pith. sign in

arxiv: 2605.28229 · v1 · pith:LN5MITOSnew · submitted 2026-05-27 · 💻 cs.CV · cs.AI

VidPrism: Heterogeneous Mixture of Experts for Image-to-Video Transfer

Rui Lin , Chuanming Wang , Huadong Ma This is my paper

Pith reviewed 2026-06-29 12:56 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords mixture of expertsvideo understandingimage-to-video transferheterogeneous expertstemporal modelingvision-language modelsexpert specialization
0
0 comments X

The pith

VidPrism deploys specialized experts in a heterogeneous MoE to prevent homogenization in video transfer from image models.

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

The paper proposes VidPrism to address expert homogenization in standard Mixture-of-Experts setups for adapting vision-language models to video tasks. It introduces functionally distinct experts for spatial and temporal roles, supported by a content-aware sampling that creates different input streams. A bidirectional fusion then combines their outputs for better video representations. This leads to state-of-the-art results on recognition benchmarks while promoting genuine specialization among experts. A sympathetic reader would care because it offers a way to make large models more efficient at handling time in videos without all parts doing the same work.

Core claim

VidPrism is a heterogeneous temporal Mixture-of-Experts framework that assigns distinct roles to experts ranging from spatial understanding to temporal modeling, fed by content-aware multi-rate sampling streams and integrated via dynamic bidirectional fusion, achieving superior video representations.

What carries the argument

The heterogeneous temporal Mixture-of-Experts with content-aware multi-rate sampling module and dynamic bidirectional fusion mechanism.

Load-bearing premise

That deploying specialized experts with multi-rate inputs will overcome homogenization without creating new training instabilities or requiring heavy tuning.

What would settle it

Running the model on a benchmark and finding no performance gain over standard MoE or observing that experts produce similar outputs instead of distinct ones.

Figures

Figures reproduced from arXiv: 2605.28229 by Chuanming Wang, Huadong Ma, Rui Lin.

Figure 1
Figure 1. Figure 1: Visualization of inter-frame attention distribution. We select a video from the Dunk category in UCF-101 [30] to compare inter-frame attention between MoTE and our method. Using the same four-frame configurations for fairness, our model exhibits clear attention peaks at key moments, while the homoge￾neous MoTE baseline shows flat, indistinct distributions. stream vision tasks. Therefore, significant perfor… view at source ↗
Figure 2
Figure 2. Figure 2: An overview of VidPrism: (1) A visual encoder processes the input video to extract a sequence of frame-level features. (2) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: A pipeline of STAM: (1) Input features are stratified into [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of Expert Usage on Training and Test Sets. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
read the original abstract

With the rapid development of pre-training technologies, adapting large-scale Vision-Language Models (VLMs) for video understanding \emph{\ie} image-to-video transfer learning has become a dominant paradigm. To achieve superior performance, it raises as an effective strategy among recent advances to employ Mixture-of-Experts (MoE) to enhance VLMs' temporal modeling capabilities. However, conventional MoE designs suffer from expert homogenization, where all experts act as identical generalists, inefficiently learning spatio-temporal features from undifferentiated video streams. To overcome this problem, we propose VidPrism, a novel heterogeneous temporal Mixture-of-Experts framework. VidPrism pioneers a division of labor by deploying functionally specialized experts, each assuming a role ranging from spatial understanding to temporal modeling. To feed these specialists appropriately, we introduce a content-aware, multi-rate sampling module that dynamically generates streams ranging from semantically rich to motion-focused representations, providing specialized inputs for experts. Furthermore, a dynamic, bidirectional fusion mechanism enables synergistic information exchange between these pathways, leading to a comprehensive video representation. Extensive experiments on various video recognition benchmarks demonstrate that VidPrism achieves state-of-the-art performance and effectively fosters expert specialization. Our source code is available at \href{https://github.com/Lrrrr549/VidPrism.git}{https://github.com/Lrrrr549/VidPrism.git}.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper proposes VidPrism, a heterogeneous temporal Mixture-of-Experts (MoE) architecture for image-to-video transfer in Vision-Language Models. It claims to overcome expert homogenization in standard MoE designs by deploying functionally specialized experts (spatial to temporal roles), a content-aware multi-rate sampling module that produces semantically rich to motion-focused streams, and a dynamic bidirectional fusion mechanism for synergistic exchange. The authors assert that extensive experiments show state-of-the-art results on video recognition benchmarks together with effective expert specialization, and release the source code.

