Pith.

Computer Vision and Pattern Recognition · arXiv:2611.20992 v1 · submitted 16 Nov 2026 · 62 references

MedJev: A System One Vision-Language Decision Model for Fast and Effective Medical Image Understanding

Fan Zhang, Hao Wu, Jinyang Wu, Hongyu Qu, Zheng Lian, Dongfu Yin, Qingsong Wen, Pheng-Ann Heng

The paper says

“Medical image understanding with vision-language models (VLMs) typically relies on autoregressively generating reasoning and answer tokens.”

more of the abstract

These unnecessary tokens make the inference cost scale with the response length and force the answer to be parsed from free-form text, incurring avoidable latency and decision errors. To this end, we introduce \textscMedJev, a system one vision-language decision model that maps medical images, questions, and candidate answers directly to decisions in a single multimodal forward pass, without any autoregressive generation process. For fast inference, \textscMedJev removes the upper language layers and the generative output pathway, while keeping the full vision pathway and all visual tokens. On this shortened backbone, a decision head scores each candidate by pretrained query–key matching and refines the scores with a lightweight multilayer perceptron (MLP) residual, coupling the pretrained matching structure with task-specific nonlinear discrimination. Built on Qwen3.5-4B, \textscMedJev-4B improves both accuracy and speed over the original model on each of four medical image classification and visual question answering benchmarks, with an absolute accuracy gain of \textbf7.93% and a \textbf{6.04×} end-to-end speedup. Further comparison with 20 cutting-edge models shows that \textscMedJev-4B achieves state-of-the-art performance and inference speed among 4B–9B general and medical VLMs, indicating that direct multimodal decision making is an effective route to fast and effective medical image understanding.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 17 Nov 2026 · read the full text (89,045 characters) · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up. MedJev turns a pretrained VLM into a fast, accurate medical decision model by removing autoregressive generation and scoring candidates with a residual head built on pretrained query-key matching, and the experiments back the claims. This line is wrong

7 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

The paper, first character to last

The reviewer was given the whole text, 89,045 characters. Not checked: figures, code, what the 62 references say.

7 minor objections

The paper says, at §1

Pith says: minor objection 1 of 7

The sentence 'and and longer traces help little once this grounding is weak' contains a duplicated 'and'.

The paper says, at §2.2

Pith says: minor objection 2 of 7

The sentence 'and and refines the pretrained query--key matching with a zero-initialized residual head' contains a duplicated 'and'.

show 5 more

The paper says, at §3.2

Pith says: minor objection 3 of 7

The depth probe figure (Fig. 2) is described but not visible. The caption says 'Balanced accuracy of the pretrained query-key match combined with a linear head over the marker states of each pretrained layer'. It would be helpful to include the numerical values or a table in the appendix for readers who cannot see the figure.

The paper says, at §4.3.4, Table 2

Pith says: minor objection 4 of 7

The latency column in Table 2 is described as 'over the three single-image benchmarks' but the table caption says 'Lat. the latency (ms) over the three single-image benchmarks'. This is reasonable but the main text should clarify why MedThinkVQA is excluded from the latency average (because it uses multiple images). The paper does explain this in the appendix but a brief note in the main text would help.

The paper says, at §4.3.5, Fig. 5

Pith says: minor objection 5 of 7

The case study figure (Fig. 5) is described but not visible. The text says 'Qwen3.5-4B spends 408.1 ms decoding ten tokens and names the bones of the forearm for an image of an ankle'. This is a qualitative example; it would be stronger to include a quantitative analysis of how often MedJev corrects backbone errors.

The paper says, at §A.3, Table A.1

Pith says: minor objection 6 of 7

The training configuration table lists 'LoRA Rank, Scaling, Dropout' as '16, 32, 0'. The dropout of 0 may lead to overfitting; a brief justification would be helpful.

The paper says, at §A.5, Fig. A.2

Pith says: minor objection 7 of 7

The calibration analysis (Fig. A.2) shows ECE of 6.18%. This is reasonable but the paper does not compare to the calibration of the backbone or other baselines. Adding such comparison would strengthen the claim that the confidence signal is usable.

