Pith. sign in

REVIEW 3 major objections 5 minor 86 references

Retrieval-augmented in-context learning lets small multimodal language models classify wound images without task-specific retraining, with the best model reaching 0.872 accuracy on a public wound dataset—beating both zero-shot prompting and

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 14:12 UTC pith:HYSFJ2RE

load-bearing objection Solid empirical ICL-for-wound-classification paper whose central 'beyond kNN' claim is conditional on unaddressed duplicate leakage risks and a single-split evaluation; worth refereeing, but the authors should be pushed to release splits and screen Medetec for near-duplicates. the 3 major comments →

arxiv 2607.18819 v1 pith:HYSFJ2RE submitted 2026-07-21 cs.CV

In-Context Learning for Wound Classification with Small Multimodal Language Models

George Martvel , Oskar Gustafsson , John Pavia , Ernst Ahlberg This is my paper
classification cs.CV
keywords in-context learningmultimodal language modelswound image classificationretrieval-augmented promptingk-nearest neighbour retrievalmaximal marginal relevanceCLIP embeddingsfew-shot classification
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 asks whether small multimodal language models, running locally, can classify wound images without being retrained for each new task. Its answer: yes, if the prompt contains a handful of labelled example images retrieved specifically for the query image. Eleven models from three families were tested on two public wound datasets; query-conditioned retrieval consistently beat zero-shot and random few-shot prompting, and the largest models outperformed a pure nearest-neighbour baseline, reaching 0.872 accuracy on one dataset and 0.678 on the other. The gains came mostly from retrieval itself and saturate at about 8 to 10 support images, suggesting compact prompts suffice. This matters because wound images are privacy-sensitive, and a retraining-free, locally runnable classifier could adapt to changing label sets without sending clinical images to external servers.

Core claim

The paper's central claim is that retrieval-augmented in-context learning lets small multimodal language models classify wound images without task-specific training. For each query, the most similar labelled support images are retrieved via CLIP ViT-B/32 embeddings (kNN), optionally reranked for diversity (MMR), and placed in the prompt. Across 11 models on two datasets, this query-conditioned prompting consistently beat zero-shot and random few-shot prompting. The best model (Qwen 3.5 27B with kNN+MMR) reached 0.872 accuracy on the Kaggle dataset and 0.678 on Medetec, above the retrieval-only weighted-kNN controls (0.708 and 0.467). Paired diagnostics show larger models preserve most retrie

What carries the argument

The central object is the retrieved support context: a small set of labelled wound images, selected by CLIP ViT-B/32 cosine similarity to the query (kNN) and optionally diversity-reranked by maximal marginal relevance (MMR), and inserted into the prompt as image–label pairs. The retrieval-only weighted-kNN classifier on the same embeddings serves as the control that isolates what the language model adds. The diagnostic machinery is a paired comparison: retention rate P(SMLM correct | control correct), rescue rate P(SMLM correct | control wrong), McNemar tests, and Cohen's kappa, which together show whether the model goes beyond nearest-neighbour voting.

Load-bearing premise

The claim collapses if CLIP ViT-B/32 cosine similarity does not track the visual cues that distinguish wound classes—then the retrieved support examples are not informative and kNN-guided ICL should fall back to random few-shot levels.

What would settle it

Run the identical ICL protocol but replace the CLIP ViT-B/32 encoder with a feature extractor that is blind to wound-specific appearance (e.g., a fixed random projection of raw pixels). If kNN retrieval still outperforms random few-shot prompting by the same margin, the retrieval mechanism is not doing the causal work.

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

If this is right

  • Wound classification can be repurposed to a new label set by replacing the support pool; no retraining or parameter updates are needed.
  • Compact contexts of about 8–10 retrieved images capture most of the ICL benefit; adding more images mostly increases latency and token count.
  • The largest SMLMs exceed the embedding-space weighted-kNN vote, so the model contributes task knowledge beyond the retrieved neighbourhood.
  • At 25% of the full labelled support, retrieval-based ICL still outperforms fine-tuned supervised baselines (e.g., 0.781 vs 0.666 top accuracy on Kaggle; 0.527 vs 0.413 on Medetec).
  • Retrieval-based gains hold across two datasets and eleven models, but absolute performance depends on model scale and dataset difficulty.

Where Pith is reading between the lines

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

  • A testable extension: because the label set is defined by the support pool, the same model could serve different wound-taxonomy tasks (e.g., tissue type vs wound type) by swapping support examples; this could be evaluated by holding the model fixed and changing only the support labels.
  • The saturation at around 10 examples suggests a deployment rule: cap the context at about 10 images, then allocate any remaining compute budget to better retrieval embeddings rather than longer prompts.
  • If the 'rescue' behaviour is real, an ensemble that combines the SMLM's prediction with the weighted-kNN vote could be more robust than either alone; a simple logistic blend on the two scores would test this.
  • The paper leaves open whether a wound-specific embedding (e.g., fine-tuned on wound images) would raise both the kNN ceiling and the ICL ceiling; replacing CLIP ViT-B/32 with a medical embedding and re-running the protocol is a direct next experiment.

