CLARIFY: A Specialist-Generalist Framework for Accurate and Lightweight Dermatological Visual Question Answering

Aranya Saha; Ismam Nur Swapnil; Mohammad Ariful Haque; Tanvir Ahmed Khan

REVIEW 3 major objections 4 minor 1 cited by

CLARIFY: A Specialist-Generalist Framework for Accurate and Lightweight Dermatological Visual Question Answering

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper proposes CLARIFY, a two-module system for dermatology visual question answering, and reports 82.1% diagnostic accuracy, 18 points above the best fine-tuned VLM, with lower memory use and latency.

desk verdict The modular architecture and dataset are worth a look, but the headline accuracy claim is a DINOv2-vs-VLM comparison on 39 images, not a validation of the framework. read the letter →

arxiv 2508.18430 v1 pith:C3WIABG5 submitted 2025-08-25 cs.CV cs.AI

Aranya Saha , Tanvir Ahmed Khan , Ismam Nur Swapnil , Mohammad Ariful Haque This is my paper

classification cs.CVcs.AI

keywords vision-languagemodelsdermatologyvisualquestionansweringspecialist-generalistframeworkretrieval-augmentedgenerationknowledgegraphmodelcompressionmedicalAI

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

CLARIFY is a modular system for dermatology visual question answering that separates perception from reasoning. A lightweight fine-tuned image classifier (the Specialist) produces the diagnosis, and a compressed conversational VLM (the Generalist) explains it, guided by the predicted label and by knowledge-graph facts retrieved about that label. On the authors' curated dataset, this division of labor yields 82.1% diagnostic accuracy versus 51-64% for end-to-end fine-tuned VLMs, while lowering VRAM by at least 20% and latency by at least 5%. The central idea is that forcing one monolithic VLM to both classify and chat causes both accuracy and fluency to suffer; routing each task to a model built for it avoids that trade-off.

What carries the argument

The load-bearing mechanism is the guided prompt built from three sources: the Specialist's class prediction, knowledge-graph facts retrieved for that class, and the user's original question. The Specialist is a DINOv2 image classifier fine-tuned for the eight disease classes; the Generalist is a structurally pruned VLM (LLaVA-1.5-7B or Qwen2.5-VL) whose layers are removed by an importance-scoring procedure. The KG-RAG module embeds a Wikipedia-derived dermatology knowledge graph and retrieves relevant entities at inference, so the VLM's answer is constrained both by the predicted label and by retrieved clinical facts.

What would settle it

Evaluate CLARIFY and the same two-stage fine-tuned Qwen-VL-7B baseline on an independent held-out set of at least 300 dermatology images with balanced classes. If the baseline matches or exceeds CLARIFY's accuracy, or the 18-point gap does not reproduce, the claimed advantage fails. A cheaper check is to repeatedly resample or relabel the existing 39-image test set and count how often the gap survives.

Watch

Extended reading notes

Core claim

The paper's central claim is that a Specialist-Generalist architecture outperforms a single fine-tuned VLM on the dermatological VQA task in both accuracy and efficiency. The Specialist—a DINOv2 classifier fine-tuned on 1,737 training images—always makes the diagnosis, which is then injected into the prompt of a pruned Qwen or LLaVA VLM together with KG-retrieved facts. This raises diagnostic accuracy to 82.1% from 64.1% for the best fine-tuned baseline, and the gain is not a model-quality effect: because the diagnosis comes from the Specialist, CLARIFY's accuracy is identical for every VLM backbone. The authors position this as evidence that perception and reasoning should be decoupled in m

Load-bearing premise

The whole accuracy comparison rests on a held-out test set of 39 images (2 to 9 per class); with a gap of just seven images between CLARIFY and the best baseline, a few different images or labels could change the headline conclusion.

