 0.052 0.189 0.011 0.164 0.201

0.559 0.598 0.405 0.475 0.349 0.578 0.348

0.302 0.290 0.281 0.303 0.216 0.287 0.275

83.08 16.82 28.59 19.44 21.47 19.41 22.95

0.224 0.011 0.006 0.077 0.011 0.061 0.006

0.289 0.311 0.189 0.310 0.171 0.272 0.250

0.362 0.413 0.350 0.398 0.297 0.302 0.264

0.940 0.820 0.540 0.740 0.580 0.800 0.530

0.443 0.437 0.399 0.310 0.146 0.383 0.552

0.548 0.457 0.521 0.593 0.391 0.600 0.548

0.250 0.367 0.142 0.215 0.201 0.241 0.219

0.449 0.417 0.336 0.475 0.312 0.415 0.447

Medical-domain MLLMs LLaVA-Med v1.5 7B HuatuoGPT-V 7B HuatuoGPT-V 34B Lingshu 7B Lingshu 32B Med Gemma 27B Med Gemma1.5 4B

0.020 0.066 0.076 0.080 0.095 0.129 0.081

0.210 0.352 0.549 0.601 0.636 0.520 0.486

0.001 0.190 0.246 0.270 0.295 0.275 0.255

32.80 20.17 37.73 19.60 18.48 21.57 31.73

0.292 0.147 0.048 0.149 0.102 0.006 0.009

0.017 0.230 0.162 0.200 0.229 0.255 0.178

0.010 0.327 0.317 0.351 0.339 0.319 0.303

0.080 0.630 0.700 0.900 0.910 0.490 0.370

0.011 0.151 0.196 0.249 0.333 0.305 0.218

0.365 0.603 0.532 0.549 0.557 0.419 0.439

0.129 0.246 0.269 0.338 0.125 0.448 0.211

0.084 0.409 0.238 0.396 0.301 0.360 0.185

Pathology-oriented MLLMs Quilt-LLaVA 7B Patho-R1 7B SlideChat ∼7B WSI-LLaVA ∼7B

0.067 0.065 0.080 0.018

0.437 0.556 0.000 0.193

0.249 0.294 0.000 0.064

59.59 48.11 24.07 72.36

0.076 0.005 0.000 0.316

0.102 0.207 0.000 0.082

0.366 0.283 0.000 0.210

0.530 0.900 0.310 0.510

0.107 0.015 0.071 0.030

0.393 0.489 0.485 0.398

0.118 0.310 0.164 0.002

0.092 0.323 0.160 0.020

PathView-tuned baseline Qwen3-VL 8B

0.323

0.829

0.305

8.32

0.790

0.631

0.673

0.920

0.617

0.741

0.693

0.783

evaluation, including Dice for grounding, accuracy for closed-set classification and spatial relation, F1 for interval or quadrant reasoning, MAE for counting, entity recall for description, and SRR for FOV-trap diagnostics.

4

Experiment

4.1

Models

We benchmark a diverse collection of MLLMs on PathView-Bench, including 1) general-purpose models with broad visual-language capabilities, 2) medical-domain models adapted with biomedical or clinical multimodal data, and 3) pathology-oriented models designed for pathology images. General-purpose MLLMs. We evaluate general-purpose MLLMs as open-domain baselines, including Gemini-2.5-Flash [40], GPT-5-mini [47], InternVL3 (8B) [48], InternVL3.5 (8B) [49], Qwen3-VL (8B and 30B) [39], and GLM-4.6V-Flash (10B) [50]. Medical-domain MLLMs. We evaluate medical-domain MLLMs adapted with biomedical or clinical multimodal data, including LLaVA-Med v1.5 (7B) [42], HuatuoGPT-V (7B and 34B) [44], Lingshu (7B and 32B) [45], Med Gemma (27B) [51], and Med Gemma1.5 (4B) [52], to test whether broad medical specialization improves pathology visual understanding. Pathology-oriented MLLMs. We further evaluate pathology-oriented models, including QuiltLLaVA (7B) [12], Patho-R1 (7B) [26], SlideChat [19], and WSI-LLaVA [20]. 7