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

3 major / 5 minor

Summary. The paper evaluates retrieval-augmented in-context learning (ICL) for wound-image classification using eleven small multimodal language models (Qwen 3.5, Ministral 3, Gemma 4) on two public datasets (Kaggle wound, 1469 images/10 classes; Medetec, 560 images/9 classes). The authors compare zero-shot prompting, random few-shot prompting, CLIP-based kNN retrieval, and kNN+MMR retrieval, with retrieval-only weighted-kNN controls. They report that query-conditioned ICL consistently outperforms zero-shot and random few-shot prompting, that the largest models (Qwen 3.5 27B, Qwen 3.5 9B, Gemma 4 26B) significantly exceed matched weighted-kNN baselines, and that gains saturate at roughly 8–10 support images. Support-set reduction experiments and comparisons with fine-tuned supervised baselines are used to support the claim that retrieval-augmented ICL is a flexible, training-free alternative that is particularly competitive under limited labelled data. The paper explicitly acknowledges the absence of cross-validation and the use of CLIP as the retrieval embedding.

Significance. If the central claim holds, the paper offers practically relevant evidence that locally deployable SMLMs with retrieved in-context examples can perform wound classification without task-specific retraining, which matters for privacy-sensitive clinical settings. The experimental design is stronger than many ICL evaluations: random few-shot prompting is a well-chosen control, the MMR control is recomputed on the final reranked support set, and the retention/rescue diagnostics directly address the question of whether the model merely echoes retrieved labels. The support-set reduction and context-size sweeps provide useful practical guidance on data and compute trade-offs. However, the headline 'beyond nearest-neighbour voting' result rests on the integrity of a single fixed split and on the assumption that CLIP similarity tracks class-relevant visual cues. The paper does not release code, split seeds, or image-level split lists, which makes the leakage concern difficult to audit. These issues are substantial but addressable within the manuscript's scope.

major comments (3)
  1. [§3.3, §4.2, §5] The integrity of the single fixed support/evaluation split is load-bearing for the central claim. §3.3 reports that mirrored duplicates were removed only for Kaggle ('half of the images were removed'), and no duplicate or near-duplicate screening is reported for Medetec. If a query image has an exact or near-duplicate in the support set, CLIP retrieval will place it at the top of the context and the SMLM can effectively copy its label. The very high any-support match rates (0.95–0.98) and retention rates (0.94–0.97) in Tables 3–4 are exactly the pattern label-copying would produce. Because the SMLM is compared with a weighted-kNN baseline that averages over K neighbours and is therefore less able to exploit a single near-duplicate, such leakage inflates the claimed 'beyond nearest-neighbour voting' advantage. The authors acknowledge in §5 that cross-validation was not used; this makes sp
  2. [§3.3, §5, Tables 1–2] The evaluation rests on a single balanced split of only 250 (Kaggle) and 90 (Medetec) images. All accuracy deltas, McNemar tests, and confidence intervals are computed on this one split and therefore do not capture split-to-split variability. The support-set reduction experiment uses five seeds for subsampling, but the evaluation split is fixed (§3.8), so it cannot estimate the variance of the main ICL-versus-baseline comparison. With 90 evaluation images on Medetec, a favourable draw or a few near-duplicates can move accuracy by several points and change which models significantly exceed the kNN baseline. Please provide repeated random support/eval splits (e.g., 5–10 seeds) or a bootstrap over evaluation images, and report the distribution of the accuracy delta between ICL and the matched weighted-kNN control.
  3. [§3.4] The retrieval pipeline assumes that CLIP ViT-B/32 cosine similarity is a valid proxy for class-relevant visual similarity among wound images. The random-support control provides indirect evidence that retrieval is informative, but it does not establish that similarity corresponds to clinically meaningful wound cues rather than to background, lighting, or duplication artefacts. This assumption is load-bearing for the practical claim that retrieved contexts can 'specify the target label set and decision criteria at inference time.' A concrete test would be to repeat the main comparison with an alternative encoder (e.g., DINOv2 or BiomedCLIP) or to report qualitative nearest-neighbour examples and class-conditional similarity distributions. At minimum, the paper should state this dependency explicitly in the abstract and conclusions, since the portability of the method to new wound tasks de
minor comments (5)
  1. [§3.3] Please report the original Kaggle image count and describe how mirrored duplicates were detected (e.g., manual review, image hashing, or embedding similarity). This will help readers assess the completeness of the de-duplication.
  2. [Tables 1–2 and throughout] The table headers render as 'F ew-shot' in several places; this appears to be a formatting artifact and should be corrected to 'Few-shot'.
  3. [§3.7] The McNemar test is described as 'exact' in the text but the equations only define the standard test. Please clarify whether the binomial exact version was used for small eval sets, and report the test statistic or the contingency table values for the key comparisons.
  4. [§5, Discussion] The sentence 'These results should not be interpreted as showing that ICL removes the need for labelled data' is an important caveat. Consider moving a version of it to the abstract, since the current abstract's phrase 'training-free' could be misread as 'data-free.'
  5. [Figure 1] The schematic images are generated with GPT Image 2 and used only for illustration. The caption says this, which is good; please ensure the figure itself is clearly labelled so readers do not confuse it with a real wound image.