What would settle it

If one were to train MedJev with the residual MLP removed (QKH-only variant) and find that its accuracy on any of the four benchmarks falls below the majority-class baseline, or if the depth probe showed that the marker states at layer 23 are not more discriminative than random for medical decisions, the central claim that pretrained matching transfers to the reduced architecture would be undermined.

from Pith's summary of the paper

How the review went

5 passes by deepseek-v4-flash · 17 Nov 2026

  1. Reader holds up confidence high
  2. Skeptic concern withdrawn: Pith's reviewer misread the paper 9 s · 31,317 tokens read, 725 written
  3. Referee 7 minor 14 s · 31,606 tokens read, 1,709 written
  4. Check against the paper kept 0 of 4 objections 18 s · 32,354 tokens read, 2,423 written
  5. Editor wrote the verdict above 2 min 23 s · 96,254 tokens read, 33,183 written

Not run on this paper: the simulated authors' defence.

Pith's passes raised 4 objections. A check against the paper's text kept 0 and withdrew 4 (4 Pith's reviewer misread the paper).

The review in full

The review in brief, machine-written

MedJev is a clean engineering contribution that does what it says. It takes a standard VLM (Qwen3.5-4B), removes the upper language layers and the generative output pathway, and replaces text generation with a single forward pass that scores candidate answers using the pretrained query-key projections plus a small MLP residual. The result is a model that is both more accurate (7.93% absolute gain on average across four benchmarks) and 6x faster than the backbone, and that beats 20 baselines in the 4B-9B range on both accuracy and speed. The experimental setup is thorough: four diverse benchmarks, 20 baselines, ablation studies, bootstrap confidence intervals, and careful latency measurements. The ablation shows that the depth reduction costs no accuracy and that the pretrained matching carries most of the decision, with the MLP residual adding a small but consistent improvement. The paper also checks that the model is not just reading candidate positions (order permutation tests) and that its confidence scores are usable (calibration plot). The main soft spot is that the method is evaluated on a single seed, so we do not know how much the results vary with initialization. The authors also assume that the pretrained query-key projections, trained on the full-depth backbone, transfer usefully to the reduced architecture; the depth probe and ablation support this, but the evidence is correlational rather than causal. These are minor concerns. The paper is clearly written, the code is promised, and the central claim is well supported. This is a solid piece of applied work that deserves a serious referee. I would accept it for peer review with the expectation that the authors release the code and consider multiple seeds in future work.

The stress test's concern

Withdrawn: Pith's reviewer misread the paper.

The paper says

“The depth reduction costs no accuracy under either head, raising the mean from 68.31\% to 68.70\% under QKH and from 68.93\% to 69.06\% under RDH while cutting about 16\,ms of latency.”

The referee's report

Not published. Pith's referee wrote this assessment before its objections were checked against the paper, and 3 of its 3 major objections did not survive that check.

Pith's reading of the paper (a machine's summary, not the authors')

The paper argues that the standard approach of using generative vision-language models (VLMs) for medical image tasks is unnecessarily slow and error-prone because it produces decisions as free-form text, requiring costly autoregressive decoding and fragile answer parsing. The authors propose MedJev, a 'system one' decision model that instead scores candidate answers directly in a single forward pass, eliminating decoding entirely. Built on Qwen3.5-4B, MedJev-4B removes the upper language layers and the generative output pathway, keeping all visual tokens, and scores candidates using the pretrained query-key matching of the last layer plus a lightweight MLP residual. On four medical benchmarks (HAM10000, PneumoniaMNIST, MedThinkVQA, PMC-VQA), MedJev-4B improves average accuracy by 7.93% over the original model while achieving a 6.04x end-to-end speedup, and it claims state-of-the-art accuracy and speed among 4B-9B VLMs. If correct, this demonstrates that direct multimodal decision-making can be both faster and more accurate than generation-based approaches for medical image tasks with fixed candidate sets.

What carries the argument. The key machinery is the residual decision head that reuses the pretrained query and key projections of the last transformer layer to compute a matching score between the decision marker and each candidate marker, then adds a nonlinear residual from a lightweight MLP (hidden dimension 256) that is initialized to zero so training starts from the pretrained matching. This is combined with language-side depth reduction that removes layers 24-31 of the 32-layer Qwen3.5-4B backbone, keeping only the first 23 layers and the query/key branches of layer 32, which shortens the path from visual evidence to decision.