Editorial extensions

If this is right

Diagnostic accuracy stops depending on the VLM chosen, so the framework can swap in a smaller or cheaper Generalist without losing diagnosis quality.
Two-stage fine-tuning of a monolithic VLM is not necessary for this task, avoiding the conversational degradation and catastrophic forgetting that the paper documents in baselines.
Since the diagnosis comes from a lightweight classifier, the model footprint can shrink enough to make on-premise deployment feasible in resource-constrained clinical settings.
KG-grounded retrieval makes the final explanation traceable to external facts about the predicted disease, which supports auditability of the system's answers.
The same decoupling blueprint can be transferred to other visually demanding medical specialties, because the design does not depend on the specific disease classes used here.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the 18-point accuracy gap holds on larger test sets, the practical implication is that fine-tuning generalist VLMs for diagnostic classification is mis-targeted; investment should go to specialist classifiers plus retrieval-grounded explanation.
A direct ablation the paper does not perform would separate how much of the conversational quality comes from the Specialist's label injection versus the KG-RAG facts; the current evaluation conflates those two components.
The framework's accuracy ceiling is set entirely by the Specialist classifier, so future gains in diagnostic performance depend on improving classifier calibration and coverage rather than on the VLM.
Replacing the Wikipedia-derived knowledge graph with a clinically validated ontology should measurably improve factual precision of explanations, since the paper acknowledges the current source is not authoritative but does not test that replacement.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The thing to know about CLARIFY is that the 18-point accuracy gain has almost nothing to do with the framework. The authors state in Section V-C that the diagnostic accuracy is the same for all VLMs because the diagnosis comes from the Specialist. So the headline number is simply the fine-tuned DINOv2 classifier's accuracy, compared to fine-tuned VLMs. The Generalist and KG-RAG only condition the explanation, not the label. That does not mean the paper is worthless—it means the central claim is misattributed.

What is genuinely useful: the modular design is a sensible engineering pattern for low-resource deployment—specialist classifier for perception, compressed VLM for conversation, KG retrieval for grounding. The paper ships a dataset (Small-Derma-VQA) with 1,776 images and 14k QA pairs, which is small but may be useful for future work. The pruning analysis in Table II is clearly presented, and the authors are transparent about the Wikipedia-derived KG not being clinically authoritative. They also acknowledge the small dataset in the Limitations section, which is more honesty than many papers.

The problems are in the evaluation. The test set is 39 images, 2-9 per class. The reported gap is 7 images (32 vs 25 correct). A Fisher exact test is not significant, and the paper reports no confidence intervals or repeats. That would be a serious weakness even if the number were genuinely about the full framework. The efficiency claim is measured on the pruned VLM alone, not the full pipeline; adding the Specialist and retrieval will change both VRAM and latency. Also, for the 3B backbone, the 7-layer pruned model reaches 19.4% VRAM reduction and 4.9% latency reduction—just short of the advertised 'at least 20% and 5%' unless you average across backbones, which is a charitable reading. The conversational evaluation uses LLM-as-judge with no human validation, and the baseline and CLARIFY use different fine-tuning protocols, so the comparison is confounded. No ablation isolates the contribution of the Specialist prompt or the KG-RAG.

Who is this for? People building low-resource medical VQA systems might get useful engineering ideas. The dataset could be a starting point, though it is too small for strong conclusions. The paper is not a valid demonstration that Specialist-Generalist improves diagnostic accuracy; it is a valid demonstration that a fine-tuned DINOv2 beats fine-tuned VLMs on a small dermatology test set.

I would send it to peer review rather than desk reject, because the architecture is clearly described and the dataset release is a concrete artifact. But a competent reviewer should flag the attribution problem and the absence of statistical support, and the authors will need a substantially larger evaluation—ideally a proper held-out set with significance testing and an ablation—before the headline claim can stand. My own verdict leans reject as written, but this is a fixable paper.

Referee Report

3 major / 4 minor

Summary. CLARIFY is a proposed specialist–generalist framework for dermatological VQA. A fine-tuned DINOv2 image classifier (the Specialist) produces the disease label, which is injected into the prompt of a structurally pruned VLM (the Generalist) together with knowledge-graph-retrieved context (KG-RAG). The Generalist then generates a conversational explanation. The paper claims an 18-point improvement in diagnostic accuracy over the best fine-tuned VLM baseline (82.1% vs. 64.1%), and at least 20% VRAM and 5% latency reductions. Evaluation is on a curated dataset of 1,776 DermNet images, of which 39 are used as the test set. The paper also reports LLM-as-a-judge conversational quality scores for pruned and unpruned backbones.

Significance. If the claims were supported, the modular specialist–generalist design would be a practical contribution for resource-constrained medical VQA, and the Small-Derma-VQA dataset could be a useful community resource. The paper is also honest in stating several limitations in Section VII. However, the central accuracy and efficiency claims are not supported by the experiments as designed: diagnostic accuracy is definitionally the Specialist's accuracy, the test set is too small for the claimed effect, and efficiency is measured only on the VLM component, not on the end-to-end pipeline. These are load-bearing issues, not presentation concerns.

major comments (3)

[Section V-C, Table III] The diagnostic accuracy of CLARIFY is definitionally the Specialist's accuracy. The paper states: 'The diagnostic accuracy in CLARIFY framework is same for all the VLMs because the diagnosis comes from the specialist module.' Therefore the Generalist and KG-RAG cannot change the predicted class, and the reported 82.1% vs. 64.1% comparison is a DINOv2-vs-VLM classifier comparison, not evidence for the Specialist–Generalist integration. No experiment reports Specialist-only accuracy, and no ablation removes the Generalist or KG-RAG from the diagnostic path. The abstract and Section VI attribute the 18-point gain to 'synergy' and the 'hierarchical design'; these claims are unsupported by the experiments as designed.
[Section IV-A, Table I, Section V-C] The test set contains only 39 images, with 2–9 images per class. The headline gap (82.1% = 32/39 vs. 64.1% = 25/39) is a difference of seven images. The paper reports no confidence intervals, significance tests, or repeated runs. A Fisher exact test on the resulting 2x2 table is not significant at the 5% level (p ≈ 0.12), so the 18-point improvement is within sampling noise. Section VII itself concedes that the small dataset 'limits the generalizability and robustness of the results.' The central accuracy claim is therefore not established.
[Section V-B, Table II, and Abstract] The efficiency claim is not measured on the CLARIFY pipeline. Table II reports VRAM and latency of the pruned VLM backbones alone; the full system in Section IV-E additionally runs the DINOv2 Specialist and KG-RAG semantic search, whose costs are excluded. For the configuration used in Section V-D (seven layers removed), Qwen2.5-3B gives 5.8 GB VRAM (a 19.4% reduction) and 137 ms/token (a 4.9% reduction), both below the advertised 'at least 20% and 5%' thresholds unless one averages over the three backbones without saying so. The abstract's efficiency claim is not supported as stated.

minor comments (4)

[Section IV-D and Section VI] The knowledge graph is built from Wikipedia, which the paper acknowledges is not clinically authoritative; yet the conclusion describes responses as 'grounded in factual medical knowledge' and 'clinically sound.' Please temper these claims to match the stated limitation.
[Table II and Section V-D] The choice of seven pruned layers as the operating point should be justified beyond the sentence in Section V-D. Also define clearly whether 'Compression Ratio' refers to parameters removed, and report full-pipeline VRAM/latency for that operating point.
[Figure 8] Figure 8 is a single qualitative example and should be labeled as illustrative; it cannot support the quantitative accuracy claim.
[General] There are minor language issues (e.g., 'to user query' in the abstract, inconsistent hyphenation of 'fine-tuned'). A careful proofread is recommended.

Circularity Check

1 steps flagged · score 6.0 of 10

CLARIFY's diagnostic accuracy is defined as the Specialist's accuracy, so the headline 18-point gain is a classifier comparison attributed to the framework.

self definitional [Section V-C, paragraph before Table III; Abstract]
"The diagnostic accuracy in CLARIFY framework is same for all the VLMs because the diagnosis comes from the specialist module. ... CLARIFY substantially outperforms all baselines, reaching an accuracy of 82.1%."

The paper defines CLARIFY's end-to-end diagnostic accuracy as the Specialist's classification output; the Generalist and KG-RAG are described as not affecting the diagnosis. Therefore the reported 18% improvement over fine-tuned VLMs is, by construction, the accuracy of the fine-tuned DINOv2 Specialist on the same 39-image test set. Since no Specialist-only accuracy or ablation of the Generalist/KG-RAG from the accuracy path is reported, the experiment equates CLARIFY diagnostic accuracy with the Specialist's accuracy and then credits the framework's hierarchical design for the gain. The framework's integration cannot be the cause of an outcome that is definitionally equal to one component's output.

full rationale

The central accuracy claim reduces by construction to a single component. Section V-C states that diagnostic accuracy is the same for all VLMs in CLARIFY because the diagnosis comes from the Specialist module. Thus the 82.1% figure in Table III is the fine-tuned DINOv2 Specialist's classification accuracy on the 39-image test set, not an emergent property of the Specialist-Generalist integration. The paper nevertheless presents this as validation of CLARIFY's hierarchical design and uses it for the headline 18-point improvement over the strongest fine-tuned VLM baseline. This is a self-definitional attribution: the measured output is defined as one module's output, so the claim that the framework improves diagnostic accuracy is not independently supported. Other claims—pruning efficiency (Table II) and conversational quality (Table IV)—are evaluated separately and do not reduce to this same definition, so the circularity is partial rather than total. The small test set is explicitly acknowledged as a limitation in Section VII; that is a statistical-validity issue, not a circularity, and does not affect this score. No load-bearing self-citation chain, imported uniqueness theorem, or ansatz-smuggling via citation was found.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the fitted Specialist weights, the post hoc choice of 7 pruned layers, and three domain assumptions about the test set, the LLM judges, and the Wikipedia-derived knowledge graph. No new physical or ontological entities are introduced.

free parameters (2)

DINOv2 fine-tuned weights
The diagnostic accuracy of CLARIFY is exactly the accuracy of this fitted Specialist; no uncertainty, confidence intervals, or calibration are reported, yet it carries the framework's headline accuracy gain.
Number of pruned layers in Generalist = 7
Chosen post hoc from Table II as the operating point for the full CLARIFY evaluation; performance drops with more pruning, so the choice trades off reported accuracy and efficiency.

assumptions (3)

domain assumption The 39-image held-out test set is representative of the 8-class dermatology distribution and large enough to estimate diagnostic accuracy.
Invoked in Section V-C and Table I; per-class test counts range from 2 to 9 images, so the 82.1% vs 64.1% comparison is driven by seven images.
domain assumption LLM-as-a-judge scores with Qwen-2.5-32B and gpt-oss-20b reflect factual and conversational quality.
Used in Section V-A and V-D as the only quantitative measure of conversational quality; no human validation or inter-judge agreement is reported.
domain assumption DermNet-derived QA pairs are correct ground truth and the KG built from Wikipedia is sufficient to ground responses.
Data curation in Section IV-A and knowledge graph construction in Section IV-D; the paper itself concedes Wikipedia is not clinically authoritative.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CLARIFY: A Specialist-Generalist Framework for Accurate and Lightweight Dermatological Visual Question Answering." pith.science (2026). https://pith.science/paper/C3WIABG5

@misc{pith2026250818430,
  author       = {Pith},
  title        = {Pith review of: CLARIFY: A Specialist-Generalist Framework for Accurate and Lightweight Dermatological Visual Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C3WIABG5}},
  note         = {Machine review of arXiv:2508.18430}
}

read the original abstract

Vision-language models (VLMs) have shown significant potential for medical tasks; however, their general-purpose nature can limit specialized diagnostic accuracy, and their large size poses substantial inference costs for real-world clinical deployment. To address these challenges, we introduce CLARIFY, a Specialist-Generalist framework for dermatological visual question answering (VQA). CLARIFY combines two components: (i) a lightweight, domain-trained image classifier (the Specialist) that provides fast and highly accurate diagnostic predictions, and (ii) a powerful yet compressed conversational VLM (the Generalist) that generates natural language explanations to user queries. In our framework, the Specialist's predictions directly guide the Generalist's reasoning, focusing it on the correct diagnostic path. This synergy is further enhanced by a knowledge graph-based retrieval module, which grounds the Generalist's responses in factual dermatological knowledge, ensuring both accuracy and reliability. This hierarchical design not only reduces diagnostic errors but also significantly improves computational efficiency. Experiments on our curated multimodal dermatology dataset demonstrate that CLARIFY achieves an 18\% improvement in diagnostic accuracy over the strongest baseline, a fine-tuned, uncompressed single-line VLM, while reducing the average VRAM requirement and latency by at least 20\% and 5\%, respectively. These results indicate that a Specialist-Generalist system provides a practical and powerful paradigm for building lightweight, trustworthy, and clinically viable AI systems.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Super-Generalist: Towards Comprehensive and Accurate Medical Image Understanding via Generalist-Specialist Synergy
cs.CV 2026-07 conditional novelty 6.0 of 10

Injecting multi-expert anatomy/lesion segmentation priors into vision–language alignment and calibrating text attention with lesion masks yields broad CT diagnosis plus specialist-level tumor performance and lesion grounding.

Reference graph

Works this paper leans on

38 extracted references · 10 canonical work pages · cited by 1 Pith paper

[1]

Visual instruction tuning,

H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” 2023. [Online]. Available: https://arxiv.org/abs/2304.08485

arXiv 2023
[2]

Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond,

J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P. Wang, J. Lin, C. Zhou, and J. Zhou, “Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond,” 2023. [Online]. Available: https://arxiv.org/abs/2308.12966

arXiv 2023
[3]

Vision-language models for vision tasks: A survey,

J. Zhang, J. Huang, S. Jin, and S. Lu, “Vision-language models for vision tasks: A survey,” IEEE transactions on pattern analysis and machine intelligence, vol. 46, no. 8, pp. 5625–5644, 2024

work page 2024
[4]

Medblip: Bootstrapping language-image pre-training from 3d medical images and texts,

Q. Chen, X. Hu, Z. Wang, and Y . Hong, “Medblip: Bootstrapping language-image pre-training from 3d medical images and texts,” 2023. [Online]. Available: https://arxiv.org/abs/2305.10799

arXiv 2023
[5]

Towards generalist biomedical ai,

T. Tu, S. Azizi, D. Driess, M. Schaekermann, M. Amin, P.-C. Chang, A. Carroll, C. Lau, R. Tanno, I. Ktena et al. , “Towards generalist biomedical ai,” Nejm Ai, vol. 1, no. 3, p. AIoa2300138, 2024

work page 2024
[6]

Multimodal large language models: A survey,

J. Wu, W. Gan, Z. Chen, S. Wan, and P. S. Yu, “Multimodal large language models: A survey,” 2023 IEEE International Conference on Big Data (BigData) , pp. 2247–2256, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:265351653

work page 2023
[7]

Enhancing medical image report generation through standard language models: Leveraging the power of llms in healthcare,

A. Santomauro, L. Portinale, and G. Leonardi, “Enhancing medical image report generation through standard language models: Leveraging the power of llms in healthcare,” in HC@AIxIA, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:266211540

work page 2023
[8]

Overcoming catastrophic forgetting in neural networks,

J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, D. Hassabis, C. Clopath, D. Kumaran, and R. Hadsell, “Overcoming catastrophic forgetting in neural networks,” Proceedings of the National Academy of Sciences , vol. 114, no. 13, pp. 3521–3526, 2017. [Online]. Available: htt...

doi:10.1073/pnas.1611835114 2017

Show all 38 references

[9]

An empirical study of catastrophic forgetting in large language models during continual fine-tuning,

Y . Luo, Z. Yang, F. Meng, Y . Li, J. Zhou, and Y . Zhang, “An empirical study of catastrophic forgetting in large language models during continual fine-tuning,” 2025. [Online]. Available: https://arxiv.org/abs/2308.08747

2025 arXiv
[10]

A simple and effective pruning approach for large language models,

M. Sun, Z. Liu, A. Bair, and J. Z. Kolter, “A simple and effective pruning approach for large language models,” 2024. [Online]. Available: https://arxiv.org/abs/2306.11695

2024 arXiv
[11]

Llm-pruner: On the structural pruning of large language models,

X. Ma, G. Fang, and X. Wang, “Llm-pruner: On the structural pruning of large language models,” 2023. [Online]. Available: https://arxiv.org/abs/2305.11627

2023 arXiv
[12]

Gptq: Accurate post-training quantization for generative pre-trained transformers,

E. Frantar, S. Ashkboos, T. Hoefler, and D. Alistarh, “Gptq: Accurate post-training quantization for generative pre-trained transformers,”
[13]

A comprehensive study on quantization techniques for large language models,

J. Lang, Z. Guo, and S. Huang, “A comprehensive study on quantization techniques for large language models,” 2024. [Online]. Available: https://arxiv.org/abs/2411.02530

2024 arXiv
[14]

Retrieval-augmented generation for knowledge-intensive nlp tasks,

P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. K ¨uttler, M. Lewis, W. tau Yih, T. Rockt ¨aschel, S. Riedel, and D. Kiela, “Retrieval-augmented generation for knowledge-intensive nlp tasks,” 2021. [Online]. Available: https://arxiv.org/abs/2005.11401