Circularity Check

0 steps flagged

No significant circularity: the paper is an empirical evaluation with independent retrieval-only controls, and no fitted parameter is presented as a prediction.

full rationale

This paper is an empirical benchmark study rather than a derivation, so the circularity patterns mostly do not apply. The central claims—that query-conditioned ICL outperforms zero-shot/random few-shot prompting and that larger SMLMs exceed weighted-kNN controls—are supported by direct experiments (Tables 1–4), not by an equation that reduces to its own inputs. The retrieval-only weighted-kNN controls are computed from the same CLIP embeddings used for support selection, but they are independent baselines rather than fitted parameters; no accuracy number is obtained by fitting the model to the test set. The kNN+MMR control is deliberately recomputed on the final reranked support set so that the comparison is matched to the prompt content, which is a methodological safeguard rather than circularity. The high any-support match rates and retention/rescue diagnostics are reported transparently as evidence that predictions are strongly shaped by the retrieved context; the paper does not claim these rates prove reasoning, and the conclusion that larger models go beyond nearest-neighbour voting is an empirical inference drawn from McNemar tests and rescue rates. The acknowledged absence of cross-validation and the lack of duplicate screening for Medetec are external-validity or leakage concerns, not circularity, because they do not make the reported outcome equivalent to the experimental inputs by construction. No load-bearing self-citation, imported uniqueness theorem, or ansatz-smuggling chain is present. The paper is therefore free of the circularity patterns defined in this analysis.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central claims are empirical, and the only hand-set quantities are hyperparameters such as lambda, K, and M; no new entities or fitted targets are introduced. The results do depend on the CLIP embedding assumption, the prompt-following behaviour of instruction-tuned models, and the integrity of the random splits.

free parameters (3)
  • MMR diversity weight lambda = 0.7
    Hand-chosen in §3.4; controls the trade-off between relevance and diversity in kNN+MMR retrieval, with no sensitivity analysis.
  • Support context size K = 10
    Hand-chosen in §3.7; the context-size sweep shows gains saturate near 8-10 images, so this is not load-bearing for the qualitative claim.
  • MMR candidate pool size M = 30
    Hand-chosen in §3.7; selects the pool from which the K reranked examples are drawn.
axioms (5)
  • domain assumption CLIP ViT-B/32 cosine similarity is a valid proxy for class-relevant visual similarity among wound images
    Invoked in §3.4 as the basis for all query-conditioned support selection; if false, the retrieval gains are an artifact of the embedding choice.
  • domain assumption Presenting retrieved images as labelled 'Example image for class y' pairs is sufficient for instruction-tuned SMLMs to use them as visual references
    Prompt design in §3.6/Figure 4; the central ICL effect depends on models following this instruction.
  • domain assumption Randomly drawn balanced support/evaluation splits contain no near-duplicate leakage
    §3.3; mirrored duplicates were removed only for Kaggle, no duplicate screening is reported for Medetec, and no cross-validation is used.
  • domain assumption Quantized checkpoints with temperature 0 faithfully represent the evaluated models
    §3.5; different quantisation schemes across models may confound cross-model comparisons.
  • standard math Standard paired statistical tests (McNemar, Fisher) are appropriate for the correctness-pair comparisons
    Used in §3.7/4.2; standard methods for paired accuracy tables.

pith-pipeline@v1.3.0-alltime-deepseek · 29590 in / 12257 out tokens · 106316 ms · 2026-08-01T14:12:34.120431+00:00 · methodology

0 comments
read the original abstract

Wound image classification is often treated as a task-specific supervised learning problem, requiring substantial amounts of manually labelled data and retraining when the label space or deployment setting changes. This study evaluated whether small multimodal language models (SMLMs) can provide a training-free alternative for wound classification through retrieval-based in-context learning (ICL). Experiments used two public wound-image datasets: the Kaggle wound dataset (1469 images, 10 classes) and the Medetec dataset (560 images, 9 classes). Eleven SMLMs from the Qwen 3.5, Ministral 3, and Gemma 4 families were evaluated under zero-shot prompting and few-shot prompting with random support examples, embedding-based k-nearest-neighbour (kNN) retrieval, and kNN retrieval followed by maximal marginal relevance reranking (MMR). Retrieval-only weighted-kNN controls, support-set reduction experiments, and support-context size sweeps were used to assess the effects of retrieval, model scale, and prompt length. Query-conditioned ICL consistently outperformed zero-shot and random few-shot prompting. On the Kaggle dataset, the best result was achieved by Qwen 3.5 27B with kNN+MMR, reaching 0.872 accuracy and 0.871 F1 score. On Medetec, Qwen 3.5 27B with kNN+MMR reached 0.678 accuracy and 0.670 F1. Larger models exceeded matched weighted-kNN controls, indicating use of retrieved examples beyond nearest-neighbour voting. Retrieval-based ICL degraded modestly under support-set reduction, and most gains saturated with 8-10 support images. Retrieval-based ICL allows SMLMs to perform adaptable wound image classification without task-specific retraining. Compact retrieved contexts may support practical and privacy-conscious deployment, although performance remains dependent on model scale, retrieval strategy, and dataset difficulty.

Figures

Figures reproduced from arXiv: 2607.18819 by Ernst Ahlberg, George Martvel, John Pavia, Oskar Gustafsson.

