Pith. sign in

REVIEW 4 major objections 5 minor 15 references

Pathology vision-language models can answer VQA questions largely without the image; the paper shows answer accuracy and visual-semantic binding diverge, and introduces PathBind, a 2,600-sample benchmark designed to measure both.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 08:31 UTC pith:URUMUOIR

load-bearing objection A credible, well-built diagnostic benchmark showing pathology VQA accuracy overstates visual grounding; the attention-based grounding metrics are the soft spot, not the central behavioral finding. the 4 major comments →

arxiv 2607.21065 v1 pith:URUMUOIR submitted 2026-07-23 cs.CV

Do Pathology Vision-Language Models Truly See Pathology?

Chengyang Zhang , Wenchuan Zhang , Bo Li , Xinyu Liu , Jiaming Yang , Mengran Li , Chenxun Deng , Jie Chen
show 5 more authors
Yang Zhang Wei Ju Yuhao Yi Hong Bu Jiancheng Lv
This is my paper
classification cs.CV
keywords pathology vision-language modelsvisual question answeringvisual groundingmultimodal gainbenchmark evaluationattention analysishistopathologydomain training
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that answer accuracy — the standard metric for pathology vision-language models — systematically overstates how much these models actually see. Across five popular pathology VQA benchmarks, strong models answer correctly even when no image, a blank image, or an unrelated image is supplied; pathology-specific training raises accuracy more than it raises genuine image dependence; and when a model is asked to ground a specific pathological entity, its attention spreads diffusely and barely changes when the query changes. To make these gaps measurable, the authors build PathBind, a 2,600-sample benchmark whose VQA questions are filtered by seven text-only LLMs and reviewed by pathologists to require visual evidence, plus expert-curated region-level grounding annotations. Evaluated on PathBind, 18 VLMs show a substantial gap between answer-side performance and visual-semantic binding — evidence that accuracy alone is not a faithful measure of pathology understanding.

Core claim

The paper's central claim is that current pathology VLMs exhibit a substantial gap between answer-side performance and visual-semantic binding, and it identifies three concrete manifestations: visual evidence is not necessary for a large share of existing benchmark questions; pathology-specific training improves accuracy without proportional gains in multimodal gain or attention-based grounding (the 'domain training illusion'); and entity-level attention is diffuse and weakly query-specific, with entropy-equivalent support above 90% of image patches and highly correlated maps across different entity queries. The claim is supported by paired comparisons between four pathology-tuned models and

What carries the argument

Two instruments carry the argument. The first is multimodal gain: the difference between a model's accuracy with the image and without it, which quantifies whether visual evidence actually contributes to a correct answer, and is paired with text-only, blank-image, and image-replacement probes. The second is entity-token attention extraction, which takes attention from the tokens of a pathological entity phrase to image patches, averages over heads and the middle third of decoder layers, smooths with a Gaussian kernel (sigma=3), and thresholds at the top 20% of pixels to form a binary attended region; that region is compared with expert bounding boxes via IoU, precision, and recall, and its c

Load-bearing premise

The load-bearing premise is that entity-token attention — aggregated over the middle third of decoder layers, smoothed with sigma=3, and thresholded at the top 20% of pixels — faithfully reflects the visual evidence the model actually uses; the paper itself notes (Appendix C.6) that spatially supervised models carry a structural advantage on absolute grounding scores, which is why its conclusions rest on pair-difference comparisons rather than raw IoU.

What would settle it

Run a counterfactual occlusion study on PathBind-Grounding: for a high-accuracy, low-IoU model such as Patho-R1, erase the top-20% attended region (and separately erase everything except that region) and measure answer accuracy. If predictions barely move when the attended region is removed, attention maps are not the model's decision evidence and the grounding-gap reading loses force; if predictions collapse, the metrics capture something causally real. A second check: find one pathology-tuned model paired with its base that shows both a large accuracy gain and a large attention-IoU gain, whi