Core claim. The paper claims that a vision-language model can be converted into a fast and accurate decision model for medical image understanding by (1) removing the upper language layers and the generative output pathway while keeping all visual tokens, and (2) scoring candidate answers using the pretrained query-key projections of the last layer, refined by a lightweight zero-initialized MLP residual. The resulting model, MedJev-4B, is claimed to achieve both higher accuracy and lower latency than the original generative model and all other 4B-9B VLMs tested on four medical benchmarks.

The premise it rests on. The paper assumes that the pretrained query and key projections from the last layer of Qwen3.5-4B, when applied to the output of layer 23 (after removing layers 24-31), provide a useful matching signal for medical decision-making, and that the attention geometry learned on general data transfers to medical candidate scoring after this architectural change.

If the paper is right (Pith's inferences, not the authors')

Medical image tasks with fixed candidate sets (diagnoses, findings, answer options) can be solved without any autoregressive generation, eliminating the cost of long reasoning traces and fragile answer parsing.

The latency of a decision becomes predictable and bounded by input length rather than response length, with no tail latency from long generations.

The confidence signal from the softmax over candidate scores can be used to defer uncertain cases to a clinician, with accuracy rising from 62.33% at full coverage to 92.67% on the most confident 14.4% of cases.

A single decision model can handle heterogeneous tasks (binary classification, 7-way diagnosis, multi-image VQA) without task-specific output heads, because the candidate set is provided as text in the input.

The approach suggests that the pretrained attention geometry of a VLM already encodes useful matching information for medical decisions, and that only a small nonlinear correction is needed to adapt it.

The depth probe analysis (Fig. 2) suggests that many VLMs may be overbuilt for decision tasks: the marker states stop becoming more discriminative well before the last layer, implying that the upper layers primarily serve generation, not representation quality for decisions.

The success of reusing pretrained query-key projections after depth reduction implies that the attention matching learned on general data transfers to medical candidate scoring, which could extend to other specialized domains with limited training data.

The 6.04x speedup is a lower bound on what is possible: the gap would grow for tasks requiring longer reasoning traces, and shrink for very short answers; the paper's choice of benchmarks with thinking disabled may understate the speed advantage in practice.

The comparison with other decision models (Table A2) shows that removing generation is necessary but not sufficient: the specific design of depth reduction plus residual matching matters, as other decision models with different architectures underperform despite sharing the non-generative interface.

Check this review

What was read

The passes were given all 89,045 characters of the paper's LaTeX source.

Fingerprint of that text (SHA-256): 0f116d478bfc30c39d3119c535a0f1d27a7635a5b3c394f2b302d3863b28b59b

Pith does not keep that text. Rebuilt from the paper's source or PDF by the same code, it has this fingerprint if it is the same text. From a copy of Pith's code:

PYTHONPATH=. python scripts/review_input_check.py 2611.20992

The passes, in the order they ran

  1. Reader deepseek-v4-flash · medium effort · $0.0040 · instructions of the Reader, fingerprint f9ead7300382…
  2. Skeptic deepseek-v4-flash · medium effort · $0.0030 · instructions of the Skeptic, fingerprint 8f88c0f61488…
  3. Referee deepseek-v4-flash · medium effort · $0.0032 · instructions of the Referee, fingerprint a852d99aea0d…
  4. Check against the paper deepseek-v4-flash · medium effort · $0.0034 · instructions of the Check against the paper, fingerprint 6c7a6883d551…
  5. Editor deepseek-v4-flash · medium effort · $0.010 · instructions of the Editor, fingerprint e7211f6a3d27…
  6. Circularity check deepseek-v4-flash · medium effort · $0.0035 · instructions of the Circularity check, fingerprint e9946e39d67b…
  7. Summary for the page deepseek-v4-flash · medium effort · temperature 0.2 · $0.0033 · instructions of the Summary for the page, fingerprint e4f56bf568e7…

The passes cost $0.030 in model calls together.

Version stamped on this review: pith-pipeline@v1.4.1-daily-deepseek-or-medium+deepseek/deepseek-v4-flash+p.d60f54226f00

Each link opens a pass's instructions as plain text, the exact bytes its fingerprint was taken over. Every stored set of instructions.

curl -s https://pith.science/methodology/prompts/f9ead730038216984cfb56d7cb34332bdf691828296a1de4245f3d6667d4bf56 | sha256sum

prints the fingerprint of the reader's instructions. The same works for every pass.

The receipt

The review's fingerprint was taken on 17 Nov 2026. The review you are reading still matches it.

