Pith. sign in

REVIEW 5 major objections 6 minor 36 references

SurgVisAgent: Multimodal Agentic Model for Versatile Surgical Visual Enhancement

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A multimodal agent that names the distortion before fixing it outperforms single-task surgical enhancement models.

desk verdict Plausible agent design that would benefit from a serious referee, but the headline numbers rest on an unspecified train/test split and a self-constructed synthetic benchmark — treat the SOTA claim as conditional, not proven. read the letter →

arxiv 2507.02252 v1 pith:OVB477QB submitted 2025-07-03 cs.CV cs.AI

Zeyu Lei , Hongyuan Yu , Jinlin Wu , Zhen Chen This is my paper
classification cs.CVcs.AI
keywords MultimodalLLMAgenticSurgicalSceneUnderstandingImageEnhancementEndoscopicdistortionChain-of-thoughtreasoningIn-contextfew-shotlearningSmokeremoval
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

This paper argues that a single multimodal agent can replace a collection of single-purpose surgical image enhancement models by first naming the distortion. It introduces SurgVisAgent, which combines a ResNet-based prior model with GPT-4o's few-shot and chain-of-thought reasoning to classify low light, overexposure, motion blur, and smoke, then invokes a matching enhancement model. On a synthesized benchmark built from endoscopic and laparoscopic frames, the agent reports 97 percent category accuracy, 84 percent severity accuracy, 83 percent combined accuracy, and the best full-reference and no-reference quality scores compared with each single-task model run alone. The reason this matters is that real surgical video often contains several distortions at once, and no single-purpose model handles that well.

What carries the argument

The load-bearing object is the surgical prior model plus the MLLM agent's chain-of-thought model-selection loop. The prior is a ResNet trained on a curated benchmark to output softened probability distributions over four distortion categories and three severity levels, smoothed by a temperature of $T=1.1$. The agent conditions its reasoning on these priors and on $k=15$ labeled in-context examples, and the chain-of-thought sequence turns the label prediction into a transparent sequence of steps that ends in the rule $L' = f_{\theta}^{\mathrm{CoT}}(I'; C_k, \mathrm{prior}_{c/s}(I'), R)$, which then selects enhancement models through $M_{\mathrm{select}} = \{M_i \mid [c_i, s_i] \in L'\}$. The argument runs on this mechanism because it is what lets a general-purpose MLLM behave like a surgical specialist without fine-tuning.

What would settle it

Take a set of real endoscopic video frames with mixed low light, overexposure, motion blur, and smoke, have surgeons label distortion category and severity, then run SurgVisAgent and the four single-task models on the same frames; if the agent's category and severity accuracy on real frames drops to the level of the ResNet prior alone, or its enhanced images are not preferred by surgeons, the routing loop is not generalizing.

Watch

Extended reading notes

Core claim

The central claim is that distortion understanding and enhancement should be coupled in one decision-making loop rather than treated as separate fixed pipelines. SurgVisAgent first produces soft-label probability distributions over distortion category and severity with a ResNet prior, then feeds those priors, together with fifteen in-context examples, into GPT-4o. The agent reasons audibly through a chain of thought—illumination analysis, anomaly analysis, final judgment—and only then selects enhancement models from a fixed set. The reported result is that this routed pipeline beats each single-task model on its own and on composite distortions, with the largest gains on no-reference and combined settings.

Load-bearing premise

The benchmark's synthetic distortions and the simplifying assumption that all smoke is severe have to stand in for real endoscopic corruption; if clinical distortions are harder or systematically different, the reported accuracy and quality gains will not transfer.

Editorial extensions

If this is right

  • One system can cover four distortion types and composites, reducing the need to deploy and maintain separate enhancement models.
  • If the 97 percent category and 84 percent severity accuracies hold on real clinical data, the agent could act as a triage layer that tells a surgeon or downstream algorithm what corruption is present and how severe it is.
  • Chain-of-thought selection means every enhancement decision carries an inspectable reason, which is useful for safety review in a clinical setting.
  • Because no fine-tuning is required, adding new enhancement models or new distortion categories may only require new examples and a new prior head.