27

0

26

1

12

1

1

2

0

0

0

7

1

0

100

54

1

5

0

17

1

2

0

0

2

0

1

0

4

6

0

100

75

30

0

7

0

1

11

0

1

0

0

100

64

0

27

0

1

2

0

1

0

3

0

0

100

11

ORG

6

6

1

3

26

8

31

0

0

3

5

1

0

15

0

0

0

29

OCL

3

3

3

3

31

6

6

1

0

2

0

0

0

3

0

0

1

31

ARE

3

9

5

4

45

7

48

0

0

1

0

73

0

11

0

1

1

100

DLO

9

12

13

0

45

3

17

0

1

0

0

52

5

4

0

0

5

94

RFT

34

7

12

2

66

74

98

0

22

17

2

66

0

29

0

0

100

26

SFT

88

96

97

62

100

98

100

0

100

92

94

100

70

8

57

18

45

99 7B

25

0

aV A

Ch

50

-L L

ide

SI W

R1 oth Pa

Sl

7B

4B

VA

1. 5

LL a iltQu

m a m

75

M

M

ed Ge

ed

Ge

Lin

m

gs

hu

m a

32

B

7B

B 34

hu gs Lin

TV

TV

GP uo at Hu

uo GP at Hu

LL

aV AM

ed

-4

.6

7B

V

B 30

M GL

3VL

en

en Qw

In

Qw

rn V

L3 .

5

3VL

8B

8B L3

-m

rn V te In

te

h las

T5

-F .5

GP

ini -2 m Ge

at

10

3

7B

69

0

27 B

2

13

7B

52

1

8B

1

1

ini

28

Refusal ratio (%)

DLO SPR

100

Region-FOV

0

4

Slide-FOV

1

Traps

RCL OCL

Figure 4: Refusal-ratio heatmap for ordinary PathView-Bench tasks and FOV-trap diagnostics. Ordinary tasks measure over-refusal; trap rows measure safe refusal under insufficient field of view. RFT: Region-FOV trap; SFT: Slide-FOV trap. 4.2

Implementation Details

We evaluate frozen models zero-shot using released checkpoints or API endpoints. For coordinatesensitive tasks, we follow each model’s documented output convention before converting predictions to the unified PathView-Bench coordinate system, e.g., Qwen3-VL [39] uses [0, 1000]-normalized coordinates. We also train a PathView-tuned baseline from Qwen3-VL-8B using only the PathViewBench training set, with LoRA in ms-swift [53] for 1 epoch, a global batch size of 16, a learning rate of 5 × 10−5 , and 4 NVIDIA A100 GPUs. 4.3

Main Results

Based on the evaluation results presented in Table 3, Figure 4, and Figure 5, there are several findings: Fine-grained multiscale pathology image understanding remains challenging for MLLMs. Our results show that current MLLMs show uneven fine-grained visual understanding across multiscale pathology images. Several frozen models perform well on closed-set recognition, including Gemini-2.5-Flash on Slide-FOV organ classification (0.940) and Lingshu 32B on Region-FOV region classification (0.636), but their performance is lower on tasks that require localization, quantification, or spatial organization. The best frozen model reaches 0.201 Dice on Region-FOV object grounding, and object counting remains difficult, with the best frozen MAE at 16.82. The PathViewtuned baseline shows that PathView-Bench-derived targets provide useful supervision signals for grounding-sensitive behavior. Domain specialization and model scale do not guarantee stronger performance. Medicaland pathology-specialized models do not consistently outperform general-purpose MLLMs on PathView-Bench, suggesting that domain-knowledge training must be balanced with basic visual grounding and spatial reasoning ability. Scaling shows a similarly uneven pattern across model families and fields of view. Separating Region-FOV and Slide-FOV results exposes these scale- and view-dependent weaknesses. For pathology-oriented models, weaker spatial scores may partly reflect slide-feature interfaces that preserve global semantics but lose pixel-level coordinate relationships. Improving coordinate-aware grounding would help extend these models to evidence navigation, lesion localization, burden estimation, and interactive review. Models show inconsistent caution under observability limits. We analyze refusal behavior because unsupported answers can be unsafe in pathology, while excessive refusal reduces utility on answerable images. As shown in Figure 4, ordinary tasks have a 12.4% average refusal rate, whereas safe-refusal rates are 30.9% for Region-FOV traps and 73.6% for Slide-FOV traps. Errors are asymmetric: SlideChat, WSI-LLaVA, and Qwen3-VL 8B over-refuse answerable tasks (50.8%, 57.2%, and 32.8%), while LLaVA-Med, Patho-R1, and MedGemma 1.5 4B often answer unsupported trap questions. The main limitation is not simply that models refuse too often or too rarely; rather, they fail to calibrate refusal to the visual evidence available at the current field of view. 8