Fingerprint of the review (SHA-256): 17b04c8cfe88afb0d4a6979024f443be91a339aa2b9e1c7a1c6744a3f205d231

curl -s 'https://pith.science/api/v1/verdicts/2611.20992?form=canonical' | sha256sum

prints the fingerprint of the review as it stands now. The rule, in four steps.

A matching fingerprint shows the review has not changed since that date. It does not show that the review is right.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, from the paper's LaTeX source.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Medical image understanding with vision-language models (VLMs) typically relies on autoregressively generating reasoning and answer tokens. These unnecessary tokens make the inference cost scale with the response length and force the answer to be parsed from free-form text, incurring avoidable latency and decision errors. To this end, we introduce \textscMedJev, a system one vision-language decision model that maps medical images, questions, and candidate answers directly to decisions in a single multimodal forward pass, without any autoregressive generation process. For fast inference, \textscMedJev removes the upper language layers and the generative output pathway, while keeping the full vision pathway and all visual tokens. On this shortened backbone, a decision head scores each candidate by pretrained query–key matching and refines the scores with a lightweight multilayer perceptron (MLP) residual, coupling the pretrained matching structure with task-specific nonlinear discrimination. Built on Qwen3.5-4B, \textscMedJev-4B improves both accuracy and speed over the original model on each of four medical image classification and visual question answering benchmarks, with an absolute accuracy gain of \textbf7.93% and a \textbf{6.04×} end-to-end speedup. Further comparison with 20 cutting-edge models shows that \textscMedJev-4B achieves state-of-the-art performance and inference speed among 4B–9B general and medical VLMs, indicating that direct multimodal decision making is an effective route to fast and effective medical image understanding.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-17

The authors' main claim, as the classifier picked it out 1.00

“Further comparison with 20 cutting-edge models shows that \textsc{MedJev}-4B achieves state-of-the-art performance and inference speed among 4B--9B general and medical VLMs, indicating that direct multimodal decision making is an effective route to fast and effective medical image understanding.”

How the main claim is stated: hedged possibility 0.62

What the abstract itself states, as probabilities: claims first 0.09 · states limitation 0.10 · human participants 0.07 · controlled comparison 0.87 · randomized assignment 0.01 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Medical image understanding with vision-language models (VLMs) typically relies on autoregressively generating reasoning and answer tokens. background 1.00 0.05 0.02 asserted 1.00
These unnecessary tokens make the inference cost scale with the response length and force the answer to be parsed from free-form text, incurring avoidable latency and decision errors. background 0.85 0.11 0.03 asserted 1.00
To this end, we introduce \textsc{MedJev}, a \emph{system one} vision-language decision model that maps medical images, questions, and candidate answers directly to decisions in a single multimodal forward pass, without any autoregressive generation process. contribution 0.97 0.90 0.18 asserted 0.99
For fast inference, \textsc{MedJev} removes the upper language layers and the generative output pathway, while keeping the full vision pathway and all visual tokens. method 1.00 0.87 0.03 asserted 1.00
On this shortened backbone, a decision head scores each candidate by pretrained query--key matching and refines the scores with a lightweight multilayer perceptron (MLP) residual, coupling the pretrained matching structure with task-specific nonlinear discrimination. method 1.00 0.84 0.04 asserted 0.99
Built on Qwen3.5-4B, \textsc{MedJev}-4B improves both accuracy and speed over the original model on each of four medical image classification and visual question answering benchmarks, with an absolute accuracy gain of \textbf{7.93\%} and a \textbf{6.04$\times$} end-to-end speedup. result 1.00 0.96 0.20 asserted 1.00
Further comparison with 20 cutting-edge models shows that \textsc{MedJev}-4B achieves state-of-the-art performance and inference speed among 4B--9B general and medical VLMs, indicating that direct multimodal decision making is an effective route to fast and effective medical image understanding. result 0.57 0.93 0.82 hedged possibility 0.62