Reading between the lines

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

  • A natural next test is whether the same routing loop works on real unpaired clinical distortions, since the synthesized low-light, overexposure, and blur images may be easier than what surgical cameras actually produce.
  • The severity prior for smoke is fixed to severe because DeSmoke-LAP has no paired mild-smoke data, so the reported 84 percent severity number could shift once mild smoke appears in a benchmark.
  • Enhancement quality is bounded by the underlying DiffLL, FECNet, MIMO-UNet+, and DeSmoke-LAP models, so SurgVisAgent's contribution is routing rather than restoration.
  • The ablation suggests sample richness matters more than sample count, which points to a testable extension: curating hard in-context examples could raise combined accuracy without growing $k$.
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.

Referee Report

5 major / 6 minor

Summary. The paper proposes SurgVisAgent, a multimodal agentic system that combines a ResNet-based surgical prior model (producing soft-label distributions over distortion category and severity) with GPT-4o in-context few-shot learning and chain-of-thought reasoning to select and apply enhancement models (DiffLL, FECNet, MIMO-UNet+, DeSmoke-LAP) for low-light, overexposure, motion blur, and smoke distortions. The authors construct a synthetic benchmark from EndoVis18 and DeSmoke-LAP and report 84% severity accuracy, 97% category accuracy, 83% combined accuracy, and improved full-reference and no-reference metrics over single-task models. The paper also includes an ablation study on few-shot sample counts.

Significance. If the reported results are reproducible and transferable, SurgVisAgent would be a useful step toward unified, interpretable enhancement of surgical endoscopic images, addressing composite distortions that single-task methods handle poorly. The architectural idea of injecting a lightweight domain prior into an MLLM agent via probabilistic soft labels, rather than expensive fine-tuning, is interesting and well-motivated. The paper provides a new synthetic benchmark and a four-task comparison, which is useful groundwork. However, the current evaluation has substantial gaps: the train/test protocol for the prior model is not stated, the benchmark is self-constructed with synthetic degradations and an unvalidated severity assumption, and the agent's routing contribution is not isolated by a baseline. These issues currently prevent the central claims from being fully supported.

major comments (5)
  1. [§3.1, §3.3, Table 1] The paper never states the train/test split used to train and evaluate the ResNet prior model or the overall benchmark. Section 3.1 mentions EndoVis18's original 2,235/999 split, but it does not say whether the synthesized distortion benchmark respects that split, whether prior-model training images are disjoint from images used to compute Tables 1 and 2, or whether any images are shared between the prior training set and the agent's few-shot context. Without this information, the reported 84%/97%/83% accuracies and the enhancement gains in Table 2 could reflect prior-model memorization rather than generalization. Please specify the exact partition, verify disjointness for all components (prior training, few-shot examples, evaluation), and report per-split results.
  2. [§3.1, Table 1] The smoke severity assumption is load-bearing and internally inconsistent with the reported numbers. The text says 'the severity of the smoke distortion is assumed to be severe' because DeSmoke-LAP has unpaired data. If all smoke images are labeled severe, then for the smoke column in Table 1 a trivial majority-class predictor would achieve 100% severity accuracy, yet the prior model reports only 0.31 and SurgVisAgent 0.52. This suggests either the severity labels are noisy, the training distribution differs from the evaluation distribution, or the labels are not actually all severe. The assumption needs validation (e.g., human annotation study or a separate severity-labeled smoke set) or the severity claims for smoke should be removed.
  3. [§3.4, Table 2] There is no baseline that isolates the contribution of the routing/selection mechanism. Table 2 compares SurgVisAgent to single-task models, but the agent's advantage could come from the underlying enhancement models or from a trivial selection rule rather than from the prior plus CoT reasoning. To support the central claim, the paper should include an oracle-routing baseline (using ground-truth distortion labels to select enhancement models), a random-selection baseline, and ablations that remove the prior and remove CoT. Without these, Table 2 does not demonstrate that the agent's perceptual/diagnostic component drives the enhancement improvement.
  4. [§3.1, §3.3] The benchmark is entirely synthetic for low-light, overexposure, and motion blur, generated with methods from [21] and [14], and no validation of realism is provided. There is no external dataset, clinical validation, or human perceptual study, and the text does not report how many images were generated per category and severity or the exact synthesis parameters. The claim that the method generalizes to real-world surgical conditions is therefore not supported. Please provide a detailed benchmark description, distribution statistics, and at least a qualitative or quantitative validation that the synthetic distortions resemble clinical endoscopy conditions.
  5. [§3.5, Table 3] The ablation study is too narrow to support the design choices. It only varies the few-shot sample count and composition; there are no ablations removing the prior soft labels, removing CoT, or varying the temperature T in Eq. (1). The formatting of Table 3 is also ambiguous (the subcolumns under 'Samples' are not clearly separated), and no error bars or significance tests are reported. The claim that 'the richness of the samples is more important than the number of samples' should be supported by a clearer experimental design and statistical analysis.
minor comments (6)
  1. [Abstract, §2.2 title] There are typographical issues: 'andchain-of-thought' in the abstract and 'F ew-Shot' in the Section 2.2 title; these should be corrected.
  2. [Eq. (1), §3.3] The temperature T is fixed at 1.1, but no sensitivity analysis or rationale for this value is provided; a brief explanation or ablation would strengthen the paper.
  3. [Table 1] The row labels 'Severity ✓ Distortion ×' and the column grouping are not self-explanatory; please expand the caption and define what each numeric cell represents (e.g., per-class accuracy vs. overall accuracy).
  4. [§3.3] The sentence 'all enhancement models are trained for an equal number of iterations' is unclear because SurgVisAgent itself is not trained; please clarify which models are trained, with what data, and what hyperparameters are used.
  5. [§2.3] The phrase 'recursive self-feedback approach' is not explained or used elsewhere; either define the recursion or remove the phrase to avoid confusion.
  6. [General] No code, data, or model weights are provided, which limits reproducibility given that the benchmark is self-constructed and GPT-4o is accessed via API; please add an availability statement or a detailed protocol.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the pipeline's prior and benchmark are empirically coupled, but no reported quantity is defined in terms of another reported quantity or reduced by construction.