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Accuracy alone is not a faithful measure: a model can top a pathology VQA leaderboard while remaining nearly image-blind, so rankings that report only accuracy conflate visual skill with language-prior skill.
  • Multimodal gain (image-conditioned minus text-only accuracy) should accompany any accuracy claim; a tuned model whose text-only accuracy rises more than its image-conditioned accuracy has improved as an answerer, not as a seer.
  • Pathology-specific training does not reliably strengthen entity-level grounding: two of the four tuned models studied regress in attention IoU and recall relative to their base models, so accuracy gains from such training cannot be assumed to be visual gains.
  • Entity-level attention on current models is diffuse (entropy-equivalent support above 90% of image patches, high cross-query correlation), so grounding evaluation must use overlap metrics like IoU and precision rather than recall alone.
  • Evaluation protocols should jointly measure answer correctness, image dependence, and localized visual evidence; PathBind is a template for doing so across six diagnostic dimensions.

Where Pith is reading between the lines

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

  • The 'domain training illusion' almost certainly generalizes beyond pathology: any VQA-fine-tuned medical vision-language model — in radiology, dermatology, or endoscopy — risks the same text-prior inflation, and this paper's multimodal-gain, pair-difference protocol transfers directly as a cheap diagnostic.
  • PathBind's text-solvability filtering is bounded by the seven LLMs used; as language models improve, residual text priors will creep back into the benchmark, so it will need periodic re-filtering or adversarial question generation to keep measuring vision rather than memorization.
  • A stricter, causal test of the grounding gap would be counterfactual: occlude the top-20% attended region and check whether the answer flips. If answers are insensitive to such occlusion, the gap between accuracy and true visual dependence is even larger than attention metrics suggest.
  • Converting PathBind-Grounding into a generation task — requiring the model to produce a region citation inside its answer — would test whether reasoning-trace improvements in reinforcement-learning-tuned pathology models translate into entity-level visual evidence.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper argues that answer accuracy on pathology VQA benchmarks overstates the visual-semantic binding of pathology vision-language models. It identifies three issues: (1) many existing benchmark questions are solvable without the image (e.g., Gemini-3-Pro reaches 53.5% average text-only accuracy on five benchmarks); (2) pathology-specific training can improve answer accuracy without proportionally improving multimodal gain or attention-based grounding (e.g., Patho-R1-7B vs Qwen2.5-VL-7B); and (3) entity-level attention is diffuse and weakly query-specific on PathVG. To support these claims, the authors introduce PathBind, a 2,600-sample benchmark with three components (VQA, PTA, Grounding), constructed with automated text-only filtering, LLM-assisted generation, and expert pathological review. They evaluate 18 VLMs on VQA and 10 on grounding, reporting accuracy, multimodal gain, and attention-derived IoU/precision/recall. The central conclusion is that current pathology VLMs exhibit a substantial gap between answer-side performance and visual-semantic binding.

Significance. If the conclusions hold, the paper provides a useful diagnostic perspective and a new benchmark that could shift how pathology VLMs are evaluated. The benchmark construction is careful: two-stage text-only filtering, expert review with high inter-rater agreement (Cohen's κ 0.85–0.91), and explicit diagnostic dimensions. The paper also ships a large evaluation (18 models) and includes sensitivity analyses for attention hyperparameters (Tables C.9, C.10) and image-perturbation experiments (Appendix A.2). The main conceptual contribution—separating answer correctness from visual dependence and entity-level grounding—is important and likely to influence future work. However, the validity of the attention-based grounding metric is not established, and several key comparisons lack uncertainty quantification.