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.92

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-17: 1 request, 5,984 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (6)
Figure 1 of the paper
Figure 1. (a) An example from PMC-VQA [55], where Qwen3.5-4B writes 1,024 tokens without selecting an option and MedJev-4B scores the candidates in one forward pass. (b) Per-request latency against generated tokens, thinking off. (c) Average accuracy and latency on four benchmarks, where MedJev-4B beats all 20 general and medical VLMs at the 4B–9B scale on both. Vision-language models (VLMs) have become a general interface for understanding medical images. Both general-purpose VLMs [3, 4, 24, 43, 58] and medical VLMs trained on biomedical image-text arXiv:2611.20992v1 [cs.CV] 16 Nov 2026
Figure 2 of the paper
Figure 2. Overview of MedJev. (a) The images, the question, and all candidates form a single prompt with learned markers, and every visual token is kept. (b) A reduced backbone returns the marker states. (c) Pretrained query–key matching and a shared MLP residual score the candidates in one pass. The example is a CT question of PMC-VQA [55] with the probabilities of MedJev-4B. pruning methods remove redundant deep layers of large language models [10, 27], yet the pruned models still generate their answers token by token. Decoding also leaves the cost of a single answer open, since a model…
Figure 3 of the paper
Figure 3. Depth probe analysis. Balanced ac￾curacy of the pretrained query-key match com￾bined with a linear head over the marker states of each pretrained layer, with the backbone frozen and all examples from the training split. A generative VLM runs its full language stack and the vocabulary projection at every decoding step, whereas a decision only needs the hidden states at the marker positions. MedJev exploits this asym￾metry. A probe of the frozen backbone confirms this, since the marker states stop becoming more discriminative well before the last layer (
Figure 4 of the paper
Figure 4. Ablation and significance. (a) Unweighted mean accuracy and (b) mean end-to-end latency over the four benchmarks, where DR is the depth reduction, QKH the query-key head, MLP the per-candidate MLP and RDH the residual decision head. (c) The accuracy gain of MedJev-4B over the generative Qwen3.5-4B, with its 95% interval from a paired bootstrap of 10,000 resamples. of every model on an NVIDIA H20 GPU at batch size one, including image loading, preprocessing, and model execution. Each baseline is timed in a separate session under the same protocol, and the latency of Qwen3.5-4B…
Figure 5 of the paper
Figure 5. Candidate order and hidden representations. (a) Accuracy under the native and three permuted orders. (b) The fate of a correct decision, kept or lost. (c, d) t-SNE of the decision marker before and after training and of the trained candidates, with the marginal distribution of each axis. 4.3.3. Candidate Order In
Figure 6 of the paper
Figure 6. One PMC-VQA case under both inter￾faces. The backbone decodes a sentence that names the wrong bones, while MedJev-4B scores all four candidates in one pass and picks the reference. In

Figures are the authors', reproduced from the paper.