Figure 1
Figure 1. Figure 1: Schematic overview of the ICL pipeline used in the current study. The wound images [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Class distributions for the Kaggle dataset (left) and the Medetec dataset (right). Images [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Zero-shot prompt structure. 7 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Few-shot (K-shot) prompt structure. 3.7 Retrieval influence diagnostics To assess how closely retrieval-augmented SMLM predictions aligned with the retrieved embedding￾space neighbourhood during ICL [49, 84], paired comparisons were conducted between each model and a retrieval-only control. The same diagnostics were applied to both kNN and kNN+MMR retrieval. For standard kNN, the global weighted-kNN baseli… view at source ↗
Figure 5
Figure 5. Figure 5: SMLM accuracy as a function of support-context size for different support-selection [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] 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

86 extracted references · 17 linked inside Pith

  1. [1]

    Can machine learning be used to discriminate between burns and pressure ulcer? InProceedings of SAI intelligent systems conference, pages 870–880

    Aliyu Abubakar, Hassan Ugail, and Ali Maina Bukar. Can machine learning be used to discriminate between burns and pressure ulcer? InProceedings of SAI intelligent systems conference, pages 870–880. Springer, 2019

    2019

  2. [2]

    Prompt leakage effect and defense strategies for multi-turn llm interactions.arXiv preprint arXiv:2404.16251, 2024

    Divyansh Agarwal, Alexander R Fabbri, Ben Risher, Philippe Laban, Shafiq Joty, and Chien- Sheng Wu. Prompt leakage effect and defense strategies for multi-turn llm interactions.arXiv preprint arXiv:2404.16251, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    Diagnostic accuracy of deep learning in medical imaging: a systematic review and meta-analysis.NPJ digital medicine, 4(1):65, 2021

    Ravi Aggarwal, Viknesh Sounderajah, Guy Martin, Daniel SW Ting, Alan Karthikesalingam, Dominic King, Hutan Ashrafian, and Ara Darzi. Diagnostic accuracy of deep learning in medical imaging: a systematic review and meta-analysis.NPJ digital medicine, 4(1):65, 2021

    2021

  4. [4]

    Flamingo: a visual language model for few-shot learning.Advances in neural information processing systems, 35:23716–23736, 2022

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning.Advances in neural information processing systems, 35:23716–23736, 2022

    2022

  5. [5]

    A novel fused multi-class deep learning approach for chronic wounds classification.Applied Sciences, 13(21):11630, 2023

    Zaid A Aldoulah, Hafiz Malik, and Richard Molyet. A novel fused multi-class deep learning approach for chronic wounds classification.Applied Sciences, 13(21):11630, 2023

    2023

  6. [6]

    Multi-modal wound classification using wound image and location by deep neural network.Scientific Reports, 12(1):17298, 2022

    D M Anisuzzaman, Behnoosh Rostami, Joseph Niezgoda, Ashwin Gopalakrishna, Zhihong Yu, and Zhiqiang Liang. Multi-modal wound classification using wound image and location by deep neural network.Scientific Reports, 12(1):17298, 2022

    2022

  7. [7]

    Image-based artificial intelligence in wound assessment: a systematic review.Advances in Wound Care, 11(12):687–709, 2022

    DM Anisuzzaman, Chuanbo Wang, Behrouz Rostami, Sandeep Gopalakrishnan, Jeffrey Niez- goda, and Zeyun Yu. Image-based artificial intelligence in wound assessment: a systematic review.Advances in Wound Care, 11(12):687–709, 2022

    2022

  8. [8]

    In-context learning for data-efficient diabetic retinopathy detection via multimodal foundation models.Ophthalmology Science, page 100934, 2025

    Murat S Ayhan, Ariel Y Ong, Eden Ruffell, Siegfried K Wagner, David A Merle, and Pearse A Keane. In-context learning for data-efficient diabetic retinopathy detection via multimodal foundation models.Ophthalmology Science, page 100934, 2025

    2025

  9. [9]

    Foundational models in medical imaging: A comprehensive survey and future vision.arXiv preprint arXiv:2310.18689, 2023

    Bobby Azad, Reza Azad, Sania Eskandari, Afshin Bozorgpour, Amirhossein Kazerouni, Islem Rekik, and Dorit Merhof. Foundational models in medical imaging: A comprehensive survey and future vision.arXiv preprint arXiv:2310.18689, 2023

    Pith/arXiv arXiv 2023

  10. [10]

    Evaluatinggeneralpurposevisionfoundationmodelsformedicalimageanalysis: An experimental study of dinov2 on radiology benchmarks.arXiv preprint arXiv:2312.02366, 2023

    Mohammed Baharoon, Waseem Qureshi, Jiahong Ouyang, Yanwu Xu, Abdulrhman Aljouie, andWeiPeng. Evaluatinggeneralpurposevisionfoundationmodelsformedicalimageanalysis: An experimental study of dinov2 on radiology benchmarks.arXiv preprint arXiv:2312.02366, 2023

    Pith/arXiv arXiv 2023

  11. [11]

    Ai privacy risks & mitigations–large language models (llms).European Data Protection Board, pages 2025–04, 2025

    Isabel Barbera. Ai privacy risks & mitigations–large language models (llms).European Data Protection Board, pages 2025–04, 2025

    2025

  12. [12]

    Artificial intelligence in medical imaging: from task-specific models to large-scale foundation models.Chinese Medical Journal, 138(06):651–663, 2025

    Yueyan Bian, Jin Li, Chuyang Ye, Xiuqin Jia, and Qi Yang. Artificial intelligence in medical imaging: from task-specific models to large-scale foundation models.Chinese Medical Journal, 138(06):651–663, 2025