Significance. If the central mechanism claims hold, the work could advance MoE usage in temporal video modeling by providing a concrete way to induce division of labor rather than homogenization. The public code release supports reproducibility and is a clear strength. However, the significance remains provisional given the absence of quantitative signatures for specialization in the presented material.

major comments (2)
  1. [Abstract] Abstract: the claim that VidPrism 'effectively fosters expert specialization' is load-bearing for the 'overcomes homogenization' thesis, yet the text supplies no quantitative signatures (routing entropy, per-expert task breakdown, or controlled ablation against a homogeneous MoE given identical streams) that would confirm the mechanism rather than capacity gains or routing changes.
  2. [Abstract] Abstract: the assertion of 'state-of-the-art performance' on 'various video recognition benchmarks' is presented without any numerical results, specific dataset names, or baseline comparisons, preventing assessment of whether the reported gains are attributable to the proposed heterogeneous design.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the constructive review and for highlighting areas where the presentation can be strengthened. We address each major comment below and will update the manuscript to improve clarity and evidentiary support.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that VidPrism 'effectively fosters expert specialization' is load-bearing for the 'overcomes homogenization' thesis, yet the text supplies no quantitative signatures (routing entropy, per-expert task breakdown, or controlled ablation against a homogeneous MoE given identical streams) that would confirm the mechanism rather than capacity gains or routing changes.

    Authors: We agree that the abstract itself does not contain these quantitative signatures. The body of the manuscript includes ablation studies and routing visualizations that illustrate differences in expert behavior under the heterogeneous design. To make the specialization claim more robust and directly responsive to the concern, we will add explicit quantitative metrics (routing entropy, per-expert breakdown, and a homogeneous-MoE control with matched streams) in the revised version. revision: yes

  2. Referee: [Abstract] Abstract: the assertion of 'state-of-the-art performance' on 'various video recognition benchmarks' is presented without any numerical results, specific dataset names, or baseline comparisons, preventing assessment of whether the reported gains are attributable to the proposed heterogeneous design.

    Authors: The abstract is intentionally concise and therefore omits specific numbers and dataset names. The experimental section of the manuscript reports results on the relevant benchmarks together with baseline comparisons. To allow readers to evaluate the abstract claim immediately, we will revise the abstract to include key numerical results, dataset references, and a brief statement on the role of the heterogeneous design. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical architecture proposal with no derivation chain or fitted predictions shown

full rationale

The paper describes a new MoE architecture (specialized experts + multi-rate sampling + bidirectional fusion) and asserts SOTA results plus specialization from experiments. No equations, no parameter-fitting steps, no predictions of quantities from fitted inputs, and no self-citation chains are present in the supplied text. The central claim rests on empirical validation rather than any closed mathematical loop, making the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no free parameters, axioms, or invented entities can be extracted or audited from the provided text.

pith-pipeline@v0.9.1-grok · 5770 in / 1080 out tokens · 30285 ms · 2026-06-29T12:56:21.509295+00:00 · methodology

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

54 extracted references · 12 canonical work pages · 5 internal anchors

  1. [1]

    Vivit: A video vi- sion transformer

    Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lu ˇci´c, and Cordelia Schmid. Vivit: A video vi- sion transformer. InICCV, pages 6836–6846, 2021. 2

    2021

  2. [2]

    Layer Normalization

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization.arXiv:1607.06450, 2016. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  3. [3]

    Frozen feature augmentation for few-shot image classification

    Andreas B ¨ar, Neil Houlsby, Mostafa Dehghani, and Manoj Kumar. Frozen feature augmentation for few-shot image classification. InCVPR, pages 16046–16057, 2024. 1

    2024

  4. [4]

    Is space-time attention all you need for video understanding? InICML, page 4, 2021

    Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? InICML, page 4, 2021. 2

    2021

  5. [5]

    Quo vadis, action recognition? a new model and the kinetics dataset

    Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. InCVPR, pages 6299–6308, 2017. 2, 6

    2017

  6. [6]