References (62)
  1. [1] Joshua Ainslie, James Lee-Thorp, Michiel de Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. GQA: Training generalized multi-query transformer models from multi-head 11 MedJev checkpoints. InConference on Empirical Methods in Natural Language Processing (EMNLP), 202... on PithPith's reviewarXiv
  2. [2] Diogo Almeida. Introducing System One models & Jev. TypeSafe AI Blog, 2026. URLhttps: //typesafe.ai/blog/introducing-system-one-models-and-jev . Published September 15, 2026 on Pith
  3. [3] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-VL technical report.arXiv preprint arXiv:2511.21631, 2025. URLhttps://arxiv.org/abs/2511.21631 on PithPith's reviewarXiv
  4. [5] Junying Chen, Chi Gui, Ruyi Ouyang, Anningzhe Gao, Shunian Chen, Guiming Hardy Chen, Xidong Wang, Ruifei Zhang, Zhenyang Cai, Ke Ji, Guangjun Yu, Xiang Wan, and Benyou Wang. HuatuoGPT-Vision, towards injecting medical visual knowledge into multimodal LLMs at scale. arXiv prepr... on PithPith's reviewarXiv
  5. [6] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. InEuropean Conference on Computer Vision (ECCV), 2024. URL https://arxi... on PithPith's reviewarXiv
  6. [7] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling.arXiv preprint arXiv:2412.05271, 2024. URLhttps://arxiv.org/abs/2412.05271 on PithPith's reviewarXiv
  7. [8] Cloudflare. IntroducingClef: Ouropen-sourcedecisionmodels, andnewRLfine-tuningplatform. Blog post, 2026. URLhttps://blog.cloudflare.com/clef-decision-models/ on Pith
  8. [9] Jonathan St. B. T. Evans and Keith E. Stanovich. Dual-process theories of higher cognition: Advancing the debate.Perspectives on Psychological Science, 8(3):223–241, 2013. doi: 10.1177/ 1745691612460685. URLhttps://doi.org/10.1177/1745691612460685 on Pithdoi:10.1177/1745691612460685
  9. [10] Andrey Gromov, Kushal Tirumala, Hassan Shapourian, Paolo Glorioso, and Daniel A. Roberts. The unreasonable ineffectiveness of the deeper layers. InInternational Conference on Learning Representations (ICLR), 2025. URLhttps://arxiv.org/abs/2403.17887 on PithPith's reviewarXiv
  10. [11] Daya Guo, Dejian Yang, Haowei Zhang, et al. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning.Nature, 645(8081):633–638, 2025. doi: 10.1038/s41586-025 -09422-z. URLhttps://www.nature.com/articles/s41586-025-09422-z on Pithdoi:10.1038/s41586-025
  11. [12] XuehaiHe, YichenZhang, LuntianMou, EricXing, andPengtaoXie. PathVQA:30000+questions for medical visual question answering.arXiv preprint arXiv:2003.10286, 2020. URLhttps: //arxiv.org/abs/2003.10286 on PithPith's reviewarXiv
  12. [13] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InInternational Conference on Learning Representations (ICLR), 2022. URLhttps://arxiv.org/abs/2106 .09685. 12 MedJev on Pith
  13. [14] Yutao Hu, Tianbin Li, Quanfeng Lu, Wenqi Shao, Junjun He, Yu Qiao, and Ping Luo. OmniMed- VQA: A new large-scale comprehensive evaluation benchmark for medical LVLM.arXiv preprint arXiv:2402.09181, 2024. URLhttps://arxiv.org/abs/2402.09181 on PithPith's reviewarXiv
  14. [15] Daniel Kahneman.Thinking, Fast and Slow. Farrar, Straus and Giroux, New York, 2011 on Pith
  15. [16] Kwai Keye Team. Kwai Keye-VL technical report.arXiv preprint arXiv:2507.01949, 2025. URL https://arxiv.org/abs/2507.01949 on PithPith's reviewarXiv
  16. [17] Yuxiang Lai, Jike Zhong, Ming Li, Shitian Zhao, Yuheng Li, Konstantinos Psounis, and Xiaofeng Yang. Med-R1: Reinforcement learning for generalizable medical reasoning in vision-language models.arXiv preprint arXiv:2503.13939, 2025. URLhttps://arxiv.org/abs/2503.139 39 on PitharXiv
  17. [18] Jason J. Lau, Soumya Gayen, Asma Ben Abacha, and Dina Demner-Fushman. A dataset of clinically generated visual questions and answers about radiology images.Scientific Data, 5: 180251, 2018. doi: 10.1038/sdata.2018.251 on Pithdoi:10.1038/sdata.2018.251
  18. [19] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. LLaVA-OneVision: Easy visual task transfer.arXiv preprint arXiv:2408.03326, 2024. URLhttps://arxiv.org/abs/2408.03326 on PithPith's reviewarXiv
  19. [20] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. LLaVA-Med: Training a large language-and- vision assistant for biomedicine in one day.arXiv preprint arXiv:2306.00890, 2023. URL https://arxiv.org/a... on PithPith's reviewarXiv