major comments (4)
  1. [Appendix C.1/C.6, Tables 3 and A.4, Figure 4c-d] The load-bearing claims about visual binding (Issues 2 and 3) rest on attention-derived grounding metrics, but the extraction protocol is not validated as a measure of the visual evidence actually used in VQA. The maps are obtained with a standalone localization prompt ('Look at the image and locate: {entity}', Appendix D.2), not with the VQA prompt. Sensitivity analyses (Tables C.9, C.10) show that model rankings are stable under changes in layer selection, threshold q, and threshold-free AP, but stability under internal perturbations does not demonstrate that the attention maps correspond to the model's causal visual evidence. The paper itself acknowledges in C.6 that the pair-difference analysis is the load-bearing claim, yet that analysis still relies on the same unvalidated attention proxy. I recommend adding a validation study: e.g., correlate attention IoU with behaviorally measur
  2. [Table 3, Figure 4, Section 5.2] All comparisons are point estimates with no confidence intervals or significance tests. For example, Table 3 reports Patho-R1-7B IoU 17.4 vs Qwen2.5-VL-7B 23.3 (a 5.9-point gap on 380 testA samples), and Figure 4c shows ΔIoU values of −4.7 and −3.7 for the two key pairs. Without per-sample variability, bootstrap CIs, or paired tests, these differences could be within sampling noise. This is particularly important for the null claim that pathology tuning does not improve visual binding: a null result requires confidence bounds to be interpretable. I recommend reporting 95% bootstrap confidence intervals for all grounding metrics and paired bootstrap or permutation tests for the ΔIoU/ΔRecall comparisons.
  3. [Section 3.3, Appendix C.4] The cross-query attention correlation metric compares attention maps for the ground-truth entity query and two alternative entity expressions 'sampled from other examples'. These alternative expressions are likely absent from the image. If a model is asked to locate an entity that is not present, it may fall back to generic or uniform attention, leading to high correlation with the ground-truth map for reasons unrelated to query-specificity. The claim that 'attention maps remain highly correlated across different entity queries' would be more convincing if alternative queries were visually present in the same image but referred to different regions, or if the analysis were stratified by whether the alternative entity is actually present. As written, the experiment does not cleanly separate query-insensitivity from a response to an absent entity.
  4. [Abstract, Section 5.2, Conclusion] The 'domain training illusion' is stated as a general phenomenon ('domain-specific training improves answer accuracy without proportional gains in visual binding'), but the evidence in Figure 4 is mixed. On PathBind, PathGen-LLaVA and HuatuoGPT-Vision improve multimodal gain by 4.1 and 2.9 points, respectively, while Quilt-LLaVA and Patho-R1 decrease. Grounding results also show gains for some pairs and losses for others. The general claim should be qualified to 'in some model families' or 'does not consistently occur', which would still be a meaningful finding. The current wording overstates the uniformity of the effect.
minor comments (5)
  1. [Abstract] The abstract reports 'Patho-R1-7B exhibits a 5.8-point lower multimodal gain and a 3.7-point lower attention IoU' without specifying the evaluation subsets. These numbers correspond to PathMMU-test-tiny and PathVG-testA, respectively; please state this to avoid ambiguity.
  2. [Figure 2] The label 'E!ective attention support ratio' contains a typo ('!' should be 'f'). Also, the caption for panel (b) is not fully self-contained; it should define what the reported values are (entropy-equivalent support ratio) as done in Appendix C.3.
  3. [Appendix C.1, Equation (2)] The layer-selection formula uses floor/ceil notation that renders ambiguously. Please use explicit \lfloor and \rceil symbols, and clarify whether indices are 0-based or 1-based.
  4. [Figure 4] The sign convention for Δ is confusing: in some places it appears as tuned−base and in others as base−tuned (e.g., the LLaVA pair in panel (a)). Please use a consistent convention and define it clearly in the caption.
  5. [Section 2/References] Several references are to future or preprint works with fictional-sounding names (e.g., GPT-5.4, Gemini-3-Pro, Qwen3.6-Max, GLM-5). While the paper appears to be a 2026 arXiv preprint, please ensure these citations are accurate and not placeholder names. If they are real, provide DOIs or version numbers.

Circularity Check

0 steps flagged

No load-bearing circularity: central claims are empirical measurements, with only minor construction-dependence and non-load-bearing self-citation.