    Ef- ficient transfer learning for video-language foundation mod- els

    Haoxing Chen, Zizheng Huang, Yan Hong, Yanshuo Wang, Zhongcai Lyu, Zhuoer Xu, Jun Lan, and Zhangxuan Gu. Ef- ficient transfer learning for video-language foundation mod- els. InCVPR, pages 29129–29138, 2025. 1, 2, 6, 7

    2025

  7. [7]

    Ost: Refining text knowledge with optimal spatio-temporal descriptor for general video recog- nition

    Tongjia Chen, Hongshan Yu, Zhengeng Yang, Zechuan Li, Wei Sun, and Chen Chen. Ost: Refining text knowledge with optimal spatio-temporal descriptor for general video recog- nition. InCVPR, pages 18888–18898, 2024. 6, 7

    2024

  8. [8]

    Multiscale vision transformers

    Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. InICCV, pages 6824–6835,

  9. [9]

    Convolutional two-stream network fusion for video action recognition

    Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. InCVPR, pages 1933–1941, 2016. 2

    1933

  10. [10]

    Slowfast networks for video recognition

    Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202–6211, 2019. 2

    2019

  11. [11]

    Separate visual pathways for perception and action.Trends in neurosciences, 15(1):20–25, 1992

    Melvyn A Goodale and A David Milner. Separate visual pathways for perception and action.Trends in neurosciences, 15(1):20–25, 1992. 2

    1992

  12. [12]

    The” something something” video database for learning and evaluating visual common sense

    Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michal- ski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The” something something” video database for learning and evaluating visual common sense. InICCV, pages 5842–5850, 2017. 6

    2017

  13. [13]

    Focus: Knowledge-enhanced adaptive visual compression for few- shot whole slide image classification

    Zhengrui Guo, Conghao Xiong, Jiabo Ma, Qichen Sun, Lishuang Feng, Jinzhuo Wang, and Hao Chen. Focus: Knowledge-enhanced adaptive visual compression for few- shot whole slide image classification. InCVPR, pages 15590–15600, 2025. 1

    2025

  14. [14]

    Tube convolu- tional neural network (t-cnn) for action detection in videos

    Rui Hou, Chen Chen, and Mubarak Shah. Tube convolu- tional neural network (t-cnn) for action detection in videos. InICCV, pages 5822–5831, 2017. 2

    2017

  15. [15]

    Prompting visual-language models for efficient video understanding

    Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. InECCV, pages 105–124. Springer, 2022. 7

    2022

  16. [16]

    Coarse-fine net- works for temporal activity detection in videos

    Kumara Kahatapitiya and Michael S Ryoo. Coarse-fine net- works for temporal activity detection in videos. InCVPR, pages 8385–8394, 2021. 2

    2021

  17. [17]

    Kuehne, H

    H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre. Hmdb: A large video database for human motion recogni- tion. InICCV, pages 2556–2563, 2011. 6

    2011

  18. [18]

    Tsm: Temporal shift module for efficient video understanding

    Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. InICCV, pages 7083–7093, 2019. 2

    2019

  19. [19]

    Frozen clip models are efficient video learners

    Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hong- sheng Li. Frozen clip models are efficient video learners. In ECCV, pages 388–404. Springer, 2022. 2

    2022

  20. [20]

    Revisiting temporal modeling for clip-based image-to-video knowledge transferring

    Ruyang Liu, Jingjia Huang, Ge Li, Jiashi Feng, Xinglong Wu, and Thomas H Li. Revisiting temporal modeling for clip-based image-to-video knowledge transferring. InCVPR, pages 6555–6564, 2023. 6

    2023

  21. [21]

    Video swin transformer

    Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer. InCVPR, pages 3202–3211, 2022. 2

    2022

  22. [22]

    Khan, and Fahad Shahbaz Khan

    Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fa- had Khan. Videogpt+: Integrating image and video encoders for enhanced video understanding.arXiv:2406.09418, 2024. 2

    work page arXiv 2024

  23. [23]

    Expanding language-image pretrained models for gen- eral video recognition

    Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for gen- eral video recognition. InECCV, pages 1–18. Springer,

  24. [24]

    St-adapter: Parameter-efficient image-to-video transfer learning.NeurIPS, 35:26462–26477, 2022

    Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hong- sheng Li. St-adapter: Parameter-efficient image-to-video transfer learning.NeurIPS, 35:26462–26477, 2022. 2, 6