  20. [21] Liquid AI. Open d1: Edge decision models for text, vision, and audio. Blog post, 2026. URL https://www.liquid.ai/blog/d1-open on Pith
show 42 more
  1. [22] Bo Liu, Li-Ming Zhan, Li Xu, Lin Ma, Yan Yang, and Xiao-Ming Wu. SLAKE: A semantically- labeledknowledge-enhanceddatasetformedicalvisualquestionanswering. InIEEEInternational Symposium on Biomedical Imaging (ISBI), 2021. URLhttps://arxiv.org/abs/2102.095 42 on Pith
  2. [23] Bo Liu, Lianyu Wang, Meng Wang, Ching-Yu Cheng, and Huazhu Fu. RetiJev: Fast and source-grounded ophthalmic decision models.arXiv preprint arXiv:2610.30947, 2026. URL https://arxiv.org/abs/2610.30947 on PithPith's reviewarXiv
  3. [24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2023. URLhttps://arxiv.or g/abs/2304.08485 on PithPith's reviewarXiv
  4. [25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), on Pith
  5. [26] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge. Blog post, 2024. URL https://llava-vl.github.io/blog/2024-01-30-llava-next/ on Pith
  6. [27] Xin Men, Mingyu Xu, Qingyu Zhang, Bingning Wang, Hongyu Lin, Yaojie Lu, Xianpei Han, and Weipeng Chen. ShortGPT: Layers in large language models are more redundant than you expect. arXiv preprint arXiv:2403.03853, 2024. URLhttps://arxiv.org/abs/2403.03853. 13 MedJev on PithPith's reviewarXiv
  7. [28] Michael Moor, Qian Huang, Shirley Wu, Michihiro Yasunaga, Cyril Zakka, Yash Dalmia, Ed- uardo Pontes Reis, Pranav Rajpurkar, and Jure Leskovec. Med-Flamingo: A multimodal medical few-shot learner.arXiv preprint arXiv:2307.15189, 2023. URLhttps://arxiv.org/abs/23 07.15189 on PithPith's reviewarXiv
  8. [31] Qwen Team. Qwen3.5-4B: Official model card. Hugging Face, 2026. URLhttps://huggin gface.co/Qwen/Qwen3.5-4B on Pith
  9. [32] URLhttps://arxiv.org/abs/2502.19634 on PithPith's reviewarXiv
  10. [33] Joshua Robinson, Christopher Michael Rytting, and David Wingate. Leveraging large lan- guage models for multiple choice question answering. InInternational Conference on Learning Representations (ICLR), 2023. URLhttps://arxiv.org/abs/2210.12353 on PithPith's reviewarXiv
  11. [34] Andrew Sellergren, Sahar Kazemzadeh, Tiam Jaroensri, Atilla Kiraly, Madeleine Traverse, Timo Kohlberger, Shawn Xu, Fayaz Jamil, Cían Hughes, Charles Lau, et al. MedGemma technical report.arXiv preprint arXiv:2507.05201, 2025. URLhttps://arxiv.org/abs/2507.05201 on PithPith's reviewarXiv
  12. [35] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. LLaVA-PruMerge: Adaptive token reduction for efficient large multimodal models. InIEEE/CVF International Conference on Computer Vision (ICCV), 2025. URLhttps://arxiv.org/abs/2403.15388 on PitharXiv
  13. [36] Qwen Team. Qwen3.5-9B: Official model card. Hugging Face, 2026. URLhttps://huggin gface.co/Qwen/Qwen3.5-9B on Pith
  14. [37] Luyao Tang and Cheng Chen. OmniMed-Jev: Calibrating LVLM confidence for trustworthy medical multimodal decisions via system one.arXiv preprint arXiv:2610.00381, 2026. URL https://arxiv.org/abs/2610.00381 on PithPith's reviewarXiv
  15. [38] CristianTommasinoandAntonioM.Rinaldi. PathJev: Decision-nativevision-languagemodelling for histopathology.arXiv preprint arXiv:2610.32478, 2026. URLhttps://arxiv.org/abs/ 2610.32478 on PithPith's reviewarXiv
  16. [39] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions.Scientific Data, 5: 180161, 2018. doi: 10.1038/sdata.2018.161. URLhttps://www.nature.com/articles/ sdata2018161 on Pithdoi:10.1038/sdata.2018.161
  17. [40] Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding.arXiv preprint arXiv:2104.09864, 2021. URLhttps://arxiv.org/abs/2104.09864 on PithPith's reviewarXiv
  18. [41] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE.Journal of Machine Learning Research, 9(86):2579–2605, 2008. URLhttps://jmlr.org/papers/v9/vande rmaaten08a.html on Pith
  19. [42] vLLM Semantic Router Team. d3-flash: A multimodal foundation decision model. Model release, on Pith
  20. [43] Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, et al. InternVL3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency.arXiv preprint arXiv:2508.18265, 2025. URL https:... on PithPith's reviewarXiv