    2025

  13. [13]

    Tabja Bortesi, Juan Pablo, Jonathan Ranisau, Shuang Di, Michael McGillion, Laura Rosella, Alistair Johnson, P. J. Devereaux, and Jeremy Petch. Machine learning approaches for the 22 image-based identification of surgical wound infections: Systematic review.Journal of Medical Internet Research, 26:e52880, 2024

    2024

  14. [14]

    Large language models and large multimodal models in medical imaging: a primer for physicians

    Tyler J Bradshaw, Xin Tie, Joshua Warner, Junjie Hu, Quanzheng Li, and Xiang Li. Large language models and large multimodal models in medical imaging: a primer for physicians. Journal of nuclear medicine, 66(2):173–182, 2025

    2025

  15. [15]

    Language models are few-shot learners.Advances in neural information processing systems, 33:1877– 1901, 2020

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhari- wal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.Advances in neural information processing systems, 33:1877– 1901, 2020

    1901

  16. [16]

    The use of mmr, diversity-based reranking for reordering documents and producing summaries

    Jaime Carbonell and Jade Goldstein. The use of mmr, diversity-based reranking for reordering documents and producing summaries. InProceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval, pages 335–336, 1998

    1998

  17. [17]

    Can multimodal large language models truly perform multimodal in-context learning? In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6000–6010

    Shuo Chen, Zhen Han, Bailan He, Jianzhe Liu, Mark Buckley, Yao Qin, Philip Torr, Volker Tresp, and Jindong Gu. Can multimodal large language models truly perform multimodal in-context learning? In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6000–6010. IEEE, 2025

    2025

  18. [18]

    Evaluating chatgpt-4o for ophthalmic image interpretation: from in-context learning to code-free clinical tool generation.Informatics and Health, 2(2):158– 169, 2025

    Joon Yul Choi and Tae Keun Yoo. Evaluating chatgpt-4o for ophthalmic image interpretation: from in-context learning to code-free clinical tool generation.Informatics and Health, 2(2):158– 169, 2025

    2025

  19. [19]

    Gemini embedding 2: Our first natively multimodal embedding model, Mar 2026

    Min Choi and Tom Duerig. Gemini embedding 2: Our first natively multimodal embedding model, Mar 2026. The Keyword Blog. Accessed: 2026-05-12

    2026

  20. [20]

    Scaling instruction-finetuned language models.Journal of Machine Learning Research, 25(70):1–53, 2024

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models.Journal of Machine Learning Research, 25(70):1–53, 2024

    2024

  21. [21]

    A coefficient of agreement for nominal scales.Educational and psychological measurement, 20(1):37–46, 1960

    Jacob Cohen. A coefficient of agreement for nominal scales.Educational and psychological measurement, 20(1):37–46, 1960

    1960

  22. [22]

    Qlora: Efficient finetuning of quantized llms.Advances in neural information processing systems, 36:10088– 10115, 2023

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms.Advances in neural information processing systems, 36:10088– 10115, 2023

    2023

  23. [23]

    A survey on in-context learning

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Baobao Chang, et al. A survey on in-context learning. InProceedings of the 2024 conference on empirical methods in natural language processing, pages 1107–1128, 2024

    2024

  24. [24]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2021

    Pith/arXiv arXiv 2010

  25. [25]

    Sedigheh Eslami, Christoph Meinel, and Gerard De Melo. Pubmedclip: How much does clip benefit visual question answering in the medical domain? InFindings of the Association for Computational Linguistics: EACL 2023, pages 1181–1193, 2023. 23

    2023

  26. [26]

    In-context learning enables multimodal large language models to classify cancer pathology images.Nature Communications, 15(1):10104, 2024

    Dyke Ferber, Georg Wölflein, Isabella C Wiest, Marta Ligero, Srividhya Sainath, Narmin Ghaffari Laleh, Omar SM El Nahhas, Gustav Müller-Franzes, Dirk Jäger, Daniel Truhn, et al. In-context learning enables multimodal large language models to classify cancer pathology images.Nature Communications, 15(1):10104, 2024

    2024

  27. [27]

    Statistical methods for research workers

    Ronald Aylmer Fisher. Statistical methods for research workers. InBreakthroughs in statistics: Methodology and distribution, pages 66–70. Springer, 1970

    1970

  28. [28]

    The rise of small language models in healthcare: A comprehensive survey.arXiv preprint arXiv:2504.17119, 2025

    Muskan Garg, Shaina Raza, Shebuti Rayana, Xingyi Liu, and Sunghwan Sohn. The rise of small language models in healthcare: A comprehensive survey.arXiv preprint arXiv:2504.17119, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    Gemma 4: Our most capable open models to date

    Gemma Team, Google DeepMind. Gemma 4: Our most capable open models to date. Google Blog, 2026. Accessed: 2026-04-15

    2026

  30. [30]

    Recognition of ischaemia and infection in diabetic foot ulcers: Dataset and techniques

    Manu Goyal, Neil D Reeves, Satyan Rajbhandari, Naseer Ahmad, Chuan Wang, and Moi Hoon Yap. Recognition of ischaemia and infection in diabetic foot ulcers: Dataset and techniques. Computers in biology and medicine, 117:103616, 2020