    2022

  25. [25]

    Dual- path adaptation from image to video transformers

    Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Dual- path adaptation from image to video transformers. InCVPR, pages 2203–2213, 2023. 6

    2023

  26. [26]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. InICML, pages 8748–8763. PmLR, 2021. 1, 2, 6, 7

    2021

  27. [27]

    Fine-tuned clip models are efficient video learners

    Hanoona Rasheed, Muhammad Uzair Khattak, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Fine-tuned clip models are efficient video learners. InCVPR, pages 6545–6554, 2023. 7

    2023

  28. [28]

    Slow-fast architecture for video multi-modal large lan- guage models.arXiv:2504.01328, 2025

    Min Shi, Shihao Wang, Chieh-Yun Chen, Jitesh Jain, Kai Wang, Junjun Xiong, Guilin Liu, Zhiding Yu, and Humphrey Shi. Slow-fast architecture for video multi-modal large lan- guage models.arXiv:2504.01328, 2025. 2

    work page arXiv 2025

  29. [29]

    Two-stream convolutional networks for action recognition in videos

    Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. NeurIPS, 27, 2014. 2

    2014

  30. [30]

    UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild

    Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild.arXiv:1212.0402, 2012. 1, 6

    work page internal anchor Pith review Pith/arXiv arXiv 2012

  31. [31]

    GLM-4.5V and GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning

    V Team, Wenyi Hong, Wenmeng Yu, et al. Glm-4.5 v and glm-4.1 v-thinking: Towards versatile multimodal reason- ing with scalable reinforcement learning.arXiv:2507.01006,

    work page internal anchor Pith review Pith/arXiv arXiv

  32. [32]

    A closer look at spatiotemporal convolutions for action recognition

    Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. InCVPR, pages 6450– 6459, 2018. 2

    2018

  33. [33]

    Step-3 is large yet afford- able: Model-system co-design for cost-effective decoding

    Bin Wang, Bojun Wang, Changyi Wan, Guanzhe Huang, Hanpeng Hu, Haonan Jia, Hao Nie, Mingliang Li, Nuo Chen, Siyu Chen, et al. Step-3 is large yet afford- able: Model-system co-design for cost-effective decoding. arXiv:2507.19427, 2025. 2

    work page arXiv 2025

  34. [34]

    Action recognition with improved trajectories

    Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. InICCV, pages 3551–3558, 2013. 2

    2013

  35. [35]

    Grounded-videollm: Sharpening fine- grained temporal grounding in video large language models

    Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yu- fan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine- grained temporal grounding in video large language models. arXiv:2410.03290, 2024. 2

    work page arXiv 2024

  36. [36]

    Temporal segment networks: Towards good practices for deep action recogni- tion

    Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaoou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recogni- tion. InECCV, pages 20–36. Springer, 2016. 2

    2016

  37. [37]

    Videomae v2: Scaling video masked autoencoders with dual masking

    Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi- nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In CVPR, pages 14549–14560, 2023. 7

    2023

  38. [38]

    arXiv preprint arXiv:2109.08472 (2021)

    Mengmeng Wang, Jiazheng Xing, and Yong Liu. Ac- tionclip: A new paradigm for video action recognition. arXiv:2109.08472, 2021. 2, 6, 7

    work page arXiv 2021

  39. [39]

    M2-clip: A multimodal, multi-task adapting framework for video action recognition.AAAI, 2024

    Mengmeng Wang, Jiazheng Xing, Boyuan Jiang, Jun Chen, Jianbiao Mei, Xingxing Zuo, Guang Dai, Jingdong Wang, and Yong Liu. M2-clip: A multimodal, multi-task adapting framework for video action recognition.AAAI, 2024. 6

    2024

  40. [40]

    Action detail matters: Refining video recognition with local action queries

    Mengmeng Wang, Zeyi Huang, Xiangjie Kong, Guojiang Shen, Guang Dai, Jingdong Wang, and Yong Liu. Action detail matters: Refining video recognition with local action queries. InCVPR, pages 19132–19142, 2025. 2, 6

    2025

  41. [41]

    Temporal pyramid pooling-based convolutional neural network for action recognition.IEEE TCSVT, 27(12):2613–2622, 2016