full rationale

The claimed derivation chain is an empirical pipeline rather than a formal derivation: a ResNet prior is trained on a curated benchmark, its soft labels are fed to GPT-4o with in-context examples, and the agent outputs category and severity labels before selecting enhancement models. No equation in the paper defines the predicted labels in terms of the evaluation metric, nor defines the prior model in terms of the agent's output. The ground-truth labels are generated from the benchmark construction, including synthetic low-light, overexposure, blur, and DeSmoke-LAP smoke with an assumed severe label, and the prior is trained on those same labels. This makes the evaluation self-referential and raises legitimate concerns about external validity and about the absence of an explicit train/test split for the prior, but it is not circularity: the prior's outputs are inputs to the agent, not reductions of the reported accuracy. The enhancement comparisons in Table 2 use the same single-task models as the agent's action set, yet the agent's orchestration remains an independent, measurable operation whose advantage is not forced by construction. The self-citations in the references are contextual and not load-bearing. Therefore no circular step can be exhibited under the required standard, and the appropriate score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central evaluation rests on a self-constructed benchmark whose severity labels are partly author-assumed, plus pretrained components whose domain suitability is asserted rather than demonstrated. No new physical entities are introduced.

free parameters (3)
  • Temperature T of prior soft-label distillation = 1.1
    Smooths the ResNet prior distributions in Eq. (1); set in Section 3.3 with no sweep or justification beyond stabilization.
  • Few-shot context size K = 15
    Number of in-context examples in Eq. (2); ablation in Table 3 shows only weak sensitivity between 10 and 20, so the choice is manual.
  • Prior model ResNet weights = trained on the constructed benchmark
    The ResNet prior models are trained on the authors' benchmark annotations, so the soft labels carry benchmark-specific supervision into the agent.
assumptions (5)
  • domain assumption EndoVis18 and DeSmoke-LAP images are suitable raw material for training and evaluating surgical enhancement.
    Section 3.1 uses EndoVis18, a segmentation dataset, as base images and DeSmoke-LAP smoke frames, without demonstrating that their content matches the distortions being studied.
  • domain assumption Synthetic degradations from [21] and [14] faithfully mimic real low-light, overexposure, and motion blur in endoscopy.
    Section 3.1 states distortions are synthesized using external methods; no clinical validation of realism is provided.
  • ad hoc to paper All DeSmoke-LAP smoke images are severe.
    Section 3.1: 'the severity of the smoke distortion is assumed to be severe' due to unpaired data; this arbitrary label shapes the severity benchmark.
  • domain assumption GPT-4o can reason correctly about surgical distortions given prior soft labels and 15 examples.
    Section 2.3 assumes the pretrained MLLM's chain-of-thought and visual perception are reliable for this domain; no systematic error analysis is given.
  • domain assumption Off-the-shelf enhancement models improve surgical images of their target distortion.
    Section 3.2 uses DiffLL, FECNet, MIMO-UNet+, and DeSmoke-LAP as the function set, assuming their outputs are beneficial on endoscopic images without task-specific retraining or clinical validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SurgVisAgent: Multimodal Agentic Model for Versatile Surgical Visual Enhancement." pith.science (2026). https://pith.science/paper/OVB477QB