full rationale

The paper's derivation chain is empirical throughout. Issue 1 (visual evidence not necessary) is established by comparing image-conditioned and text-only accuracy on existing benchmarks (Table 1), not by any fitted parameter. Issue 2 (domain training illusion) is established by paired differences in multimodal gain and attention-based IoU/recall (Figure 2a, Figure 4); these are direct measurements on PathVG/PathBind-Grounding. Issue 3 (diffuse, weakly query-specific attention) is measured by entropy-equivalent support and cross-query Pearson correlation (Appendix C.3-C.4). The only self-citation is Patho-R1, which is evaluated as a subject rather than invoked as a theorem; even if omitted, Quilt-LLaVA/PathGen-LLaVA comparisons carry the same conclusion. The text-only LLM filtering in PathBind construction is a design choice that could influence multimodal-gain magnitudes, but the paper does not present these gains as a prediction; it reports them as benchmark properties and checks model rankings with threshold-free AP (Table C.10). The attention-extraction protocol is an unvalidated proxy for grounding, but the paper acknowledges this in Appendix C.6 and restricts its load-bearing claim to pair-difference analysis; that is a validity limitation, not circularity. No equation reduces to its input by construction. Score 2 reflects the minor construction-dependence and self-citation, not load-bearing circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper's contributions are empirical: a new benchmark and evaluation protocol. The main burden is the attention-as-grounding proxy and the text-only filtering thresholds, not new theoretical entities. No invented physical or conceptual entities are introduced.

free parameters (5)
  • Attention threshold q = 0.2 = 0.2
    Chosen to match the 90th percentile of ground-truth bounding-box areas and to sit at the IoU peak for the strongest models (Appendix C.1, Table C.10). A sensitivity sweep shows model ranking is invariant.
  • Gaussian smoothing sigma = 3 = 3
    Applied to attention heatmaps before thresholding (Algorithm 1). No sensitivity sweep is reported for sigma.
  • Middle-third decoder layer aggregation = middle third of layers
    Selected based on HiDe's empirical peak-layer observation, with a sanity check on six models (Table C.9). Affects all absolute grounding numbers.
  • Text-only filtering thresholds = remove if 2/2 closed-source correct; remove if >=3/5 additional LLMs correct
    Two-stage filtering determines which samples enter PathBind-VQA (Appendix B.1). These thresholds are design choices rather than fitted to the target claim, but they shape the benchmark.
  • K = 3 entity queries for cross-query attention correlation = 3
    Number of alternative queries used in Figure 2(c); fixed seed for comparability, but the choice of K is arbitrary.
axioms (5)
  • domain assumption Entity-token attention extracted from decoder layers is a valid proxy for visual grounding.
    The entire grounding analysis (Tables 3, A.4, Figure 2) depends on this assumption. The authors partially hedge in C.6 but do not validate attention against causal interventions.
  • domain assumption Pathologist annotations and expert curation provide accurate ground truth for region-level grounding.
    Ground-truth bounding boxes on PathVG and expert-reviewed PathBind-Grounding are treated as correct without external verification.
  • domain assumption Paired base/tuned model differences isolate the effect of pathology-specific tuning.
    Figure 4 and the 'domain training illusion' rely on this. Confounded by differing training data, RL objectives, and tokenizers, as the authors partly acknowledge in C.6.
  • domain assumption Text-only LLM accuracy approximates the degree to which a VQA sample is solvable via textual shortcuts.
    The filtering pipeline assumes LLMs with no image access are a good proxy for shortcut solvability; the evaluated VLMs may have different text priors.
  • domain assumption Benchmark answer keys and source labels are correct.
    All accuracy computations inherit the correctness of PathMMU, Path-VQA, Quilt-VQA, MedXpert-Path, OmniMed-Bright, and the private atlas labels.

pith-pipeline@v1.3.0-alltime-deepseek · 31085 in / 7867 out tokens · 86139 ms · 2026-08-01T08:31:29.575568+00:00 · methodology