0.60

0.70

0

0.42

0.36 0.27

PathView score

0.35

L E D O LG OC AR DL

&

&

&

& TCGA-A2-A0YE

TCGA-G3-A25S

TCGA-A2-A04Y

TCGA-AO-A1KP

WSI-VQA

WSI-VQA

WSI-VQA

WSI-VQA

which margin showed involvement by the tumor? A. medial surgical margin Pred: B B. lateral surgical margin

C. superior surgical margin

D. inferior surgical margin.

what is the recommended treatment for the patient? A. lumpectomy

Pred: C B. chemotherapy

C. total mastectomy

D. radiation therapy

what are the hormone receptors status ...? A. positive for her2

Pred: B B. positive for estrogen and progesterone

C. negative for all hormone receptors

D. positive for estrogen only

What is the histological grade of the hepatocellular carcinoma in this patient, based on differentiation? A. Grade I ...

Pred: B B. Grade II ...

C. Grade III ...

D. Grade IV ...

Pathview-Area Estimation

Pathview-Lesion Grounding

Pathview-Density Localization

Pathview-Region Classification

What percentage of the tissue area in the image is occupied by malignant lesion? A. 50%-75%

B. 25%-50%

Pred: A C. 75%-100%

Locate all regions of malignant lesion.

Which quadrant contains the densest distribution of malignant lesion?

Which category best describes the predominant finding in this region <box>(98,361),(935,969)</box>?

D. 0%-25%

E. ...not contain

sufficient information...

GT: [0.065,

0.003,

0.869,

0.656]

Pred:

GT

Dice

0.11 Pred

A. Top-left quadrant

Pred: B. Top-right quadrant

C. Bottom-left quadrant

D GT D. Bottom-right quadrant

E. ...not contain

sufficient information...

Pred

A. non-cancer tissue

Pred: A B. malignant lesion

C. ...not contain

sufficient information...

Figure 5: Left: For each task and model, we first select matched model-slide pairs where the model correctly answers an external SlideBench or WSI-VQA question, and then score the corresponding PathView-Bench visual-understanding task on the same slides. Each bar reports the macro-average across 16 open-source models. Right: shows Lingshu-32B cases in which external benchmark questions are answered correctly but matched PathView-Bench visual-understanding questions fail on the same Slide-FOV slides.

Correct high-level answers do not imply correct basic visual understanding. To test whether high-level WSI QA performance transfers to basic visual operations, we matched PathView-Bench Slide-FOV samples to slides that also appear in SlideBench [19] or WSI-VQA [31]. For each model and task, we conditioned the analysis on external success by retaining only matched slides on which the same model correctly answered at external benchmark question, and then scored the corresponding PathView-Bench task on those slides. The conditional PathView-Bench scores remained limited for tasks requiring localization or quantification: 0.274 for lesion grounding, 0.360 for area estimation, and 0.422 for density localization (Figure 5). The object-classification subset is less affected because most matched samples ask only for binary cancer-region recognition. Even after conditioning on externally correct high-level answers, the qualitative pairs in Figure 5b show failures in basic visual interpretation on the same Slide-FOV slides. Overall, this matched-slide analysis suggests that correct high-level recognition does not reliably imply fine-grained localization, burden estimation, or spatial reasoning.