@misc{pith2026250702252,
  author       = {Pith},
  title        = {Pith review of: SurgVisAgent: Multimodal Agentic Model for Versatile Surgical Visual Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OVB477QB}},
  note         = {Machine review of arXiv:2507.02252}
}
read the original abstract

Precise surgical interventions are vital to patient safety, and advanced enhancement algorithms have been developed to assist surgeons in decision-making. Despite significant progress, these algorithms are typically designed for single tasks in specific scenarios, limiting their effectiveness in complex real-world situations. To address this limitation, we propose SurgVisAgent, an end-to-end intelligent surgical vision agent built on multimodal large language models (MLLMs). SurgVisAgent dynamically identifies distortion categories and severity levels in endoscopic images, enabling it to perform a variety of enhancement tasks such as low-light enhancement, overexposure correction, motion blur elimination, and smoke removal. Specifically, to achieve superior surgical scenario understanding, we design a prior model that provides domain-specific knowledge. Additionally, through in-context few-shot learning and chain-of-thought (CoT) reasoning, SurgVisAgent delivers customized image enhancements tailored to a wide range of distortion types and severity levels, thereby addressing the diverse requirements of surgeons. Furthermore, we construct a comprehensive benchmark simulating real-world surgical distortions, on which extensive experiments demonstrate that SurgVisAgent surpasses traditional single-task models, highlighting its potential as a unified solution for surgical assistance.

Figures

Figures reproduced from arXiv: 2507.02252 by the authors.

Figure 1
Figure 1. The overview of the SurgVisAgent framework. The surgical visual in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The illustration of our benchmark. (a) Image comparisons illustrating [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison of SurgVisAgent with other single-task founda [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 25 canonical work pages

  1. [21]

    IEEE Transactions on Pattern Analysis and Machine Intelligence44(12), 9396–9416 (2021)

    Li, C., Guo, C., Han, L., Jiang, J., Cheng, M.M., Gu, J., Loy, C.C.: Low-light image and video enhancement using deep learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence44(12), 9396–9416 (2021)

    work page 2021

  2. [14]

    NeurIPS 33, 975–985 (2020)

    Guo, Q., Juefei-Xu, F., Xie, X., Ma, L., Wang, J., Yu, B., Feng, W., Liu, Y.: Watch out! motion is blurring the vision of your deep neural networks. NeurIPS 33, 975–985 (2020)

    work page 2020

  3. [1]

    arXiv preprint arXiv:2001.11190 (2020)

    Allan,M.,Kondo,S.,Bodenstedt,S.,Leger,S.,Kadkhodamohammadi,R.,Luengo, I., Fuentes, F., Flouty, E., Mohammed, A., Pedersen, M., et al.: 2018 robotic scene segmentation challenge. arXiv preprint arXiv:2001.11190 (2020)

    arXiv 2020

  4. [2]

    arXiv preprint arXiv:2308.12966 (2023)

    Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023)

    arXiv 2023

  5. [3]

    In: MICCAI

    Bai, L., Chen, T., Tan, Q., Nah, W.J., Li, Y., He, Z., Yuan, S., Chen, Z., Wu, J., Islam,M.,etal.:Endouic:Promptablediffusiontransformerforunifiedillumination correction in capsule endoscopy. In: MICCAI. pp. 296–306. Springer (2024) SurgVisAgent: Multimodal Agentic Model for Surgical Visual Enhancement 11

    work page 2024

  6. [4]

