Pith. sign in

REVIEW 4 major objections 6 minor 22 references

Agentic Surgical AI: Surgeon Style Fingerprinting and Privacy Risk Quantification via Discrete Diffusion in a Vision-Language-Action Framework

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper establishes that surgeon-specific embeddings improve gesture prediction in robotic surgery while simultaneously increasing identity-leakage risk under membership inference.

desk verdict The gesture-diffusion application is plausible and clearly described, but the headline privacy result is an artifact of a broken membership-inference attack and should not be published as evidence of identity leakage. read the letter →

arxiv 2506.08185 v2 pith:Y7NNZLK2 submitted 2025-06-09 cs.CV cs.AI

Huixin Zhan , Jason H. Moore This is my paper
classification cs.CVcs.AI
keywords discretediffusionsurgicalgesturepredictionvision-language-actionsurgeonstylefingerprintingmembershipinferenceprivacy-utilitytrade-offpersonalizedembeddingsroboticsurgery
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

Surgeons leave a style signal in their gestures, and the paper's thesis is that modeling this signal improves prediction but creates a measurable privacy cost. The authors build a discrete diffusion model that denoises gesture sequences, conditioning on video, task language, and surgeon embeddings created from natural-language prompts such as "Surgeon ID: 3, average skill score: 3.75". On the JIGSAWS dataset, the most expressive embedding (LLM with ID and GRS) reaches 83.89% Top-1 accuracy and 0.8447 F1, above the non-private baseline's 82.40% and 0.8237. The same embedding is also the most re-identifiable, with membership inference AUC hitting 1.000, so the paper's central result is a performance-privacy trade-off in personalized surgical AI.

What carries the argument

The load-bearing mechanism is a discrete denoising diffusion process over a vocabulary of 15 gesture tokens plus a [MASK] token, with a multinomial corruption schedule $Q_t$ and a transformer reverse model $p_\theta(x_{t-1}|x_t,c,t)$. The conditioning context $c$ fuses ResNet visual features, BERT-encoded task language, a timestep embedding, and the surgeon embedding. The surgeon embedding itself is the privacy-critical object: it is produced by feeding a natural-language prompt (surgeon ID, optionally with averaged GRS) into a frozen third-party sentence model such as Sentence-BERT or MiniLM, then projecting it. That embedding carries both the personalization signal and the identity-leakage signal, which is why the same representation drives both the accuracy gain and the membership inference success.

What would settle it

If a membership inference classifier trained on real held-out surgeons' embeddings (or on embeddings from unseen trials of the same surgeons) drops to near-chance AUC for the ID+GRS condition, then the claimed identity leakage is an artifact of the synthetic non-member proxy rather than evidence of personalization leaking training identity.

Watch

Extended reading notes

Core claim

Stated on the paper's own terms, the discovery is that personalizing a gesture-prediction model through frozen language-model embeddings of surgeon identity and skill improves reconstruction accuracy while making the embedding space more separable from non-members. Table 1 shows the LLM (ID + GRS) condition winning on Top-1 accuracy and weighted F1 over both the ID-only embedding and the non-private learnable baseline. Table 2 shows that the same winning condition yields a membership inference AUC of 1.000, meaning an XGBoost classifier can perfectly separate these embeddings from synthetic non-members. The paper reads the non-private baseline's low attack performance (AUC 0.469) not as genuine privacy protection but as a sign that unstructured embeddings carry little identity signal. The conclusion is that richer personalization and privacy protection pull in opposite directions.

Load-bearing premise

The attack's 'non-member' set is synthetic out-of-distribution embeddings generated from unseen prompt values, so the perfect AUC may measure distribution shift rather than true identity leakage from training data.