5

Limitations and Future Work

PathView-Bench is built from publicly available pathology datasets and is intended as a visioncentered evaluation resource rather than a complete clinical benchmark. It tests whether pathology MLLMs can ground structured answers in visible image content, but it does not assess full diagnostic reporting quality, clinical deployment readiness, or human-pathologist equivalence. In particular, PathView-Bench does not include private longitudinal records, multi-timepoint diagnostic workflows, or follow-up information. Future extensions will incorporate clinical workflow data, guideline-driven diagnostic observations, and pathologists’ visual decision trajectories while preserving PathViewBench as an auditable benchmark for multiscale visual understanding.

6

Conclusion

We introduced PathView-Bench, a vision-centered benchmark for evaluating pathology MLLMs across region and slide fields of view. Built from standardized public datasets, it uses task-ready labels, geometry, prompts, and deterministic metrics to test localization, classification, quantification, spatial reasoning, and refusal under insufficient visual context. Across models, domain specialization, scale, and high-level WSI QA performance do not reliably translate into fine-grained visual grounding. The PathView-tuned baseline further indicates that structured PathView-Bench-derived targets provide learnable supervision signals under the benchmark protocol. These findings argue for evidencegrounded evaluation, rather than answer plausibility alone, and for models that preserve spatial evidence across fields of view. 9

References [1] Anant Madabhushi and George Lee. Image analysis and machine learning in digital pathology: Challenges and opportunities. Medical image analysis, 33:170–175, 2016. [2] Muhammad Khalid Khan Niazi, Anil V Parwani, and Metin N Gurcan. Digital pathology and artificial intelligence. The lancet oncology, 20(5):e253–e261, 2019. [3] Gabriele Campanella, Matthew G Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J Busam, Edi Brogi, Victor E Reuter, David S Klimstra, and Thomas J Fuchs. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine, 25(8):1301–1309, 2019. [4] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering, 5(6):555–570, 2021. [5] Wenhao Tang, Fengtao Zhou, Sheng Huang, Xiang Zhu, Yi Zhang, and Bo Liu. Feature re-embedding: Towards foundation model-level performance in computational pathology. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11343–11352, 2024. [6] Junyu Li, Ye Zhang, Wen Shu, Xiaobing Feng, Yingchun Wang, Pengju Yan, Xiaolin Li, Chulin Sha, and Min He. M4: Multi-proxy multi-gate mixture of experts network for multiple instance learning in histopathology image analysis. Medical Image Analysis, 103:103561, 2025. [7] Zeyu Gao, Anyu Mao, Yuxing Dong, Hannah Clayton, Jialun Wu, Jiashuai Liu, ChunBao Wang, Kai He, Tieliang Gong, Chen Li, et al. Smmile enables accurate spatial quantification in digital pathology using multiple-instance learning. Nature Cancer, pages 1–17, 2025. [8] Xiyue Wang, Junhan Zhao, Eliana Marostica, Wei Yuan, Jietian Jin, Jiayu Zhang, Ruijiang Li, Hongping Tang, Kanran Wang, Yu Li, et al. A pathology foundation model for cancer diagnosis and prognosis prediction. Nature, 634(8035):970–978, 2024. [9] Jinxi Xiang, Xiyue Wang, Xiaoming Zhang, Yinghua Xi, Feyisope Eweje, Yijiang Chen, Yuchen Li, Colin Bergstrom, Matthew Gopaulchan, Ted Kim, et al. A vision–language foundation model for precision oncology. Nature, 638(8051):769–778, 2025. [10] Jiabo Ma, Zhengrui Guo, Fengtao Zhou, Yihui Wang, Yingxue Xu, Jinbang Li, Fang Yan, Yu Cai, Zhengjie Zhu, Cheng Jin, et al. A generalizable pathology foundation model using a unified knowledge distillation pretraining framework. Nature Biomedical Engineering, pages 1–20, 2025. [11] 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 towards artificial general intelligence of pathology. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5034–5042, 2024. [12] Mehmet Saygin Seyfioglu, Wisdom O Ikezogwo, Fatemeh Ghezloo, Ranjay Krishna, and Linda Shapiro. Quilt-llava: Visual instruction tuning by extracting localized narratives from open-source histopathology videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13183–13192, 2024. [13] 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 computational pathology. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10360–10371, 2025. [14] Wenchuan Zhang, Jingru Guo, Hengzhe Zhang, Penghao Zhang, Jie Chen, Shuwan Zhang, Zhang Zhang, Yuhao Yi, and Hong Bu. Patho-agenticrag: towards multimodal agentic retrieval-augmented generation for pathology vlms via reinforcement learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 29921–29929, 2026. [15] Pingyi Chen, Honglin Li, Chenglu Zhu, Sunyi Zheng, Zhongyi Shui, and Lin Yang. Wsicaption: Multiple instance generation of pathology reports for gigapixel whole-slide images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 546–556. Springer, 2024. [16] Zhengrui Guo, Jiabo Ma, Yingxue Xu, Yihui Wang, Liansheng Wang, and Hao Chen. Histgen: Histopathology report generation via local-global feature encoding and cross-modal context interaction. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 189–199. Springer, 2024.