0 comments
read the original abstract

Pathology vision-language models (VLMs) have recently progressed rapidly and are commonly evaluated by answer accuracy on pathology VQA benchmarks. However, we dig into current evaluations and identify three overlooked issues: 1) Visual evidence is not always necessary. For instance, Gemini-3-Pro achieves 53.5% average accuracy across 5 VQA benchmarks without any visual input. 2) Domain training can improve accuracy without proportional gains in visual binding. Compared with Qwen2.5-VL-7B, Patho-R1-7B exhibits a 5.8-point lower multimodal gain and a 3.7-point lower attention IoU. 3) Entity-level attention is diffuse and weakly query-specific. On PathVG, attention maps remain highly correlated across different entity queries. These issues can lead to substantial misjudgments of pathology VLMs' actual multimodal capabilities. To this end, we present PathBind, a benchmark comprising 2,600 samples: PathBind-VQA with 1,500 questions across six dimensions, PathBind-PTA with 600 questions from a private pathology teaching atlas, and PathBind-Grounding with 500 expert-curated region-level samples. Each component undergoes task-specific automated filtering and expert review to reduce textual shortcuts and improve entity-region correspondence. We evaluate 18 representative VLMs on VQA samples of PathBind and five existing pathology VQA benchmarks, and further evaluate 10 VLMs on PathBind-Grounding and PathVG. Results show that current pathology VLMs still exhibit a substantial gap between answer-side performance and visual-semantic binding.

Figures

Figures reproduced from arXiv: 2607.21065 by Bo Li, Chengyang Zhang, Chenxun Deng, Hong Bu, Jiaming Yang, Jiancheng Lv, Jie Chen, Mengran Li, Wei Ju, Wenchuan Zhang, Xinyu Liu, Yang Zhang, Yuhao Yi.

Figure 1
Figure 1. Figure 1: (a) Pathology-tuned VLM answering correctly un [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a) Paired deltas on PathMMU-test-tiny for answer [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the PathBind data curation process. PathBind integrates three components: a filtered VQA set, a private [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: General-VLM → pathology-tuned pairs across four metrics. For each pair, we compare (a) Multimodal Gain on five raw public benchmarks; (b) MG on PathBind (VQA + PTA); (c) IoU and (d) Recall on PathBind-Grounding. Green/red lines mark pairs where the tuned model scores above/below its general model. Observation from visual dependence. Although Path￾Bind removes samples that can be consistently solved by mul￾… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

15 extracted references · 8 linked inside Pith

  1. [2]

    InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, 23431–23441

    Pathfinder: A multi-modal multi-agent system for medical diagnostic decision-making applied to histopathology. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, 23431–23441. Gilal, N. U.; Zegour, R.; Al-Thelaya, K.; Özer, E.; Agus, M.;Schneider,J.;andBoughorbel,S.2025. PathVLM-Eval: Evaluationofopenvisionlanguagemodelsinhistopat...

    2025

  2. [3]

    strategy

    A.1 More Results on Pathology VQA Benchmarks and PathBind Full evaluation results on five pathology VQA bench- marks.Table A.1 extends Table 1 in the main paper by including additional general-purpose VLMs. Following the same protocol, we evaluate each model in both image- conditionedandtext-onlysettingsacrossmultiplepathology VQA benchmarks. Consistent w...

    2025

  3. [6]

    Liu, H.; Li, C.; Li, Y.; and Lee, Y

    PathReasoner-R1:InstillingStructuredReasoninginto Pathology Vision-Language Model via Knowledge-Guided Policy Optimization.arXiv preprint arXiv:2601.21617. Liu, H.; Li, C.; Li, Y.; and Lee, Y. J

    arXiv

  4. [7]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 26296–26306

    Improved baselines with visual instruction tuning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 26296–26306. Liu,X.;Hu,Y.;Zou,Y.;Wu,L.;Xu,J.;andZheng,B.2025. HiDe: Rethinking The Zoom-IN method in High Resolu- tion MLLMs via Hierarchical Decoupling.arXiv preprint arXiv:2510.00054. Lu,M.Y.;Chen,B.;Williamson,D.F.;Che...

    Pith/arXiv arXiv 2025

  5. [8]

    How good is my histopathologyvision-languagefoundationmodel?aholistic benchmark.arXiv preprint arXiv:2503.12990. Meta AI

    Pith/arXiv arXiv

  6. [9]

    arXiv preprint arXiv:2502.13923

    Qwen2.5-VL Technical Report. arXiv preprint arXiv:2502.13923. Seyfioglu, M. S.; Ikezogwo, W. O.; Ghezloo, F.; Krishna, R.; and Shapiro, L

    Pith/arXiv arXiv

  7. [10]

    Advances in Neural Information Processing Systems, 38: 101673–101731

    CPathAgent: An Agent-based Foun- dation Model for Interpretable High-Resolution Pathology Image Analysis Mimicking Pathologists’ Diagnostic Logic. Advances in Neural Information Processing Systems, 38: 101673–101731. Sun,Y.;Wu,H.;Zhu,C.;Si,Y.;Chen,Q.;Zhang,Y.;Zhang, K.; Li, J.; Cai, J.; Wang, Y.; et al. 2025b. PathBench: Ad- vancing the Benchmark of Large...