    2020

  31. [31]

    Domain adaptation for medical image analysis: a survey.IEEE Transactions on Biomedical Engineering, 69(3):1173–1185, 2021

    Hao Guan and Mingxia Liu. Domain adaptation for medical image analysis: a survey.IEEE Transactions on Biomedical Engineering, 69(3):1173–1185, 2021

    2021

  32. [32]

    Deep residual learning for im- age recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for im- age recognition. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016

    2016

  33. [33]

    Training compute-optimal large language models.arXiv preprint arXiv:2203.15556, 10, 2022

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, DDL Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models.arXiv preprint arXiv:2203.15556, 10, 2022

    Pith/arXiv arXiv 2022

  34. [34]

    Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022

    2022

  35. [35]

    Many-shot in-context learning in multimodal foundation models.arXiv preprint arXiv:2405.09798, 2024

    Yixing Jiang, Jeremy Irvin, Ji Hun Wang, Muhammad Ahmed Chaudhry, Jonathan H Chen, and Andrew Y Ng. Many-shot in-context learning in multimodal foundation models.arXiv preprint arXiv:2405.09798, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Scaling laws for neural language models.arXiv preprint arXiv:2001.08361, 2020

    JaredKaplan, SamMcCandlish, TomHenighan, TomBBrown, BenjaminChess, RewonChild, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprint arXiv:2001.08361, 2020

    Pith/arXiv arXiv 2001

  37. [37]

    Exploring the role of diversity in example selection for in-context learning

    Janak Kapuriya, Manit Kaushik, Debasis Ganguly, and Sumit Bhatia. Exploring the role of diversity in example selection for in-context learning. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2962– 2966, 2025

    2025

  38. [38]

    Investigating the impact of class-dependent label noise in medical image classification

    Bidur Khanal, SM Kamrul Hasan, Bishesh Khanal, and Cristian A Linte. Investigating the impact of class-dependent label noise in medical image classification. InProceedings of SPIE– the International Society for Optical Engineering, volume 12464, page 1246437, 2023

    2023

  39. [39]

    Learning to select visual in-context demonstrations

    Eugene Lee, Yu-Chi Lin, and Jiajie Diao. Learning to select visual in-context demonstrations. arXiv preprint arXiv:2603.26775, 2026. 24

    arXiv 2026

  40. [40]

    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. InInternational conference on machine learning, pages 19730–19742. PMLR, 2023

    2023

  41. [41]

    How to configure good in- context sequence for visual question answering

    Li Li, Jiawei Peng, Huiyi Chen, Chongyang Gao, and Xu Yang. How to configure good in- context sequence for visual question answering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26710–26720, 2024

    2024

  42. [42]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892–34916, 2023

    2023

  43. [43]

    Aconvnetforthe2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. Aconvnetforthe2020s. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976–11986, 2022

    2022

  44. [44]

    Small language models: Survey, measurements, and insights.arXiv preprint arXiv:2409.15790, 2024

    Zhenyan Lu, Xiang Li, Dongqi Cai, Rongjie Yi, Fangming Liu, Xiwen Zhang, Nicholas D Lane, and Mengwei Xu. Small language models: Survey, measurements, and insights.arXiv preprint arXiv:2409.15790, 2024

    Pith/arXiv arXiv 2024

  45. [45]

    Yang Luo, Zangwei Zheng, Zirui Zhu, and Yang You. How does the textual information affect the retrieval of multimodal in-context learning? InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 5321–5335, 2024

    2024

  46. [46]

    Mul- timodal machine learning to predict surgical site infection with healthcare workload impact assessment.npj Digital Medicine, 8(1):121, 2025

    Kenneth A McLean, Alessandro Sgrò, Leo R Brown, Louis F Buijs, Katie E Mountain, Cather- ine A Shaw, Thomas M Drake, Riinu Pius, Stephen R Knight, Cameron J Fairfield, et al. Mul- timodal machine learning to predict surgical site infection with healthcare workload impact assessment.npj Digital Medicine, 8(1):121, 2025

    2025

  47. [47]

    Note on the sampling error of the difference between correlated proportions or percentages.Psychometrika, 12(2):153–157, 1947

    Quinn McNemar. Note on the sampling error of the difference between correlated proportions or percentages.Psychometrika, 12(2):153–157, 1947

    1947

  48. [48]

    Arterial vs venous ulcers: What’s the difference?, 2023

    Medline Industries. Arterial vs venous ulcers: What’s the difference?, 2023. Accessed: 2026- 06-27

    2023

  49. [49]

    Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. Rethinking the role of demonstrations: What makes in-context learning work? InProceedings of the 2022 conference on empirical methods in natural language pro- cessing, pages 11048–11064, 2022

    2022

  50. [50]

    Ministral 3.arXiv preprint arXiv:2601.08584, 2026

    Mistral AI. Ministral 3.arXiv preprint arXiv:2601.08584, 2026

    Pith/arXiv arXiv 2026

  51. [51]

    Med-flamingo: a multimodal medical few-shot learner

    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. InMachine learning for health (ML4H), pages 353–367. PMLR, 2023

    2023

  52. [52]

    Multimodal large language models in medical imaging: current state and future directions.Korean Journal of Radiology, 26(10):900, 2025

    Yoojin Nam, Dong Yeong Kim, Sunggu Kyung, Jinyoung Seo, Jeong Min Song, Jimin Kwon, Jihyun Kim, Wooyoung Jo, Hyungbin Park, Jimin Sung, et al. Multimodal large language models in medical imaging: current state and future directions.Korean Journal of Radiology, 26(10):900, 2025. 25

    2025

  53. [53]

    Can generalist foundation models outcompete special-purpose tuning? case study in medicine.arXiv preprint arXiv:2311.16452, 2023

    Harsha Nori, Yin Tat Lee, Sheng Zhang, Dean Carignan, Richard Edgar, Nicolo Fusi, Nicholas King, Jonathan Larson, Yuanzhi Li, Weishung Liu, et al. Can generalist foundation models outcompete special-purpose tuning? case study in medicine.arXiv preprint arXiv:2311.16452, 2023

    Pith/arXiv arXiv 2023

  54. [54]

    Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khali- dov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    Pith/arXiv arXiv 2023

  55. [55]

    Traininglanguagemodelsto follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, ChongZhang, SandhiniAgarwal, KatarinaSlama, AlexRay, etal. Traininglanguagemodelsto follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

    2022

  56. [56]

    Integrated image and location analysis for wound classification: a deep learning approach.Scientific Reports, 14(1):7043, 2024

    YashPatel, TirthShah, MrinalKantiDhar, TaiyuZhang, JeffreyNiezgoda, SandeepGopalakr- ishnan, and Zeyun Yu. Integrated image and location analysis for wound classification: a deep learning approach.Scientific Reports, 14(1):7043, 2024

    2024

  57. [57]

    Fine-tuning smaller language models for question answering over financial documents

    Karmvir Singh Phogat, Sai Akhil Puranam, Sridhar Dasaratha, Chetan Harsha, and Shashishekar Ramakrishna. Fine-tuning smaller language models for question answering over financial documents. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 10528–10548, 2024

    2024

  58. [58]

    Qwen3.5: Towards native multimodal agents, February 2026

    Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026

    2026

  59. [59]

    Learning transferable visual models from natural language supervision

    AlecRadford, JongWookKim, ChrisHallacy, AdityaRamesh, GabrielGoh, SandhiniAgarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021

    2021

  60. [60]

    Fully automated wound tissue segmentation using deep learning on mobile devices: Cohort study.JMIR mHealth and uHealth, 10(4):e36977, 2022

    Dhanesh Ramachandram, Jose Luis Ramirez-GarciaLuna, Robert DJ Fraser, Mario Aurelio Martínez-Jiménez, Jesus E Arriaga-Caballero, and Justin Allport. Fully automated wound tissue segmentation using deep learning on mobile devices: Cohort study.JMIR mHealth and uHealth, 10(4):e36977, 2022

    2022

  61. [61]

    Multiclass wound image classification using an ensemble deep cnn-based classifier.Computers in Biology and Medicine, 134:104536, 2021

    Behrouz Rostami, DM Anisuzzaman, Chuanbo Wang, Sandeep Gopalakrishnan, Jeffrey Niez- goda, and Zeyun Yu. Multiclass wound image classification using an ensemble deep cnn-based classifier.Computers in Biology and Medicine, 134:104536, 2021

    2021

  62. [62]

    Detect-and-segment: A deep learning approach to automate wound image segmentation.Informatics in Medicine Unlocked, 29:100884, 2022

    Gaetano Scebba, Jia Zhang, Sabrina Catanzaro, Carina Mihai, Oliver Distler, Martin Berli, and Walter Karlen. Detect-and-segment: A deep learning approach to automate wound image segmentation.Informatics in Medicine Unlocked, 29:100884, 2022

    2022

  63. [63]

    In-context learning for label-efficient cancer image classification in oncology.Informatics in Medicine Unlocked, page 101683, 2025

    Mobina Shrestha, Bishwas Mandal, Vishal Mandal, Asis Shrestha, and Amir Babu Shrestha. In-context learning for label-efficient cancer image classification in oncology.Informatics in Medicine Unlocked, page 101683, 2025

    2025

  64. [64]

    Oriane Siméoni, Huy V Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, et al. Dinov3. arXiv preprint arXiv:2508.10104, 2025. 26

    Pith/arXiv arXiv 2025

  65. [65]

    Medical multimodal foundation models in clinical diagnosis and treatment: Applications, challenges, and future directions.Artificial Intelligence in Medicine, page 103265, 2025

    Kai Sun, Siyan Xue, Fuchun Sun, Haoran Sun, Yu Luo, Ling Wang, Siyuan Wang, Na Guo, Lei Liu, Tian Zhao, et al. Medical multimodal foundation models in clinical diagnosis and treatment: Applications, challenges, and future directions.Artificial Intelligence in Medicine, page 103265, 2025

    2025

  66. [66]

    Visual-language foundation models in medical imaging: A systematic review and meta-analysis of diagnostic and analytical applications

    Yiyao Sun, Xinran Wen, Yan Zhang, Lijun Jin, Chunna Yang, Qianhui Zhang, Mingchen Jiang, Zhaoyang Xu, Wei Guo, Juan Su, et al. Visual-language foundation models in medical imaging: A systematic review and meta-analysis of diagnostic and analytical applications. Computer Methods and Programs in Biomedicine, 268:108870, 2025

    2025

  67. [67]