    IEEE Transactions on Medical Imaging (2023)

    Ban,Y.,Eckhoff,J.A.,Ward,T.M.,Hashimoto,D.A.,Meireles,O.R.,Rus,D.,Ros- man, G.: Concept graph neural networks for surgical video understanding. IEEE Transactions on Medical Imaging (2023)

    work page 2023

  7. [5]

    NeurIPS 33, 1877–1901 (2020)

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Nee- lakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS 33, 1877–1901 (2020)

    work page 2020

  8. [6]

    In: MICCAI

    Chen, T., Lyu, Q., Bai, L., Guo, E., Gao, H., Yang, X., Ren, H., Zhou, L.: Lightdiff: Surgical endoscopic image low-light enhancement with t-diffusion. In: MICCAI. pp. 369–379. Springer (2024)

    work page 2024

Show all 36 references
  1. [7]

    In: MICCAI

    Chen, Z., Guo, Q., Yeung, L.K., Chan, D.T., Lei, Z., Liu, H., Wang, J.: Surgical video captioning with mutual-modal concept alignment. In: MICCAI. pp. 24–34. Springer (2023)

    2023

  2. [8]

    IEEE Transactions on Medical Imaging 40(5), 1377–1389 (2021)

    Chen, Z., Guo, X., Woo, P.Y., Yuan, Y.: Super-resolution enhanced medical image diagnosis with sample affinity interaction. IEEE Transactions on Medical Imaging 40(5), 1377–1389 (2021)

    2021

  3. [9]

    In: MICCAI

    Chen, Z., Guo, X., Yang, C., Ibragimov, B., Yuan, Y.: Joint spatial-wavelet dual- stream network for super-resolution. In: MICCAI. pp. 184–193. Springer (2020)

    2020

  4. [10]

    In: BIBM

    Chen, Z., Luo, X., Wu, J., Chan, D.T., Lei, Z., Ourselin, S., Liu, H.: Surgfc: Multi- modal surgical function calling framework on the demand of surgeons. In: BIBM. pp. 3076–3081. IEEE (2024)

    2024

  5. [11]

    In: BIBM

    Chen, Z., Zhai, Y., Zhang, J., Wang, J.: Surgical temporal action-aware network with sequence regularization for phase recognition. In: BIBM. pp. 1836–1841. IEEE (2023)

    2023

  6. [12]

    In: IROS

    Chen, Z., Zhang, Z., Guo, W., Luo, X., Bai, L., Wu, J., Ren, H., Liu, H.: Asi-seg: Audio-driven surgical instrument segmentation with surgeon intention understand- ing. In: IROS. pp. 13773–13779. IEEE (2024)

    2024

  7. [13]

    In: ICCV

    Cho, S.J., Ji, S.W., Hong, J.P., Jung, S.W., Ko, S.J.: Rethinking coarse-to-fine approach in single image deblurring. In: ICCV. pp. 4641–4650 (2021)

    2021

  8. [15]

    Journal of Orthopaedic Surgery and Research14, 1–21 (2019)

    Hammad, A., Wirries, A., Ardeshiri, A., Nikiforov, O., Geiger, F.: Open versus minimally invasive tlif: literature review and meta-analysis. Journal of Orthopaedic Surgery and Research14, 1–21 (2019)

    2019

  9. [16]

    arXiv preprint arXiv:1503.02531 (2015)

    Hinton, G.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)

    2015 arXiv

  10. [17]

    In: ECCV

    Huang, J., Liu, Y., Zhao, F., Yan, K., Zhang, J., Huang, Y., Zhou, M., Xiong, Z.: Deep fourier-based exposure correction network with spatial-frequency interaction. In: ECCV. pp. 163–180. Springer (2022)

    2022

  11. [18]

    Electronics Letters44(13), 800–801 (2008)

    Huynh-Thu, Q., Ghanbari, M.: Scope of validity of psnr in image/video quality assessment. Electronics Letters44(13), 800–801 (2008)

    2008

  12. [19]

    ACM Transactions on Graphics42(6), 1–14 (2023)