    2025

  8. [11]

    Wu, J.; Yang, H.; Zeng, X.; He, G.; Chen, Z.; Li, Z.; Zhang, X.; Ma, Y.; Fang, R.; and Liu, Y

    Qwen2-vl: En- hancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191. Wu, J.; Yang, H.; Zeng, X.; He, G.; Chen, Z.; Li, Z.; Zhang, X.; Ma, Y.; Fang, R.; and Liu, Y

    Pith/arXiv arXiv

  9. [12]

    Zafar, A.; Murali, L

    Pathvlm-r1: A re- inforcement learning-driven reasoning model for pathology visual-language tasks.arXiv preprint arXiv:2504.09258. Zafar, A.; Murali, L. K.; and Vashist, A

    Pith/arXiv arXiv

  10. [13]

    Zhang, W.; Guo, J.; Zhang, H.; Zhang, P.; Chen, J.; Zhang, S.;Zhang,Z.;Yi,Y.;andBu,H.2026a

    Beyond ac- curacy: Evaluating visual grounding in multimodal medical reasoning.arXiv preprint arXiv:2603.03437. Zhang, W.; Guo, J.; Zhang, H.; Zhang, P.; Chen, J.; Zhang, S.;Zhang,Z.;Yi,Y.;andBu,H.2026a. Patho-AgenticRAG: towards multimodal agentic retrieval-augmented generation for pathology VLMs via reinforcement learning. InPro- ceedings of the AAAI Co...

    arXiv 2025

  11. [14]

    A Additional Experimental Results and Analyses In this section, we provide additional experimental results and analyses that complement the findings in the main pa- per

    Medxpertqa: Benchmarking expert-level medical reasoning and understanding.arXiv preprint arXiv:2501.18362. A Additional Experimental Results and Analyses In this section, we provide additional experimental results and analyses that complement the findings in the main pa- per. Specifically, we report (1) expanded evaluation results on additional VLMs, (2) ...

    Pith/arXiv arXiv

  12. [2020]

    Hu, Y.; Li, T.; Lu, Q.; Shao, W.; He, J.; Qiao, Y.; and Luo, P

    Pathvqa: 30000+ questions for medical visual question an- swering.arXiv preprint arXiv:2003.10286. Hu, Y.; Li, T.; Lu, Q.; Shao, W.; He, J.; Qiao, Y.; and Luo, P

    Pith/arXiv arXiv 2003

  13. [2024]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 22170–22183

    Omnimedvqa: A new large-scale comprehensive evaluation benchmark for medical lvlm. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 22170–22183. Ikezogwo, W.; Seyfioglu, S.; Ghezloo, F.; Geva, D.; Sheikh Mohammed, F.; Anand, P. K.; Krishna, R.; and Shapiro,L.2023. Quilt-1m:Onemillionimage-textpairsfor histopathology.Adv...

    2023

  14. [2025]

    Chen, J.; Gui, C.; Ouyang, R.; Gao, A.; Chen, S.; Chen, G

    Pathagent: Toward inter- pretable analysis of whole-slide pathology images via large language model-based agentic reasoning.arXiv preprint arXiv:2511.17052. Chen, J.; Gui, C.; Ouyang, R.; Gao, A.; Chen, S.; Chen, G. H.; Wang, X.; Cai, Z.; Ji, K.; Wan, X.; et al. 2024a. Towardsinjectingmedicalvisualknowledgeintomultimodal llmsatscale. InProceedingsoftheCon...

    arXiv

  15. [2026]

    Google DeepMind

    GLM-5: From Vibe Coding to Agentic Engineering.arXiv preprint arXiv:2602.15763. Google DeepMind

    Pith/arXiv arXiv