2021 arXiv
[15]

Unifying large language models and knowledge graphs: A roadmap,

S. Pan, L. Luo, Y . Wang, C. Chen, J. Wang, and X. Wu, “Unifying large language models and knowledge graphs: A roadmap,” IEEE Transactions on Knowledge and Data Engineering , vol. 36, no. 7, p. 3580–3599, Jul. 2024. [Online]. Available: http://dx.doi.org/10.1109/ TKDE.2024.3352100

2024
[16]

Improving retrieval-augmented generation in medicine with iterative follow-up questions,

G. Xiong, Q. Jin, X. Wang, M. Zhang, Z. Lu, and A. Zhang, “Improving retrieval-augmented generation in medicine with iterative follow-up questions,” 2024. [Online]. Available: https://arxiv.org/abs/2408.00727

2024 arXiv
[17]

Dermatology resource,

DermNet, “Dermatology resource,” https://dermnetnz.org, 2025, ac- cessed: 2025-08-19

2025
[18]

Learning transferable visual models from natural language supervision,

A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” 2021. [Online]. Available: https://arxiv.org/abs/2103.00020

2021 arXiv
[19]

Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation,

J. Li, D. Li, C. Xiong, and S. Hoi, “Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation,”
[20]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” 2015. [Online]. Available: https://arxiv.org/abs/1512.03385

2015 arXiv
[21]

An image is worth 16x16 words: Transformers for image recognition at scale,

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” 2021. [Online]. Available: https://arxi...

2021 arXiv
[22]

Dermatologist-level classification of skin cancer with deep neural networks,

A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, and S. Thrun, “Dermatologist-level classification of skin cancer with deep neural networks,” nature, vol. 542, no. 7639, pp. 115–118, 2017

2017
[23]

A survey on model compression and accel- eration for pretrained language models,

C. Xu and J. McAuley, “A survey on model compression and accel- eration for pretrained language models,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 37, no. 9, 2023, pp. 10 566– 10 575

2023
[24]

Shortened llama: A simple depth pruning for large language models,

B.-K. Kim, G. Kim, T.-H. Kim, T. Castells, S. Choi, J. Shin, and H.- K. Song, “Shortened llama: A simple depth pruning for large language models,” arXiv preprint arXiv:2402.02834 , vol. 11, 2024

2024 arXiv
[25]

Shortgpt: Layers in large language models are more redundant than you expect,

X. Men, M. Xu, Q. Zhang, B. Wang, H. Lin, Y . Lu, X. Han, and W. Chen, “Shortgpt: Layers in large language models are more redundant than you expect,” 2024. [Online]. Available: https://arxiv.org/abs/2403.03853

2024 arXiv
[26]

Medrag: Enhancing retrieval-augmented generation with knowledge graph-elicited reasoning for healthcare copilot,

X. Zhao, S. Liu, S.-Y . Yang, and C. Miao, “Medrag: Enhancing retrieval-augmented generation with knowledge graph-elicited reasoning for healthcare copilot,” 2025. [Online]. Available: https://arxiv.org/abs/ 2502.04413

2025 arXiv
[27]

Kg-rag: Bridging the gap between knowledge and creativity,

D. Sanmartin, “Kg-rag: Bridging the gap between knowledge and creativity,” 2024. [Online]. Available: https://arxiv.org/abs/2405.12035

2024 arXiv
[28]

Gregg and D

F. Gregg and D. Eder, “Dedupe,” https://github.com/dedupeio/dedupe, 2022, accessed: 2025-08-19

2022
[29]

Small-derma-vqa,

A. Saha, T. A. Khan, and I. N. Swapnil, “Small-derma-vqa,” 2025. [Online]. Available: https://www.kaggle.com/dsv/12845315

2025
[30]

Dinov2: Learning robust visual features without supervision,

M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby et al. , “Dinov2: Learning robust visual features without supervision,” arXiv preprint arXiv:2304.07193, 2023

2023 arXiv
[31]

Llm-pruner: On the structural pruning of large language models,

X. Ma, G. Fang, and X. Wang, “Llm-pruner: On the structural pruning of large language models,” Advances in neural information processing systems, vol. 36, pp. 21 702–21 720, 2023

2023
[32]

Kggen: Extracting knowledge graphs from plain text with language models,

B. Mo, K. Yu, J. Kazdan, P. Mpala, L. Yu, C. Cundy, C. Kanatsoulis, and S. Koyejo, “Kggen: Extracting knowledge graphs from plain text with language models,” arXiv preprint arXiv:2502.09956 , 2025

2025
[33]

Gemini: a family of highly capable multimodal models,

G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Millican et al., “Gemini: a family of highly capable multimodal models,”arXiv preprint arXiv:2312.11805, 2023

2023 arXiv
[34]

Dspy: Compiling declarative language model calls into self-improving pipelines,

O. Khattab, A. Singhvi, P. Maheshwari, Z. Zhang, K. Santhanam, S. Vardhamanan, S. Haq, A. Sharma, T. T. Joshi, H. Moazam et al. , “Dspy: Compiling declarative language model calls into self-improving pipelines,” arXiv preprint arXiv:2310.03714 , 2023

2023 arXiv
[35]

Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers,

W. Wang, F. Wei, L. Dong, H. Bao, N. Yang, and M. Zhou, “Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers,” 2020. [Online]. Available: https://arxiv. org/abs/2002.10957

2020 arXiv
[36]

A survey on llm-as-a-judge,

J. Gu, X. Jiang, Z. Shi, H. Tan, X. Zhai, C. Xu, W. Li, Y . Shen, S. Ma, H. Liu et al. , “A survey on llm-as-a-judge,” arXiv preprint arXiv:2411.15594, 2024. IX. B IOGRAPHY SECTION Aranya Saha received the B.Sc. degree in Electrical and Electronic Engineering, majoring in Comm...

2024 arXiv
[2022]

Available: https://arxiv.org/abs/2201.12086 10

[Online]. Available: https://arxiv.org/abs/2201.12086 10

arXiv
[2023]

Available: https://arxiv.org/abs/2210.17323

[Online]. Available: https://arxiv.org/abs/2210.17323

arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Visual instruction tuning,

H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” 2023. [Online]. Available: https://arxiv.org/abs/2304.08485

arXiv 2023

[2] [2]

Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond,

J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P. Wang, J. Lin, C. Zhou, and J. Zhou, “Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond,” 2023. [Online]. Available: https://arxiv.org/abs/2308.12966

arXiv 2023

[3] [3]

Vision-language models for vision tasks: A survey,

J. Zhang, J. Huang, S. Jin, and S. Lu, “Vision-language models for vision tasks: A survey,” IEEE transactions on pattern analysis and machine intelligence, vol. 46, no. 8, pp. 5625–5644, 2024

work page 2024

[4] [4]

Medblip: Bootstrapping language-image pre-training from 3d medical images and texts,

Q. Chen, X. Hu, Z. Wang, and Y . Hong, “Medblip: Bootstrapping language-image pre-training from 3d medical images and texts,” 2023. [Online]. Available: https://arxiv.org/abs/2305.10799

arXiv 2023

[5] [5]

Towards generalist biomedical ai,

T. Tu, S. Azizi, D. Driess, M. Schaekermann, M. Amin, P.-C. Chang, A. Carroll, C. Lau, R. Tanno, I. Ktena et al. , “Towards generalist biomedical ai,” Nejm Ai, vol. 1, no. 3, p. AIoa2300138, 2024

work page 2024

[6] [6]

Multimodal large language models: A survey,

J. Wu, W. Gan, Z. Chen, S. Wan, and P. S. Yu, “Multimodal large language models: A survey,” 2023 IEEE International Conference on Big Data (BigData) , pp. 2247–2256, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:265351653

work page 2023

[7] [7]

Enhancing medical image report generation through standard language models: Leveraging the power of llms in healthcare,

A. Santomauro, L. Portinale, and G. Leonardi, “Enhancing medical image report generation through standard language models: Leveraging the power of llms in healthcare,” in HC@AIxIA, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:266211540

work page 2023

[8] [8]

Overcoming catastrophic forgetting in neural networks,

J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, D. Hassabis, C. Clopath, D. Kumaran, and R. Hadsell, “Overcoming catastrophic forgetting in neural networks,” Proceedings of the National Academy of Sciences , vol. 114, no. 13, pp. 3521–3526, 2017. [Online]. Available: htt...

doi:10.1073/pnas.1611835114 2017