Editorial extensions

If this is right

  • If the trade-off is real, surgical AI systems that add per-surgeon embeddings should treat those embeddings as privacy-sensitive and run membership inference as a standard evaluation before deployment.
  • Adding clinically grounded cues like GRS to the prompt improves gesture prediction, so skill scores become doubly useful: they improve behavior modeling and also sharpen the identity signal that attackers can exploit.
  • The non-private baseline's low attack scores suggest that weak personalization hides identity well, meaning a privacy-preserving system might deliberately degrade embedding expressiveness unless it adds protections.
  • The framework's design—frozen third-party LLM encoders with learnable projections—means the privacy risk can be tuned by choosing the encoder and the prompt template, not just by adding noise.

Reading between the lines

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

  • The membership inference 'non-member' set is synthetic out-of-distribution embeddings from unseen prompt values; a stronger test would use real held-out surgeons or same-surgeon unseen trials, and that could lower the reported AUC substantially.
  • The perfect AUC of 1.000 may reflect distribution shift between the LLM prompt embeddings and the synthetic non-members more than memorization of individual training sequences, so the privacy claim should be validated with a calibrated attack.
  • The same personalization–privacy tension probably appears in any user-conditioned generative model, such as style-adaptive text or motor-behavior models, where richer user embeddings improve fidelity and raise re-identification risk.
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

4 major / 6 minor

Summary. The paper proposes a discrete diffusion-based vision-language-action framework for personalized surgical gesture prediction on the JIGSAWS dataset. Surgeon-specific conditioning is provided by embeddings derived from frozen language models applied to prompts containing surgeon ID and Global Rating Score, and the authors report that this improves gesture prediction accuracy (Table 1) while simultaneously increasing susceptibility to a membership inference attack (Table 2). The central claim is therefore a privacy-utility trade-off: richer personalized embeddings improve task performance but make identity leakage easier. The reviewer finds that the gesture-prediction component is a plausible if modest empirical contribution, but the privacy claim rests on a membership inference protocol whose 'non-member' set is synthetic and out-of-distribution, which invalidates the reported AUC values as evidence of identity leakage.

Significance. If the privacy-utility trade-off were established, the paper would be a useful contribution to the emerging discussion of personalization and privacy in surgical AI. The code release, the use of a standard public dataset, and the framing of gesture prediction as discrete sequence denoising are positive aspects. However, the load-bearing privacy result is not supported by the current experimental design: the membership inference attack compares real embeddings against synthetic out-of-distribution embeddings, so the reported AUC of 1.000 for the LLM (ID + GRS) condition is better explained as prompt-real-versus-synthetic separation than as training-data membership leakage. In addition, the gesture accuracy gains in Table 1 are small and unreplicated, and the non-private baseline row of Table 2 is internally inconsistent with a working attack protocol. The significance of the paper as a whole therefore depends on a claim that the experiments do not currently establish.

major comments (4)
  1. [§2.6 and Table 2] The membership inference attack defines 'non-member' embeddings as 'synthetic out-of-distribution' ones, but the paper never specifies how these synthetic embeddings are generated. If they are produced from unseen surgeon IDs or random GRS values, the classifier's task is not 'was this embedding produced by a model trained on this surgeon's gestures?' but 'is this a real surgeon prompt or a fabricated one?' Because the third-party LLM embedding is a frozen deterministic function of the prompt string, a classifier can separate real prompts from synthetic prompts without ever using gesture, video, or training data. The AUC = 1.000 for the LLM (ID + GRS) condition is therefore a distribution-shift artifact rather than evidence of identity leakage, and the central privacy claim is unsupported.
  2. [Table 2, non-private baseline row] The non-private baseline uses a learnable embedding layer that directly maps surgeon ID to a vector, which should make membership trivially identifiable if the attack measured membership in the training data. Yet the reported accuracy is 0.889, precision is 0.000, recall is 0.000, and AUC is 0.469, meaning the classifier has learned to predict none of the positive (member) class. This internal inconsistency indicates that the attack protocol itself is broken, rather than that the baseline is privacy-preserving; a valid attack should perform at least as well on the most identity-explicit representation as on the LLM-based representations.
  3. [Table 1] The gesture prediction improvement that motivates the personalization claim is small (Top-1 accuracy 83.89% vs 82.40%, weighted F1 0.8447 vs 0.8237) and is reported without error bars, multiple seeds, or significance tests. There is also no comparison against existing surgical gesture recognition or sequence prediction baselines, so the reader cannot determine whether the reported differences are within run-to-run variability or whether the diffusion-based formulation is competitive with prior work. The claim that 'personalized embeddings improve gesture prediction accuracy' is therefore not statistically supported.
  4. [§3.4 and Table 2] The paper interprets the results as showing that 'more expressive embeddings improve task performance but increase susceptibility to identity inference attacks,' yet both LLM conditions achieve AUC = 1.000. The claimed privacy gradient rests only on precision (0.998 vs 0.500) and F1 (0.988 vs 0.667), which are not meaningful if the non-member proxy is invalid. Moreover, the ID+GRS and ID-only conditions differ in both the prompt content and the specific sentence encoder used (§3.1), so the observed differences cannot be attributed cleanly to the addition of GRS information.
minor comments (6)
  1. [§2.3 and §3.1] The prompt 'Surgeon ID: 3' includes a raw identifier, so the statement in §2.3 that the approach avoids 'exposing explicit identity' or 'direct integration of raw identifiers' is misleading; the identity is exposed to the frozen encoder, and the claimed privacy property would need a precise threat model to be meaningful.
  2. [§2.4] The text says gradients are backpropagated through the entire diffusion trajectory, but the loss in Equation (1) is written as an expectation over a single sampled timestep, which is the standard denoising objective; the wording should be corrected to avoid implying full-trajectory backpropagation.
  3. [§3.1 and Figure 3] The manuscript does not specify the train/test split protocol for JIGSAWS (e.g., by surgeon, by trial, or by task), which is essential for interpreting both the gesture prediction numbers and the membership inference results; this should be stated explicitly.
  4. [§2.3] Sentence-BERT and MiniLM are sentence encoders, not large language models in the usual sense; repeatedly calling them 'third-party LLMs' inflates the technical claim and should be corrected.
  5. [Figure 3] The three panels (a), (b), and (c) are presented as separate figures or are split across pages, which makes the qualitative comparison difficult; they should be placed together on one page with shared axes.
  6. [References] Reference [1] (Carlini et al.) is cited as an example of privacy-utility trade-offs in embedding models, but that work concerns extracting training data from language models, which is not the same setting; a more directly relevant citation would clarify the relationship.

Circularity Check

1 steps flagged · score 6.0 of 10

Privacy 'identity leakage' result is forced by construction: the membership-inference non-member set is synthetic, and the real embeddings are deterministic functions of the ID/GRS prompt.

  1. self definitional [Section 2.6 (Privacy Risk Analysis via Membership Inference) and Section 3.4 (Privacy Evaluation), Table 2]
    "Given the learned surgeon embeddings {s_i}, we simulate an adversary that trains a binary classifier (e.g., XGBoost [2]) to distinguish between in-training (member) embeddings and synthetic out-of-distribution (non-member) ones. High classifier performance—measured by AUC, accuracy, and F1-score—indicates a greater risk of identity leakage. ... the resulting embeddings are easily distinguishable from synthetic (non-member) embeddings, with an AUC of 1.000."

    The 'non-member' set is synthetic (generation unspecified), not actual non-member data. Since the frozen third-party LLM deterministically maps the prompt string ('Surgeon ID: i, GRS: 3.75', Sec. 2.3/3.1) to an embedding, the classifier's task is to separate real-prompt embeddings from synthetic-prompt embeddings—prompt distribution shift, not membership in the gesture-training set. AUC=1.000 is forced by construction. Supporting anomaly: the non-private baseline, which directly encodes surgeon ID, gives zero precision/recall and AUC 0.469, which would be impossible if the attack measured membership. Hence the 'identity leakage' claim is built into the input definition.

full rationale

The gesture-prediction half of the paper is a straightforward empirical comparison: three surgeon-embedding strategies are trained under identical diffusion conditions, and Table 1 reports Top-1 accuracy and weighted F1. That part is self-contained and not circular. No self-citations appear in the reference list, and no load-bearing claim rests on the authors' prior work. The circularity is confined to the privacy half. The membership-inference protocol in Section 2.6 defines the 'non-member' class as 'synthetic out-of-distribution' embeddings, while real 'member' embeddings are deterministic functions of a frozen LLM applied to prompts that literally contain 'Surgeon ID: i' and optionally GRS. The classifier therefore separates real prompts from synthetic prompts; AUC=1.000 for the LLM (ID+GRS) condition is a consequence of this input construction, not evidence that gesture training data leaks identity. The non-private baseline's zero precision/recall under the same protocol reinforces that the attack is not measuring membership. Because the paper's headline trade-off depends on this privacy result, the central claim is partially circular: the performance gain is empirical, but the claimed 'increased identity leakage' is forced by design.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The main free choices are the diffusion timesteps, sequence length, hidden dimension, and training schedule. The most important unstated assumption is that synthetic non-members are a valid baseline for membership inference, which is false and drives the privacy result.

free parameters (4)
  • Number of diffusion timesteps T = 10
    Chosen by hand to set the noise schedule; no external justification or sensitivity analysis.
  • Gesture sequence length L = 5 tokens
    Chosen to capture short temporal contexts; the entire 'sequence prediction' task is restricted to 5-gesture windows.
  • Shared hidden dimension = 512
    Manual architecture choice for projection layers, embeddings, and denoiser; not central but still a hand-set number.
  • Training schedule = lr=1e-3, batch=32, epochs=20
    Set without sensitivity analysis; standard choices but manually selected.
assumptions (4)
  • domain assumption JIGSAWS gesture annotations are accurate and capture surgeon-specific style.
    The entire evaluation treats gesture labels as ground truth for style; label noise would affect fingerprinting.
  • domain assumption Frozen third-party LLM embeddings of textual prompts provide a meaningful surrogate for surgeon identity and skill.
    The method assumes MiniLM or Sentence-BERT embeddings separate surgeons usefully; no validation of this assumption is provided.
  • ad hoc to paper Synthetic out-of-distribution prompts are a valid non-member set for membership inference.
    Section 2.6 generates non-members synthetically rather than from real held-out surgeons, which makes the attack a distribution-shift detector, not a membership test.
  • domain assumption The 5-token window captures enough context for personalized prediction.
    Section 3 defines each input sequence as 5 gesture tokens; longer-term style is not evaluated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Agentic Surgical AI: Surgeon Style Fingerprinting and Privacy Risk Quantification via Discrete Diffusion in a Vision-Language-Action Framework." pith.science (2026). https://pith.science/paper/Y7NNZLK2

@misc{pith2026250608185,
  author       = {Pith},
  title        = {Pith review of: Agentic Surgical AI: Surgeon Style Fingerprinting and Privacy Risk Quantification via Discrete Diffusion in a Vision-Language-Action Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y7NNZLK2}},
  note         = {Machine review of arXiv:2506.08185}
}
read the original abstract

Surgeons exhibit distinct operating styles shaped by training, experience, and motor behavior-yet most surgical AI systems overlook this personalization signal. We propose a novel agentic modeling approach for surgeon-specific behavior prediction in robotic surgery, combining a discrete diffusion framework with a vision-language-action (VLA) pipeline. Gesture prediction is framed as a structured sequence denoising task, conditioned on multimodal inputs including surgical video, intent language, and personalized embeddings of surgeon identity and skill. These embeddings are encoded through natural language prompts using third-party language models, allowing the model to retain individual behavioral style without exposing explicit identity. We evaluate our method on the JIGSAWS dataset and demonstrate that it accurately reconstructs gesture sequences while learning meaningful motion fingerprints unique to each surgeon. To quantify the privacy implications of personalization, we perform membership inference attacks and find that more expressive embeddings improve task performance but simultaneously increase susceptibility to identity leakage. These findings demonstrate that while personalized embeddings improve performance, they also increase vulnerability to identity leakage, revealing the importance of balancing personalization with privacy risk in surgical modeling. Code is available at: https://github.com/huixin-zhan-ai/Surgeon_style_fingerprinting.

Figures

Figures reproduced from arXiv: 2506.08185 by the authors.

Figure 1
Figure 1. Personalized gesture sequence prediction using diffusion models. (a) Vision features are extracted from surgical videos using ResNet, while task-level semantic cues are obtained from the language prompt (e.g., ‘suturing task”) using BERT. Surgeon-specific behavior embeddings are generated by encoding the surgeon ID and GRS using a third-party large language model (LLM). These visual, linguistic, and behavioral signa… view at source ↗
Figure 2
Figure 2. t-SNE of surgeon embeddings. (Left) Trained using surgeon ID only. (Right) Trained with both ID and GRS supervision. Colors indicate average skill score (GRS) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Predicted gesture token distributions across surgeons. Rows represent surgeon IDs and columns denote gesture tokens. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 3
Figure 3. Figure 3: (continued) Panel (c) shows structured, personalized distributions under ID + GRS embedding. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 18 canonical work pages

  1. [1]

    Extract- ing training data from large language models

    Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-V oss, Katherine Lee, Adam Roberts, Tom B Brown, Dawn Song, Ulfar Erlingsson, et al. Extract- ing training data from large language models. InUSENIX Security Symposium, 2021. 6

    work page 2021

  2. [2]

    Xgboost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. InProceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785–794, 2016. 3

    work page 2016

  3. [3]

    Bert: Pre-training of deep bidirectional trans- formers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional trans- formers for language understanding. InProceedings of the 2019 conference of the North American chapter of the asso- ciation for computational linguistics: human language tech- nologies, volume 1 (long and short papers), pages 4171– 4186, 2019. 3

    2019

  4. [4]

    Visual question answering in robotic surgery: A comprehensive re- view.IEEE Access, 2025

    Di Ding, Tianliang Yao, Rong Luo, and Xusen Sun. Visual question answering in robotic surgery: A comprehensive re- view.IEEE Access, 2025. 5

    work page 2025

  5. [5]

    Segmental spatiotemporal cnns for fine-grained ac- tion segmentation

    Robert DiPietro, Colin Lea, Anand Malpani, and Gregory D Hager. Segmental spatiotemporal cnns for fine-grained ac- tion segmentation. InEuropean Conference on Computer Vision (ECCV), pages 36–52. Springer, 2019. 6

    work page 2019

  6. [6]

    The algorithmic founda- tions of differential privacy.Foundations and Trends in The- oretical Computer Science, 9(3-4):211–407, 2014

    Cynthia Dwork and Aaron Roth. The algorithmic founda- tions of differential privacy.Foundations and Trends in The- oretical Computer Science, 9(3-4):211–407, 2014. 6

    work page 2014

  7. [7]

    Model inversion attacks that exploit confidence information and basic countermeasures

    Matthew Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. InProceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Se- curity (CCS), pages 1322–1333, 2015. 6

    work page 2015

  8. [8]

    Video-based surgical skill assessment using 3d convolu- tional neural networks

    Ilja Funke, Simon T Mees, J ¨urgen Weitz, and Stefanie Spei- del. Video-based surgical skill assessment using 3d convolu- tional neural networks. InInternational Conference on Med- ical Image Computing and Computer-Assisted Intervention (MICCAI), pages 101–108. Springer, 2019. 6

    work page 2019

Show all 22 references
  1. [9]

    Jhu-isi gesture and skill assessment working set (jigsaws): A surgi- cal activity dataset for human motion modeling

    Yixin Gao, S Swaroop Vedula, Carol E Reiley, Narges Ah- midi, Balakrishnan Varadarajan, Henry C Lin, Lingling Tao, Luca Zappella, Benjamın B ´ejar, David D Yuh, et al. Jhu-isi gesture and skill assessment working set (jigsaws): A surgi- cal activity dataset for human motion mo...

    2014

  2. [10]

    Global rating scales in residency education

    Jean D Gray. Global rating scales in residency education. Academic Medicine, 71(1):S55–63, 1996. 1

    1996

  3. [11]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 3

    2016

  4. [12]

    Vidlpro: A video-language pre- training framework for robotic and laparoscopic surgery

    Mohammadmahdi Honarmand, Muhammad Abdullah Ja- mal, and Omid Mohareri. Vidlpro: A video-language pre- training framework for robotic and laparoscopic surgery. In Advancements In Medical Foundation Models: Explainabil- ity, Robustness, Security, and Beyond, 2024. 6

    2024

  5. [13]

    A vision transformer for decoding surgeon activity from surgical videos.Nature biomedical en- gineering, 7(6):780–796, 2023

    Dani Kiyasseh, Runzhuo Ma, Taseen F Haque, Brian J Miles, Christian Wagner, Daniel A Donoho, Animashree Anandku- mar, and Andrew J Hung. A vision transformer for decoding surgeon activity from surgical videos.Nature biomedical en- gineering, 7(6):780–796, 2023. 6

    2023

  6. [14]

    Transsg: A spatial-temporal transformer for surgical gesture recognition

    Le Ma, Hangyeol Kang, Nadia Magnenat-Thalmann, and Katarzyna Wac. Transsg: A spatial-temporal transformer for surgical gesture recognition. InComputer Graphics Interna- tional Conference, pages 151–165. Springer, 2024. 6

    2024

  7. [15]

    Sentence-bert: Sentence embeddings using siamese bert-networks.arXiv preprint arXiv:1908.10084, 2019

    Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks.arXiv preprint arXiv:1908.10084, 2019. 3

    1908 arXiv

  8. [16]

    Recogni- tion and prediction of surgical gestures and trajectories us- ing transformer models in robot-assisted surgery

    Chang Shi, Yi Zheng, and Ann Majewicz Fey. Recogni- tion and prediction of surgical gestures and trajectories us- ing transformer models in robot-assisted surgery. In2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8017–8024. IEEE, 2022. 6

    2022

  9. [17]

    Membership inference attacks against machine learning models

    Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In2017 IEEE symposium on security and privacy (SP), pages 3–18. IEEE, 2017. 3, 6

    2017

  10. [18]

    Gesture recognition in robotic surgery: a review

    Beatrice van Amsterdam, Matthew J Clarkson, and Danail Stoyanov. Gesture recognition in robotic surgery: a review. IEEE Transactions on Biomedical Engineering, 68(6), 2021. 1

    2021

  11. [19]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017. 3

    2017

  12. [20]

    Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers.Ad- vances in neural information processing systems, 33:5776– 5788, 2020

    Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers.Ad- vances in neural information processing systems, 33:5776– 5788, 2020. 3

    2020

  13. [21]

    From decision to action in surgical autonomy: Multi-modal large language models for robot-assisted blood suction.IEEE Robotics and Automation Letters, 2025

    Sadra Zargarzadeh, Maryam Mirzaei, Yafei Ou, and Mahdi Tavakoli. From decision to action in surgical autonomy: Multi-modal large language models for robot-assisted blood suction.IEEE Robotics and Automation Letters, 2025. 5

    2025

  14. [22]

    Automated assessment of surgi- cal skills using frequency analysis.IEEE Transactions on Biomedical Engineering, 65(9):2155–2166, 2018

    Aneeq Zia and Irfan Essa. Automated assessment of surgi- cal skills using frequency analysis.IEEE Transactions on Biomedical Engineering, 65(9):2155–2166, 2018. 6 (a) Non-private baseline (b) Third-party LLM (ID only) Figure 3. Predicted gesture token distributions across surg...

    2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.