10

[17] Manuel Tran, Paul Schmidle, Ruifeng Ray Guo, Sophia J Wagner, Valentin Koch, Valerio Lupperger, Brenna Novotny, Dennis H Murphree, Heather D Hardway, Marina D’Amato, et al. Generating dermatopathology reports from gigapixel whole slide images with histogpt. Nature communications, 16(1): 4886, 2025. [18] Rundong Wang, Wei Ba, Ying Zhou, Yingtai Li, Bowen Liu, Baizhi Wang, Yuhao Wang, Zhidong Yang, Kun Zhang, Rui Yan, et al. Qcagent: An agentic framework for quality-controllable pathology report generation from whole slide image. arXiv preprint arXiv:2603.01647, 2026. [19] Ying Chen, Guoan Wang, Yuanfeng Ji, Yanjun Li, Jin Ye, Tianbin Li, Ming Hu, Rongshan Yu, Yu Qiao, and Junjun He. Slidechat: A large vision-language assistant for whole-slide pathology image understanding. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 5134–5143, 2025. [20] Yuci Liang, Xinheng Lyu, Wenting Chen, Meidan Ding, Jipeng Zhang, Xiangjian He, Song Wu, Xiaohan Xing, Sen Yang, Xiyue Wang, et al. Wsi-llava: A multimodal large language model for whole slide image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22718–22727, 2025. [21] Yuxuan Sun, Yixuan Si, Chenglu Zhu, Kai Zhang, Zhongyi Shui, Bowen Ding, Tao Lin, and Lin Yang. Cpathagent: An agent-based foundation model for interpretable high-resolution pathology image analysis mimicking pathologists’ diagnostic logic. arXiv preprint arXiv:2505.20510, 2025. [22] Thomas A Buckley, Kian R Weihrauch, Katherine Latham, Andrew Z Zhou, Padmini A Manrai, and Arjun K Manrai. Navigating gigapixel pathology images with large multimodal models. arXiv preprint arXiv:2511.19652, 2025. [23] Sheng Wang, Ruiming Wu, Charles Herndon, Yihang Liu, Shunsuke Koga, Jeanne Shen, and Zhi Huang. Pathology-cot: Learning visual chain-of-thought agent from expert whole slide image diagnosis behavior. arXiv preprint arXiv:2510.04587, 2025. [24] Jingyun Chen, Linghan Cai, Zhikang Wang, Yi Huang, Songhan Jiang, Shenjin Huang, Hongpeng Wang, and Yongbing Zhang. Pathagent: Toward interpretable analysis of whole-slide pathology images via large language model-based agentic reasoning. arXiv preprint arXiv:2511.17052, 2025. [25] Shengyi Hua, Jianfeng Wu, Tianle Shen, Kangzhe Hu, Zhongzhen Huang, Shujuan