    Jiang, H., Luo, A., Fan, H., Han, S., Liu, S.: Low-light image enhancement with wavelet-based diffusion models. ACM Transactions on Graphics42(6), 1–14 (2023)

    2023

  13. [20]

    In: MICCAI

    Li, C., Liu, H., Liu, Y., Feng, B.Y., Li, W., Liu, X., Chen, Z., Shao, J., Yuan, Y.: Endora: Video generation models as endoscopy simulators. In: MICCAI. pp. 230–240. Springer (2024)

    2024

  14. [22]

    NeurIPS36 (2024) 12 Z

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. NeurIPS36 (2024) 12 Z. Lei et al

    2024

  15. [23]

    In: ISBI

    Luo, X., Pang, Y., Chen, Z., Wu, J., Zhang, Z., Lei, Z., Liu, H.: Surgplan: Surgical phase localization network for phase recognition. In: ISBI. pp. 1–5. IEEE (2024)

    2024

  16. [24]

    IEEE Transactions on Image Processing21(12), 4695–4708 (2012)

    Mittal, A., Moorthy, A.K., Bovik, A.C.: No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing21(12), 4695–4708 (2012)

    2012

  17. [25]

    completely blind

    Mittal, A., Soundararajan, R., Bovik, A.C.: Making a “completely blind” image quality analyzer. IEEE Signal Processing Letters20(3), 209–212 (2012)

    2012

  18. [26]

    International Journal of Computer Assisted Radiology and Surgery17(5), 885–893 (2022)

    Pan, Y., Bano, S., Vasconcelos, F., Park, H., Jeong, T.T., Stoyanov, D.: Desmoke- lap: improved unpaired image-to-image translation for desmoking in laparoscopic surgery. International Journal of Computer Assisted Radiology and Surgery17(5), 885–893 (2022)

    2022

  19. [27]

    van der Sluis, P.C., van der Horst, S., May, A.M., Schippers, C., Brosens, L.A., Joore, H.C., Kroese, C.C., Mohammad, N.H., Mook, S., Vleggaar, F.P., et al.: Robot-assisted minimally invasive thoracolaparoscopic esophagectomy versus open transthoracic esophagectomy for resecta...

    2019

  20. [28]

    arXiv preprint arXiv:2307.09288 (2023)

    Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bash- lykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)

    2023 arXiv

  21. [29]

    arXiv preprint arXiv:2401.10727 (2024)

    Wang, C., Luo, W., Chen, Q., Mai, H., Guo, J., Dong, S., Li, Z., Ma, L., Gao, S., et al.: Mllm-tool: A multimodal large language model for tool agent learning. arXiv preprint arXiv:2401.10727 (2024)

    2024 arXiv

  22. [30]

    In: AAAI

    Wang, J., Chan, K.C., Loy, C.C.: Exploring clip for assessing the look and feel of images. In: AAAI. vol. 37, pp. 2555–2563 (2023)

    2023

  23. [31]

    IEEE Transactions on Image Process- ing 13(4), 600–612 (2004)

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Process- ing 13(4), 600–612 (2004)

    2004

  24. [32]

    NeurIPS 35, 24824–24837 (2022)

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. NeurIPS 35, 24824–24837 (2022)

    2022

  25. [33]

    In: Big Data

    Wu, J., Liang, X., Bai, X., Chen, Z.: Surgbox: Agent-driven operating room sand- box with surgery copilot. In: Big Data. pp. 2041–2048. IEEE (2024)

    2024

  26. [34]

    In: ECCV

    Wu, R., Zhang, Z., Zhang, S., Gou, L., Chen, H., Zhang, L., Chen, H., Zuo, W.: Self-supervised video desmoking for laparoscopic surgery. In: ECCV. pp. 307–324. Springer (2024)

    2024

  27. [35]

    International Journal of Computer Assisted Radiology and Surgery19(2), 345–353 (2024)

    Zhai, Y., Chen, Z., Zheng, Z., Wang, X., Yan, X., Liu, X., Yin, J., Wang, J., Zhang, J.: Artificial intelligence for automatic surgical phase recognition of laparo- scopic gastrectomy in gastric cancer. International Journal of Computer Assisted Radiology and Surgery19(2), 345...

    2024

  28. [36]

    In: CVPR

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586–595 (2018)

    2018

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.