  21. [44] Tao Tu, Shekoofeh Azizi, Danny Driess, Mike Schaekermann, Mohamed Amin, Pi-Chuan Chang, Andrew Carroll, Chuck Lau, Ryutaro Tanno, Ira Ktena, et al. Towards generalist biomedical AI. arXiv preprint arXiv:2307.14334, 2023. URLhttps://arxiv.org/abs/2307.14334. 14 MedJev on PithPith's reviewarXiv
  22. [45] Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Towards generalist foundation model for radiology by leveraging web-scale 2D&3D medical data.arXiv preprint arXiv:2308.02463, 2023. URLhttps://arxiv.org/abs/2308.02463 on PithPith's reviewarXiv
  23. [46] Xiaomi LLM-Core Team. MiMo-VL technical report.arXiv preprint arXiv:2506.03569, 2025. URLhttps://arxiv.org/abs/2506.03569 on PithPith's reviewarXiv
  24. [47] URLhttps://huggingface.co/vllm-sr/d3-flash on Pith
  25. [49] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. InAdvances in Neural Information Processing Systems (NeurIPS), 2022. URLhttps: //arxiv.org/abs... on PithPith's reviewarXiv
  26. [50] Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pfister, and Bingbing Ni. MedMNIST v2—a large-scale lightweight benchmark for 2D and 3D biomedical image classification.Scientific Data, 2023. doi: 10.1038/s41597-022-01721-8. URLhttps: //arxiv.or... on PithPith's reviewarXiv
  27. [51] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. VisionZip: Longer is better but not necessary in vision language models.arXiv preprint arXiv:2412.04467, 2024. URLhttps://arxiv.org/abs/2412.04467 on PitharXiv
  28. [52] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Conical visual concentration for efficient large vision- language models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), on Pith
  29. [53] URLhttps://arxiv.org/abs/2410.17247 on PithPith's reviewarXiv
  30. [54] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. BiomedCLIP: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs.arXiv preprint arXiv:2303.0091... on PithPith's reviewarXiv
  31. [55] URLhttps://arxiv.org/abs/2506.07044 on PithPith's reviewarXiv
  32. [56] Biao Yang, Bin Wen, Boyang Ding, Changyi Liu, et al. Kwai Keye-VL 1.5 technical report.arXiv preprint arXiv:2509.01563, 2025. URLhttps://arxiv.org/abs/2509.01563 on PithPith's reviewarXiv
  33. [57] Xunlan Zhou, Xianliang Yang, and Li Zhao. From text decisions to pixels: An study of Jev-style visual choice model.arXiv preprint arXiv:2609.29283, 2026. URLhttps://arxiv.org/ab s/2609.29283 on PithPith's reviewarXiv
  34. [59] Zonghai Yao, Benlu Wang, Yifan Zhang, Junda Wang, Iris Xia, Zhipeng Tang, Shuo Han, Feiyun Ouyang, Zhichao Yang, Arman Cohan, and Hong Yu. Medical thinking with multiple images. InInternational Conference on Learning Representations (ICLR), 2026. URLhttps: //arxiv.org/abs/2604.16506 on PithPith's reviewarXiv
  35. [60] Guanxu Yu and Yuhang Yao. Visual Jev: Accurate and efficient decisions from shared visual context.arXiv preprint arXiv:2609.25845, 2026. URLhttps://arxiv.org/abs/2609.258 45. 15 MedJev on PithPith's reviewarXiv
  36. [62] Xiaoman Zhang, Chaoyi Wu, Ziheng Zhao, Weixiong Lin, Ya Zhang, Yanfeng Wang, and Weidi Xie. PMC-VQA: Visual instruction tuning for medical visual question answering.arXiv preprint arXiv:2305.10415, 2023. URLhttps://arxiv.org/abs/2305.10415 on PithPith's reviewarXiv
  37. [63] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, and Shanghang Zhang. Sparse- VLM: Visual token sparsification for efficient vision-language model inference. InInternational Conference on M... on PithPith's reviewarXiv
  38. [66] URLhttps://arxiv.org/abs/2504.10479 on PithPith's reviewarXiv
  39. [67] Yuhan Zhu, Xiangyu Zeng, Qingyi Si, Chenxu Yang, Yifei Li, Luozheng Qin, Junhao Zhou, Dingyu Yao, Jun Zhang, Chuanyu Qin, Zhiqiu Zhang, Yuandong Yang, Yuqi Xu, Tianyu Zhang, Hongcheng Guo, Jiaqi Wang, Nan Duan, and Limin Wang. VideoJev: System One model for video understanding... on PithPith's reviewarXiv
  40. [2024] URLhttps://arxiv.org/abs/2310.03744 on PithPith's reviewarXiv
  41. [2025] URLhttps://arxiv.org/abs/2502.13923 on PithPith's reviewarXiv
  42. [2026] URLhttps://arxiv.org/abs/2610.22276 on PithPith's reviewarXiv
Cited by 1 Pith paper

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment