Pith. sign in

REVIEW 3 major objections 5 minor 61 references

This paper claims that adding CBAM attention and DIoU loss to YOLO11n yields a gesture detector with 98.9% precision, and that pairing it with cloud LLM/VLM agents lets a small robot follow single, composite, and vision-dependent commands.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 01:22 UTC pith:QTSIWULW

load-bearing objection Competent integration of known components; the headline detector metrics are likely optimistic because the same validation split is used for model selection and reporting. the 3 major comments →

arxiv 2607.14675 v3 pith:QTSIWULW submitted 2026-07-16 cs.RO cs.AI

An Intelligent-Cloud Edge Multimodal Interaction System for Robots

Zihan Guo , Xiaoqi Li This is my paper
classification cs.RO cs.AI MSC 68T4068T45
keywords gesture detectionhuman-robot interactionYOLO-DCCBAMDIoU losscloud-edge architecturemultimodal agentstask planning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to show that a resource-constrained robot can handle rich multimodal human-robot interaction if heavy perception and reasoning are pushed to the cloud while the robot keeps only sensing, communication, and motion control. Its central empirical claim is that a small modification to YOLO11n — CBAM attention in the neck and Distance-IoU (DIoU) loss — produces a gesture detector (YOLO-DC) that reaches 98.9% precision and 90.7% mean average precision at 0.5 IoU threshold (mAP@0.5) on a public dataset, and 95.0% precision and 92.7% mAP@0.5 on a custom dataset, outperforming the baseline by up to 16 percentage points in precision. At the system level, the paper reports 95%, 88%, and 82% task success for single-action, composite-action, and vision-dependent commands on the TonyPi robot, along with a mean user satisfaction of 3.69 out of 5. A sympathetic reader would care because this is a concrete demonstration that LLM/VLM-style reasoning can be made available to a small robot through a cloud-edge split rather than onboard deployment.

Core claim

YOLO-DC is a one-stage detector made by inserting the Convolutional Block Attention Module (CBAM) into the neck of YOLO11n and swapping the regression objective to Distance-IoU (DIoU) loss. The paper argues that CBAM's channel and spatial attention sharpen features for small or partially occluded gestures in cluttered scenes, while DIoU's center-distance penalty keeps gradients informative even when boxes do not overlap, improving localization. Around this detector, the system runs a cloud-edge pipeline: YOLO-DC, a vision-language model for scene understanding, and a large language model for intent parsing and planning all run in the cloud; a rule engine and finite-state machines validate ou

What carries the argument

The load-bearing pieces are (1) CBAM, an attention module that sequentially applies channel and spatial attention to re-weight feature maps; (2) DIoU loss, which adds a normalized center-distance penalty to standard IoU; and (3) the cloud-edge dual-agent pipeline with a unified JSON schema, a rule engine, and finite-state machines that turn perception output into validated action sequences. The CBAM+DIoU pair is what the paper credits for the detection gains; the JSON/FSM layer is what makes the multimodal agents safe enough to drive a physical robot.

Load-bearing premise

The headline accuracy numbers are computed on the validation split that was also used for model selection, so they may reflect selection on that split rather than true generalization to new data.

What would settle it

Hold out a completely unseen test set (or use k-fold cross-validation) and report YOLO-DC's precision and mAP@0.5; if the numbers fall well below 98.9%/95.0% precision and 90.7%/92.7% mAP@0.5, the paper's central detection claim is not supported. Similarly, reproducing the system-level 95%/88%/82% success rates on a fresh set of commands and users would test the interaction claim.

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

If this is right

  • If the detection results hold, a lightweight one-stage detector can be upgraded to 98.9%/95.0% precision with roughly 0.1 million extra parameters, small enough for edge-friendly deployment.
  • If the system-level results hold, cloud-offloaded LLM/VLM agents can drive a small robot through single-action, composite-action, and vision-dependent tasks at 95%, 88%, and 82% success, with a rule engine rejecting about 90% of invalid command combinations.
  • The architecture suggests a reusable division of labor: keep sensing and motion control on the robot, and move all heavy inference and reasoning to the cloud, connected by structured JSON messages.
  • The reported mean user satisfaction of 3.69/5, combined with sub-5% response repetition, indicates the interaction is acceptable but with room for improvement in naturalness and reliability.

Where Pith is reading between the lines

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

  • Because the reported precision numbers come from the same validation split used to choose the model, the true out-of-sample performance could be lower; an independent test set or cross-validation would be needed to confirm the 16-point precision gain over YOLO11n.
  • The cloud-edge split trades latency and privacy for compute; the paper measures VLM response latency at about 210 ms, but in real deployments network bandwidth and cloud cost would likely be the practical bottlenecks.
  • The recipe of 'lightweight detector + attention + cloud reasoning' may transfer to other edge-robot perception tasks, such as object or person detection, where onboard compute is the constraint.
  • The paper's Section 3.3.2 states that the validation mechanisms are functional checks rather than a complete cybersecurity assessment; whether the JSON protocol would survive adversarial or malformed inputs is left untested.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper proposes a cloud–edge multimodal interaction system for a TonyPi humanoid robot. The main technical component is YOLO-DC, a YOLO11n detector modified by inserting CBAM into the neck and replacing the bounding-box regression loss with DIoU. The cloud layer runs YOLO-DC, a VLM for scene understanding, and an LLM for intent interpretation and task-planning; outputs are fused and validated by a rule engine and FSMs before being sent as structured JSON to the robot. Reported experiments include detector comparisons on a 600-image public dataset and a custom dataset of unreported size, an ablation on the public dataset, physical task-execution success rates, and a 30-participant satisfaction study. The headline numbers are precision 98.9%/95.0% and mAP@0.5 90.7%/92.7%, with system success rates of 95%, 88%, and 82%.

Significance. If the quantitative claims can be supported, this is a useful systems-integration paper: it combines a standard attention/loss modification with off-the-shelf LLM/VLM components, evaluates by physical execution rather than simulation, includes an ablation, and introduces message-level validation and command whitelisting. The detector modification is incremental, and the 'cloud–edge' split is mostly a thin-client arrangement, but the end-to-end demonstration is of potential interest to the robotics community. The current evaluation, however, does not yet establish the claimed improvements: the detector metrics are computed on the validation split that was also used for model selection, and the system-level results lack trial counts and error estimates. The central quantitative claims are therefore conditional on a substantially strengthened experimental protocol.

major comments (3)
  1. [§4.1.1, Table 1] The paper states: 'The validation subset is used for model selection and for reporting the results in this paper.' Because the same split is used both to select the model (including the CBAM/DIoU choices and checkpoint selection) and to compute the headline numbers, the precision/mAP values in Table 1 are selection statistics rather than unbiased estimates of generalization. The custom-dataset improvement over YOLO11n (precision 79.0→95.0, mAP 86.7→92.7) is especially large relative to the public-dataset gains, and the custom dataset size is never reported. A held-out test set, multiple seeds, and confidence intervals/error bars are required before the claimed detector improvements can be accepted.
  2. [§4.3.1, Table 3] Task-execution success rates are reported as 95%, 88%, and 82%, but no trial counts are given, success is not defined, and no error bars or statistical comparisons are provided. The statement that the rule engine rejects 90% of tested invalid action combinations also lacks the number of tests. Without these details the system-level feasibility claim cannot be evaluated; the impact of detector uncertainty on the system-level numbers is therefore unknown.
  3. [§4.2.2, Table 2] The ablation is a single run on the public dataset only. Adding CBAM alone changes mAP@0.5 by +0.5 points (88.4→88.9) while slightly decreasing precision and recall; at this scale the effects are within plausible random-seed/optimization noise. Multiple-seed runs with standard deviations (or paired significance tests) are needed to support the claim that both modifications contribute to the final YOLO-DC result.
minor comments (5)
  1. [§4.1.1] The custom dataset's total number of images, class distribution, and annotation protocol are omitted. Since the custom dataset is one of the two central evaluation sets, this information should be provided.
  2. [§3.2.1] The statement that CBAM adds 'approximately 0.1 million parameters' cannot be verified from the paper. Provide a small model-complexity table (parameters/GFLOPs) for YOLO11n and YOLO-DC.
  3. [§3.2.2, Eq. (3)] The notation ρ(b,b gt) should be typeset as ρ(b,b^gt), and the symbols b and b^gt should be defined as bounding-box center coordinates. Minor grammatical issues include 'The proposed detector, incorporates' in the Abstract.
  4. [§4.3.2] The user-satisfaction study lacks details on participant recruitment, scenario presentation order, and whether ratings are task-specific or overall. The overall mean of 3.69/5 is not clearly positive and should be interpreted in the text.
  5. [§2.3] The security-related discussion cites blockchain/software-security works [57–61] but no corresponding security experiments are reported in the system evaluation. Consider shortening this part or adding a basic security validation.

Circularity Check

0 steps flagged

No circular derivation: YOLO-DC is evaluated against external baselines and the system is tested by physical execution; the few self-citations are tangential and non-load-bearing.

full rationale

The paper's claimed derivation chain is self-contained. YOLO-DC is constructed by adding CBAM to the neck of YOLO11n and replacing the regression loss with DIoU, and its reported gains are obtained by comparison with external baselines (YOLOv5n, YOLOv8n, YOLO11n) plus an ablation study, so the central quantitative claim does not reduce by definition to its inputs. The system-level success rates come from physical task execution on the TonyPi platform and are not derived from the detector metrics in a circular way. The self-citations in §2.3 (references [57]–[61]) motivate message-schema validation by analogy to security research, but they are not used as evidence for the detector's accuracy or the task success rates, so they are not load-bearing. The statement in §4.1.1 that the validation subset is used both for model selection and for reporting results is a legitimate methodological concern about possible selection bias, but it is not a circularity: the reported numbers are not identical to the selection criterion by construction, and the paper does not rename a fitted parameter as a prediction. Thus no circular step can be quoted or exhibited, and the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

Central claim rests on empirical training and evaluation rather than a parameter-free derivation. No ad hoc constants are introduced for the detector beyond the standard training regimen, and no new entities (forces, particles, conserved quantities) are proposed. The key load-bearing assumptions are that the validation split yields unbiased metrics and that cloud LLM/VLM services behave reliably.

axioms (4)
  • standard math DIoU loss (Eq. 3) is a valid bounding-box regression objective, as established by [21]
    The paper's detector-improvement claim depends on DIoU behaving as described; the loss is adopted without modification or re-derivation.
  • domain assumption The public 600-image gesture dataset has correct labels and its 2:1 split is representative
    Section 4.1.1 uses the validation subset for both model selection and reporting; if the split is unrepresentative, headline metrics do not generalize.
  • domain assumption Cloud-provided LLM/VLM inference is available and adequately reliable during system evaluation
    Section 3.3.2 routes perception/reasoning to the cloud; reported success rates depend on cloud service behavior not characterized beyond mean VLM latency.
  • domain assumption YOLO11n is a proper baseline and its default training recipe is stable
    Section 3.2 builds on YOLO11n without independent verification; all comparisons inherit this baseline.

pith-pipeline@v1.3.0-alltime-deepseek · 8412 in / 13550 out tokens · 130140 ms · 2026-08-02T01:22:59.868659+00:00 · methodology

0 comments
read the original abstract

Robust human-robot interaction in complex environments requires accurate gesture perception, semantic scene understanding, and reliable task planning under limited onboard computing resources. This paper presents a cloud-edge multimodal interaction framework that integrates an enhanced YOLO-based gesture detector with coordinated large language model (LLM) and vision-language model (VLM) agents. The proposed detector, incorporates the Convolutional Block Attention Module (CBAM) into the neck and replaces the baseline bounding-box regression objective with Distance-IoU (DIoU) loss. These modifications improve feature discrimination and localization for small or partially occluded gestures in complex backgrounds. The cloud layer performs gesture detection, scene understanding, multimodal fusion, and action planning, whereas the TonyPi robot locally handles data acquisition, communication, action execution, and feedback. Experiments on a public gesture dataset and a custom dataset show that YOLO-DC achieves precision values of 98.9% and 95.0%, with mAP@0.5 values of 90.7% and 92.7%, respectively. System-level evaluation yields success rates of 95%, 88%, and 82% for single-action, composite-action, and vision-dependent tasks. A 30 participant evaluation yields an overall mean satisfaction score of 3.69 out of 5. These results demonstrate the feasibility of combining refined gesture detection with multimodal agents for resource-constrained robotic interaction.

Figures

Figures reproduced from arXiv: 2607.14675 by Xiaoqi Li, Zihan Guo.

Figure 1
Figure 1. Figure 1: Cloud–edge architecture of the proposed multimodal human–robot interaction framework [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Network architecture of YOLO-DC. 4 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The Architecture of CBAM Although the C2PSA module in YOLO11n already provides spatial attention, its feature aggregation may remain insufficient for small gestures. We therefore place CBAM after the C2PSA stage in the neck. Channel attention emphasizes feature channels associated with gesture cues, and spatial attention highlights target regions. According to the model summary, the modification adds appro… view at source ↗
Figure 4
Figure 4. Figure 4: For each dataset, the images are divided into training and validation subsets at a ratio of 2:1. The validation subset is used for model selection and for reporting the results in this paper [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Representative samples from the public dataset. 4.1.2 Implementation Details Experiments are conducted using an NVIDIA GeForce RTX 4090 GPU and PyTorch 2.1.0. The models are trained for 300 epochs with stochastic gradient descent, a momentum of 0.937, a batch size of 32, and an initial learning rate of 0.01. Input images are resized to 640 × 640 pixels. A dropout rate of 0.5 is used in the training configu… view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of detection results on the public dataset: YOLO-DC (left) and YOLO11n (right) [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of detection results on the custom dataset: YOLO-DC (left) and YOLO11n (right). 8 [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Training and validation curves of YOLO-DC on the public dataset. 4.3 System-Level Evaluation The multimodal interaction system is evaluated on the TonyPi platform in terms of task￾execution success, response diversity, and user satisfaction. 4.3.1 Task-Execution Success Rate Three command categories are evaluated: single-action, composite-action, and vision￾dependent tasks [PITH_FULL_IMAGE:figures/full_fi… view at source ↗
Figure 9
Figure 9. Figure 9: English-language example of the scene-understanding and structured-response output generated by the multimodal agent. 4.3.2 Feedback Diversity and User Satisfaction To assess response diversity, 500 responses are generated from 50 predefined templates. The resulting repetition rate is below 5%. In addition, 30 participants rate four predefined interaction scenarios—Wave, Kick Ball, Twist, and Celebrate—on … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

61 extracted references · 13 linked inside Pith

  1. [1]

    You Only Look Once: Unified, Real- Time Object Detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You Only Look Once: Unified, Real- Time Object Detection,” inProc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  2. [2]

    YOLO9000: Better, Faster, Stronger,

    J. Redmon and A. Farhadi, “YOLO9000: Better, Faster, Stronger,” inProc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2017

    2017

  3. [3]

    YOLOv3: An Incremental Improvement,

    J. Redmon and A. Farhadi, “YOLOv3: An Incremental Improvement,”arXiv preprint arXiv:1804.02767, 2018

    Pith/arXiv arXiv 2018

  4. [4]

    YOLOv4: Optimal Speed and Accuracy of Object Detection,

    A. Bochkovskiy, C.-Y. Wang, and H.-Y. M. Liao, “YOLOv4: Optimal Speed and Accuracy of Object Detection,”arXiv preprint arXiv:2004.10934, 2020

    Pith/arXiv arXiv 2004

  5. [5]

    YOLOv6: A Single-Stage Object Detection Framework for Industrial Applica- tions,

    C. Li et al., “YOLOv6: A Single-Stage Object Detection Framework for Industrial Applica- tions,”arXiv preprint arXiv:2209.02976, 2022

    Pith/arXiv arXiv 2022

  6. [6]

    YOLOv7: Trainable Bag-of-Freebies Sets New State-of-the-Art for Real-Time Object Detectors,

    C.-Y. Wang, A. Bochkovskiy, and H.-Y. M. Liao, “YOLOv7: Trainable Bag-of-Freebies Sets New State-of-the-Art for Real-Time Object Detectors,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  7. [7]

    YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information,

    C.-Y. Wang, I.-H. Yeh, and H.-Y. M. Liao, “YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information,”arXiv preprint arXiv:2402.13616, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    YOLOv10: Real-Time End-to-End Object Detection,

    A. Wang et al., “YOLOv10: Real-Time End-to-End Object Detection,”arXiv preprint arXiv:2405.14458, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks,” inAdvances in Neural Information Processing Systems, 2015

    2015

  10. [10]

    SSD: Single Shot MultiBox Detector,

    W. Liu et al., “SSD: Single Shot MultiBox Detector,” inProc. European Conf. Computer Vision (ECCV), 2016. 11

    2016

  11. [11]

    Focal Loss for Dense Object Detection,

    T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, “Focal Loss for Dense Object Detection,” inProc. IEEE Int. Conf. Computer Vision (ICCV), 2017

    2017

  12. [12]

    End-to-End Object Detection with Transformers,

    N. Carion et al., “End-to-End Object Detection with Transformers,” inProc. European Conf. Computer Vision (ECCV), 2020

    2020

  13. [13]

    Deformable DETR: Deformable Transformers for End-to-End Object Detec- tion,

    X. Zhu et al., “Deformable DETR: Deformable Transformers for End-to-End Object Detec- tion,” inProc. Int. Conf. Learning Representations (ICLR), 2021

    2021

  14. [14]

    Mask R-CNN,

    K. He, G. Gkioxari, P. Dollár, and R. Girshick, “Mask R-CNN,” inProc. IEEE Int. Conf. Computer Vision (ICCV), 2017

    2017

  15. [15]

    CBAM: Convolutional Block Attention Module,

    S. Woo, J. Park, J.-Y. Lee, and I. S. Kweon, “CBAM: Convolutional Block Attention Module,” inProc. European Conf. Computer Vision (ECCV), 2018

    2018

  16. [16]

    Squeeze-and-Excitation Networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-Excitation Networks,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  17. [17]

    Attention Is All You Need,

    A. Vaswani et al., “Attention Is All You Need,” inAdvances in Neural Information Processing Systems, 2017

    2017

  18. [18]

    Non-Local Neural Networks,

    X. Wang, R. Girshick, A. Gupta, and K. He, “Non-Local Neural Networks,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  19. [19]

    Coordinate Attention for Efficient Mobile Network Design,

    Q. Hou, D. Zhou, and J. Feng, “Coordinate Attention for Efficient Mobile Network Design,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  20. [20]

    Generalized Intersection over Union: A Metric and a Loss for Bounding Box Regression,

    H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese, “Generalized Intersection over Union: A Metric and a Loss for Bounding Box Regression,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  21. [21]

    Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression,

    Z. Zheng, P. Wang, W. Liu, J. Li, R. Ye, and D. Ren, “Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression,” inProc. AAAI Conf. Artificial Intelligence, 2020

    2020

  22. [22]

    SIoU Loss: More Powerful Learning for Bounding Box Regression,

    Z. Gevorgyan, “SIoU Loss: More Powerful Learning for Bounding Box Regression,”arXiv preprint arXiv:2205.12740, 2022

    Pith/arXiv arXiv 2022

  23. [23]

    Focaler-IoU: More Focused Intersection over Union Loss,

    Y. Zhang, “Focaler-IoU: More Focused Intersection over Union Loss,”arXiv preprint, 2024

    2024

  24. [24]

    UnitBox: An Advanced Object Detection Network,

    J. Yu, Y. Jiang, Z. Wang, Z. Cao, and T. Huang, “UnitBox: An Advanced Object Detection Network,” inProc. ACM Int. Conf. Multimedia, 2016

    2016

  25. [25]

    Language Models Are Few-Shot Learners,

    T. B. Brown et al., “Language Models Are Few-Shot Learners,” inAdvances in Neural Information Processing Systems, 2020

    2020

  26. [26]

    LLaMA: Open and Efficient Foundation Language Models,

    H. Touvron et al., “LLaMA: Open and Efficient Foundation Language Models,”arXiv preprint arXiv:2302.13971, 2023

    Pith/arXiv arXiv 2023

  27. [27]

    Llama 2: Open Foundation and Fine-Tuned Chat Models,

    H. Touvron et al., “Llama 2: Open Foundation and Fine-Tuned Chat Models,”arXiv preprint arXiv:2307.09288, 2023

    Pith/arXiv arXiv 2023

  28. [28]

    GPT-4 Technical Report,

    OpenAI, “GPT-4 Technical Report,”arXiv preprint arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  29. [29]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,

    J. Wei et al., “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,” inAdvances in Neural Information Processing Systems, 2022

    2022

  30. [30]

    PaLM: Scaling Language Modeling with Pathways,

    A. Chowdhery et al., “PaLM: Scaling Language Modeling with Pathways,”arXiv preprint arXiv:2204.02311, 2022. 12

    Pith/arXiv arXiv 2022

  31. [31]

    LearningTransferableVisualModelsfromNaturalLanguageSupervision,

    A.Radfordetal., “LearningTransferableVisualModelsfromNaturalLanguageSupervision,” inProc. Int. Conf. Machine Learning (ICML), 2021

    2021

  32. [32]

    BLIP: Bootstrapping Language-Image Pre-Training for Unified Vision-Language Understanding and Generation,

    J. Li et al., “BLIP: Bootstrapping Language-Image Pre-Training for Unified Vision-Language Understanding and Generation,” inProc. Int. Conf. Machine Learning (ICML), 2022

    2022

  33. [33]

    BLIP-2: Bootstrapping Language-Image Pre-Training with Frozen Image Encoders and Large Language Models,

    J. Li et al., “BLIP-2: Bootstrapping Language-Image Pre-Training with Frozen Image Encoders and Large Language Models,” inProc. Int. Conf. Machine Learning (ICML), 2023

    2023

  34. [34]

    Visual Instruction Tuning,

    H. Liu et al., “Visual Instruction Tuning,” inAdvances in Neural Information Processing Systems, 2023

    2023

  35. [35]

    Improved Baselines with Visual Instruction Tuning,

    H. Liu et al., “Improved Baselines with Visual Instruction Tuning,”arXiv preprint arXiv:2310.03744, 2023

    Pith/arXiv arXiv 2023

  36. [36]

    Flamingo: A Visual Language Model for Few-Shot Learning,

    J.-B. Alayrac et al., “Flamingo: A Visual Language Model for Few-Shot Learning,” in Advances in Neural Information Processing Systems, 2022

    2022

  37. [37]

    MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models,

    D. Zhu et al., “MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models,”arXiv preprint arXiv:2304.10592, 2023

    Pith/arXiv arXiv 2023

  38. [38]

    AnImageIsWorth16 ×16Words: TransformersforImageRecognition at Scale,

    A.Dosovitskiyetal., “AnImageIsWorth16 ×16Words: TransformersforImageRecognition at Scale,” inProc. Int. Conf. Learning Representations (ICLR), 2021

    2021

  39. [39]

    Masked Autoencoders Are Scalable Vision Learners,

    K. He et al., “Masked Autoencoders Are Scalable Vision Learners,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  40. [40]

    Do As I Can, Not As I Say: Grounding Language in Robotic Affordances,

    M. Ahn et al., “Do As I Can, Not As I Say: Grounding Language in Robotic Affordances,” inProc. Conf. Robot Learning (CoRL), 2022

    2022

  41. [41]

    VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models,

    W. Huang, C. Wang, R. Zhang, Y. Li, J. Wu, and L. Fei-Fei, “VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models,” inProc. Conf. Robot Learning (CoRL), 2023

    2023

  42. [42]

    RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control,

    A. Brohan et al., “RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control,” inProc. Conf. Robot Learning (CoRL), 2023

    2023

  43. [43]

    CLIPort: What and Where Pathways for Robotic Manipulation,

    M. Shridhar, L. Manuelli, and D. Fox, “CLIPort: What and Where Pathways for Robotic Manipulation,” inProc. Conf. Robot Learning (CoRL), 2022

    2022

  44. [44]

    ReAct: Synergizing Reasoning and Acting in Language Models,

    S. Yao et al., “ReAct: Synergizing Reasoning and Acting in Language Models,” inProc. Int. Conf. Learning Representations (ICLR), 2023

    2023

  45. [45]

    Generative Agents: Interactive Simulacra of Human Behavior,

    J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein, “Generative Agents: Interactive Simulacra of Human Behavior,” inProc. ACM Symp. User Interface Software and Technology (UIST), 2023

    2023

  46. [46]

    Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows,

    Z. Liu et al., “Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows,” inProc. IEEE Int. Conf. Computer Vision (ICCV), 2021

    2021

  47. [47]

    A ConvNet for the 2020s,

    Z. Liu et al., “A ConvNet for the 2020s,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  48. [48]

    Deep Residual Learning for Image Recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in Proc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2016. 13

    2016

  49. [49]

    MobileNetV2: Inverted Residuals and Linear Bottlenecks,

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “MobileNetV2: Inverted Residuals and Linear Bottlenecks,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  50. [50]

    Searching for MobileNetV3,

    A. Howard et al., “Searching for MobileNetV3,” inProc. IEEE Int. Conf. Computer Vision (ICCV), 2019

    2019

  51. [51]

    EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,

    M. Tan and Q. V. Le, “EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,” inProc. Int. Conf. Machine Learning (ICML), 2019

    2019

  52. [52]

    Edge AI: On-Demand Accelerating Deep Neural Network Inference via Edge Computing,

    E. Li, L. Zeng, Z. Zhou, and X. Chen, “Edge AI: On-Demand Accelerating Deep Neural Network Inference via Edge Computing,”IEEE Transactions on Wireless Communications, vol. 19, no. 1, pp. 447–457, 2020

    2020

  53. [53]

    MLPerf Tiny Benchmark,

    C. Banbury et al., “MLPerf Tiny Benchmark,” inNeurIPS Datasets and Benchmarks Track, 2021

    2021

  54. [54]

    Very Deep Convolutional Networks for Large-Scale Image Recognition,

    K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition,” inProc. Int. Conf. Learning Representations (ICLR), 2015

    2015

  55. [55]

    Rethinking the Inception Architecture for Computer Vision,

    C. Szegedy et al., “Rethinking the Inception Architecture for Computer Vision,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  56. [56]

    DINOv2: Learning Robust Visual Features without Supervision,

    M. Oquab et al., “DINOv2: Learning Robust Visual Features without Supervision,”Trans- actions on Machine Learning Research, 2023

    2023

  57. [57]

    Penetration Testing for System Security: Methods and Practical Approaches,

    W. Zhang, J. Xing, and X. Li, “Penetration Testing for System Security: Methods and Practical Approaches,”arXiv preprint arXiv:2505.19174, 2025

    arXiv 2025

  58. [58]

    SmartBugBert: BERT-Enhanced Vulnerability Detection for Smart Contract Bytecode,

    J. Bu, W. Li, Z. Li, Z. Zhang, and X. Li, “SmartBugBert: BERT-Enhanced Vulnerability Detection for Smart Contract Bytecode,”arXiv preprint arXiv:2504.05002, 2025

    arXiv 2025

  59. [59]

    Exploring Vulnerabilities and Concerns in Solana Smart Contracts,

    X. Wu, J. Xing, and X. Li, “Exploring Vulnerabilities and Concerns in Solana Smart Contracts,”arXiv preprint arXiv:2504.07419, 2025

    Pith/arXiv arXiv 2025

  60. [60]

    Interaction-aware vulnerability detection in smart contract bytecodes,

    W. Li, X. Li, Y. Mao, and Y. Zhang, “Interaction-aware vulnerability detection in smart contract bytecodes,”IEEE Transactions on Dependable and Secure Computing, 2025

    2025

  61. [61]

    Penetrating the hostile: Detecting DeFi protocol exploits through cross-contract analysis,

    X. Li, W. Li, Z. Liu, Y. Zhang, and Y. Mao, “Penetrating the hostile: Detecting DeFi protocol exploits through cross-contract analysis,”IEEE Transactions on Information Forensics and Security, vol. 20, pp. 11759–11774, 2025. 14

    2025