    Visualpromptselectionforin-contextlearningsegmentation.arXiv preprint arXiv:2407.10233, 2024

    Wei Suo, Lanqing Lai, Mengyang Sun, Hanwang Zhang, Peng Wang, and Yanning Zhang. Visualpromptselectionforin-contextlearningsegmentation.arXiv preprint arXiv:2407.10233, 2024

    Pith/arXiv arXiv 2024

  68. [68]

    Efficientnetv2: Smaller models and faster training.arXiv preprint arXiv:2104.00298, 2021

    Mingxing Tan and Quoc V Le. Efficientnetv2: Smaller models and faster training.arXiv preprint arXiv:2104.00298, 2021

    Pith/arXiv arXiv 2021

  69. [69]

    Multimodal few-shot learning with frozen language models.Advances in Neural Information Processing Systems, 34:200–212, 2021

    Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models.Advances in Neural Information Processing Systems, 34:200–212, 2021

    2021

  70. [70]

    A survey on small language models

    Chien Van Nguyen, Xuan Shen, Ryan Aponte, Yu Xia, Samyadeep Basu, Zhengmian Hu, Jian Chen, Mihir Parmar, Sasidhar Kunapuli, Joe Barrow, et al. A survey on small language models. InProceedings of the 15th International Conference on Recent Advances in Natural Language Processing-Natural Language Processing in the Generative AI Era, pages 807–821, 2025

    2025

  71. [71]

    A unified framework for automatic wound segmentation and analysis with deep convolutional neural networks

    Changhan Wang, Xinchen Yan, Max Smith, Kanika Kochhar, Marcie Rubin, Stephen M War- ren, James Wrobel, and Honglak Lee. A unified framework for automatic wound segmentation and analysis with deep convolutional neural networks. In2015 37th annual international con- ference of the ieee engineering in medicine and biology society (EMBC), pages 2415–2418. IEEE, 2015

    2015

  72. [72]

    Fully automatic wound segmenta- tion with deep convolutional neural networks.Scientific reports, 10(1):21897, 2020

    Chuanbo Wang, DM Anisuzzaman, Victor Williamson, Mrinal Kanti Dhar, Behrouz Rostami, Jeffrey Niezgoda, Sandeep Gopalakrishnan, and Zeyun Yu. Fully automatic wound segmenta- tion with deep convolutional neural networks.Scientific reports, 10(1):21897, 2020

    2020

  73. [73]

    A com- prehensive survey on deep active learning in medical image analysis.Medical Image Analysis, 95:103201, 2024

    Haoran Wang, Qiuye Jin, Shiman Li, Siyu Liu, Manning Wang, and Zhijian Song. A com- prehensive survey on deep active learning in medical image analysis.Medical Image Analysis, 95:103201, 2024

    2024

  74. [74]

    Medclip: Contrastive learning from unpaired medical images and text

    Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3876–3887, 2022

    2022

  75. [75]

    Deep learning with noisy labels in medical prediction problems: a scoping review.Journal of the American Medical Informatics Association, 31(7):1596–1607, 2024

    Yishu Wei, Yu Deng, Cong Sun, Mingquan Lin, Hongmei Jiang, and Yifan Peng. Deep learning with noisy labels in medical prediction problems: a scoping review.Journal of the American Medical Informatics Association, 31(7):1596–1607, 2024

    2024

  76. [76]

    A uni- fied framework for automatic detection of wound infection with artificial intelligence.Applied Sciences, 10(15):5353, 2020

    Jin-Ming Wu, Chia-Jui Tsai, Te-Wei Ho, Feipei Lai, Hao-Chih Tai, and Ming-Tsan Lin. A uni- fied framework for automatic detection of wound infection with artificial intelligence.Applied Sciences, 10(15):5353, 2020. 27

    2020

  77. [77]

    Retrieving counterfactuals improves visual in-context learning.arXiv preprint arXiv:2603.16737, 2026

    Guangzhi Xiong, Sanchit Sinha, Zhenghao He, and Aidong Zhang. Retrieving counterfactuals improves visual in-context learning.arXiv preprint arXiv:2603.16737, 2026

    arXiv 2026

  78. [78]

    On-device language models: A comprehensive review.arXiv preprint arXiv:2409.00088, 2024

    Jiajun Xu, Zhiyuan Li, Wei Chen, Qun Wang, Xin Gao, Qi Cai, and Ziyuan Ling. On-device language models: A comprehensive review.arXiv preprint arXiv:2409.00088, 2024

    Pith/arXiv arXiv 2024

  79. [79]

    Feature extraction based machine learning for human burn diagnosis from burn images.IEEE journal of translational engineering in health and medicine, 7:1–7, 2019

    DP Yadav, Ashish Sharma, Madhusudan Singh, and Ayush Goyal. Feature extraction based machine learning for human burn diagnosis from burn images.IEEE journal of translational engineering in health and medicine, 7:1–7, 2019

    2019

  80. [80]

    A survey of wound image analysis using deep learning: classification, detection, and segmentation.IEEE Access, 10:79502–79515, 2022

    Ruyi Zhang, Dingcheng Tian, Dechao Xu, Wei Qian, and Yudong Yao. A survey of wound image analysis using deep learning: classification, detection, and segmentation.IEEE Access, 10:79502–79515, 2022

    2022

Showing first 80 references.