Ni, Zhihong Zhang, Yuan Li, Zhe Wang, and Xiaofan Zhang. Pathfound: An agentic multimodal model activating evidenceseeking pathological diagnosis. arXiv preprint arXiv:2512.23545, 2025. [26] Wenchuan Zhang, Penghao Zhang, Jingru Guo, Tao Cheng, Jie Chen, Shuwan Zhang, Zhang Zhang, Yuhao Yi, and Hong Bu. Patho-r1: A multimodal reinforcement learning-based pathology expert reasoner. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 28418–28426, 2026. [27] Songhan Jiang, Fengchun Liu, Ziyue Wang, Linghan Cai, and Yongbing Zhang. Pathreasoner-r1: Instilling structured reasoning into pathology vision-language model via knowledge-guided policy optimization. arXiv preprint arXiv:2601.21617, 2026. [28] Xinheng Lyu, Yuci Liang, Wenting Chen, Meidan Ding, Jiaqi Yang, Guolin Huang, Daokun Zhang, Xiangjian He, and Linlin Shen. Wsi-agents: A collaborative multi-agent system for multi-modal whole slide image analysis. arXiv preprint arXiv:2507.14680, 2025. [29] Xuehai He, Yichen Zhang, Luntian Mou, Eric Xing, and Pengtao Xie. Pathvqa: 30000+ questions for medical visual question answering. arXiv preprint arXiv:2003.10286, 2020. [30] Yuxuan Sun, Hao Wu, Chenglu Zhu, Sunyi Zheng, Qizi Chen, Kai Zhang, Yunlong Zhang, Dan Wan, Xiaoxiao Lan, Mengyue Zheng, et al. Pathmmu: A massive multimodal expert-level benchmark for understanding and reasoning in pathology. In European Conference on Computer Vision, pages 56–73. Springer, 2024. [31] Pingyi Chen, Chenglu Zhu, Sunyi Zheng, Honglin Li, and Lin Yang. Wsi-vqa: Interpreting whole slide images by generative visual question answering. In European Conference on Computer Vision, pages 401–417. Springer, 2024. [32] Alejandro Lozano, Jeffrey Nirschl, James Burgess, Sanket R Gupte, Yuhui Zhang, Alyssa Unell, and Serena Yeung-Levy. Micro-bench: A microscopy benchmark for vision-language understanding. Advances in Neural Information Processing Systems, 37:30670–30685, 2024. [33] Yuxuan Sun, Hao Wu, Chenglu Zhu, Yixuan Si, Qizi Chen, Yunlong Zhang, Kai Zhang, Jingxiong Li, Jiatong Cai, Yuhan Wang, et al. Pathbench: Advancing the benchmark of large multimodal models for pathology image understanding at patch and whole slide level. IEEE Transactions on Medical Imaging, 2025.