    Peng Wang, Yuanzhouhan Cao, Chunhua Shen, Lingqiao Liu, and Heng Tao Shen. Temporal pyramid pooling-based convolutional neural network for action recognition.IEEE TCSVT, 27(12):2613–2622, 2016. 2

    2016

  42. [42]

    Internvid: A large-scale video-text dataset for multimodal understanding and generation.ICLR, 2023

    Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation.ICLR, 2023. 1, 2, 7

    2023

  43. [43]

    InternVideo2.5: Empowering Video MLLMs with Long and Rich Context Modeling

    Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xi- angyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Internvideo2. 5: Empowering video mllms with long and rich context modeling.arXiv:2501.12386,

    work page internal anchor Pith review Pith/arXiv arXiv

  44. [44]

    Vita-clip: Video and text adaptive clip via multimodal prompting

    Syed Talal Wasim, Muzammal Naseer, Salman Khan, Fa- had Shahbaz Khan, and Mubarak Shah. Vita-clip: Video and text adaptive clip via multimodal prompting. InCVPR, pages 23034–23044, 2023. 6

    2023

  45. [45]

    Revisiting clas- sifier: Transferring vision-language models for video recog- nition

    Wenhao Wu, Zhun Sun, and Wanli Ouyang. Revisiting clas- sifier: Transferring vision-language models for video recog- nition. InAAAI, pages 2847–2855, 2023. 1, 6

    2023

  46. [46]

    Videoclip: Contrastive pre-training for zero-shot video-text understanding

    Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. InEMNLP, 2021. 1, 2

    2021

  47. [47]

    Slowfast-llava: A strong training-free baseline for video large language models.arXiv:2407.15841, 2024

    Mingze Xu, Mingfei Gao, Zhe Gan, Hong-You Chen, Zhengfeng Lai, Haiming Gang, Kai Kang, and Afshin De- hghan. Slowfast-llava: A strong training-free baseline for video large language models.arXiv:2407.15841, 2024. 2

    work page arXiv 2024

  48. [48]

    Slowfast-llava-1.5: A family of token- efficient video large language models for long-form video understanding.arXiv:2503.18943, 2025

    Mingze Xu, Mingfei Gao, Shiyu Li, Jiasen Lu, Zhe Gan, Zhengfeng Lai, Meng Cao, Kai Kang, Yinfei Yang, and Afshin Dehghan. Slowfast-llava-1.5: A family of token- efficient video large language models for long-form video understanding.arXiv:2503.18943, 2025. 2

    work page arXiv 2025

  49. [49]

    Larger norm more transferable: An adaptive feature norm approach for unsupervised domain adaptation

    Ruijia Xu, Guanbin Li, Jihan Yang, and Liang Lin. Larger norm more transferable: An adaptive feature norm approach for unsupervised domain adaptation. InICCV, pages 1426– 1435, 2019. 3

    2019

  50. [50]

    Qwen3 Technical Report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chen- gen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv:2505.09388, 2025. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  51. [51]

    Temporal pyramid network for action recognition

    Ceyuan Yang, Yinghao Xu, Jianping Shi, Bo Dai, and Bolei Zhou. Temporal pyramid network for action recognition. In CVPR, pages 591–600, 2020. 2

    2020

  52. [52]

    Aim: Adapting image models for efficient video action recognition.ICLR, 2023

    Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. Aim: Adapting image models for efficient video action recognition.ICLR, 2023. 2, 6

    2023

  53. [53]

    Rethink- ing the smaller-norm-less-informative assumption in channel pruning of convolution layers.ICLR, 2018

    Jianbo Ye, Xin Lu, Zhe Lin, and James Z Wang. Rethink- ing the smaller-norm-less-informative assumption in channel pruning of convolution layers.ICLR, 2018. 3

    2018

  54. [54]

    Mote: Reconciling generalization with specialization for visual-language to video knowledge transfer.NeurIPS, 37: 55403–55424, 2024

    Minghao Zhu, Zhengpu Wang, Mengxian Hu, Ronghao Dang, Xiao Lin, Xun Zhou, Chengju Liu, and Qijun Chen. Mote: Reconciling generalization with specialization for visual-language to video knowledge transfer.NeurIPS, 37: 55403–55424, 2024. 1, 2, 6, 7

    2024