11

[34] Chunlin Zhong, Shuang Hao, Junhua Wu, Xiaona Chang, Jiwei Jiang, Xiu Nie, He Tang, and Xiang Bai. Pathvg: A new benchmark and dataset for pathology visual grounding. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 454–463. Springer, 2025. [35] Wisdom Ikezogwo, Saygin Seyfioglu, Fatemeh Ghezloo, Dylan Geva, Fatwir Sheikh Mohammed, Pavan Kumar Anand, Ranjay Krishna, and Linda Shapiro. Quilt-1m: One million image-text pairs for histopathology. Advances in neural information processing systems, 36:37995–38017, 2023. [36] Yuxuan Sun, Yunlong Zhang, Yixuan Si, Chenglu Zhu, Kai Zhang, Zhongyi Shui, Jingxiong Li, Xuan Gong, Xinheng Lyu, Tao Lin, et al. Pathgen-1.6 m: 1.6 million pathology image-text pairs generation through multi-agent collaboration. In The Thirteenth International Conference on Learning Representations, 2024. [37] Mohammad Asadi, Jack W O’Sullivan, Fang Cao, Tahoura Nedaee, Kamyar Fardi, Fei-Fei Li, Ehsan Adeli, and Euan Ashley. Mirage the illusion of visual understanding. arXiv preprint arXiv:2603.21687, 2026. [38] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. [39] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631, 2025. [40] Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261, 2025. [41] Michael Moor, Qian Huang, Shirley Wu, Michihiro Yasunaga, Yash Dalmia, Jure Leskovec, Cyril Zakka, Eduardo Pontes Reis, and Pranav Rajpurkar. Med-flamingo: a multimodal medical few-shot learner. In Machine learning for health (ML4H), pages 353–367. PMLR, 2023. [42] 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, 2023. [43] Tao Tu, Shekoofeh Azizi, Danny Driess, Mike Schaekermann, Mohamed Amin, Pi-Chuan Chang, Andrew Carroll, Charles Lau, Ryutaro Tanno, Ira Ktena, et al. Towards generalist biomedical ai. Nejm Ai, 1(3): AIoa2300138, 2024. [44] Junying Chen, Chi Gui, Ruyi Ouyang, Anningzhe Gao, Shunian Chen, Guiming Hardy Chen, Xidong Wang, Zhenyang Cai, Ke Ji, Xiang Wan, et al. Towards injecting medical visual knowledge into multimodal llms at scale. In Proceedings of the 2024 conference on empirical methods in natural language processing, pages 7346–7370, 2024. [45] Weiwen Xu, Hou Pong Chan, Long Li, Mahani Aljunied, Ruifeng Yuan, Jianyu Wang, Chenghao Xiao, Guizhen Chen, Chaoqun Liu, Zhaodonghui Li, et al. Lingshu: A generalist foundation model for unified multimodal medical understanding and reasoning. arXiv preprint arXiv:2506.07044, 2025. [46] Jeroen Van der Laak, Geert Litjens, and Francesco Ciompi. Deep learning in histopathology: the path to the clinic. Nature medicine, 27(5):775–784, 2021. [47] Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, et al. Openai gpt-5 system card. arXiv preprint arXiv:2601.03267, 2025. [48] Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479, 2025. [49] 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. [50] Wenyi Hong, Wenmeng Yu, Xiaotao Gu, Guo Wang, Guobing Gan, Haomiao Tang, Jiale Cheng, Ji Qi, Junhui Ji, Lihang Pan, et al. Glm-4.5 v and glm-4.1 v-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning. arXiv preprint arXiv:2507.01006, 2025.

12

[51] Andrew Sellergren, Sahar Kazemzadeh, Tiam Jaroensri, Atilla Kiraly, Madeleine Traverse, Timo Kohlberger, Shawn Xu, Fayaz Jamil, Cían Hughes, Charles Lau, et al. Medgemma technical report. arXiv preprint arXiv:2507.05201, 2025. [52] Andrew Sellergren, Chufan Gao, Fereshteh Mahvar, Timo Kohlberger, Fayaz Jamil, Madeleine Traverse, Alberto Tono, Bashir Sadjad, Lin Yang, Charles Lau, et al. Medgemma 1.5 technical report. arXiv preprint arXiv:2604.05081, 2026. [53] Yuze Zhao, Jintao Huang, Jinghan Hu, Xingjun Wang, Yunlin Mao, Daoze Zhang, Zeyinzi Jiang, Zhikai Wu, Baole Ai, Ang Wang, et al. Swift: a scalable lightweight infrastructure for fine-tuning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 29733–29735, 2025.

13

Related documents

• From Compound Figures to Medical Multi-image Reasoning: Scaling Multimodal Large Language Models with Biomedical Literature#1142587
• SLICEChat: Progressive In-Encoder Token Pruning for Whole-Slide Pathology Language Models#1051291
• PathAgentBench: Benchmarking Evidence-Seeking Vision-Language Models on Whole-Slide Pathology Image#386909
• Lingshu: A Generalist Foundation Model for Unified Multimodal Medical Understanding and Reasoning#1050758
• Lost in the Hype: Revealing and Dissecting the Performance Degradation of Medical Multimodal Large Language Models in Image Classification#2607