pith. machine review for the scientific record. sign in

arxiv: 2605.13202 · v1 · submitted 2026-05-13 · 💻 cs.CV · cs.AI

Recognition: unknown

STAR: Semantic-Temporal Adaptive Representation Learning for Few-Shot Action Recognition

Hongli Liu , Yu Wang , Shengjie Zhao
Authors on Pith no claims yet

Pith reviewed 2026-05-14 20:14 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords few-shot action recognitionsemantic-temporal alignmenttemporal prototype refinementMamba state-space modelcross-modal attentionvideo understandinglimited supervisionvision-language models
0
0 comments X

The pith

STAR resolves semantic-temporal misalignment in few-shot action recognition by aligning frame-level cues with text and refining prototypes via Mamba blocks.

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

The paper establishes that few-shot action recognition improves when models fix the gap between static text prompts and the sparse, time-varying visual signals in videos, while also handling both short bursts and long dependencies in sequences. It introduces a unified framework that performs cross-modal attention at the frame level and refines class prototypes with semantic guidance and state-space modeling. This matters for a sympathetic reader because current vision-language methods often lose decisive moments or blur temporal structure when only a few examples are available. Experiments across five benchmarks show the approach yields higher accuracy in 1-shot settings than prior methods.

Core claim

The authors claim that semantic-temporal misalignment and inadequate multi-scale temporal modeling in few-shot action recognition are addressed by a Temporal Semantic Attention module that performs frame-level cross-modal alignment and a Semantic Temporal Prototype Refiner that integrates semantic-guided Mamba blocks with multi-frequency sampling and bidirectional refinement, together with LLM-derived class descriptors, to produce temporally consistent and discriminative prototypes for novel actions.

What carries the argument

The Temporal Semantic Attention (TSA) mechanism for frame-level semantic alignment and the Semantic Temporal Prototype Refiner (STPR) that uses semantic-guided Mamba blocks with multi-frequency temporal sampling.

If this is right

  • Higher accuracy on SSv2-Full, SSv2-Small, and HMDB51 under the 1-shot protocol compared with prior few-shot methods.
  • Better preservation of both short-term discriminative cues and long-range temporal dependencies in video sequences.
  • Successful transfer of sequence-modeling capabilities from state-space models into the few-shot action recognition setting.
  • Additional long-range semantic guidance supplied by temporally dependent class descriptors from large language models.

Where Pith is reading between the lines

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

  • The same alignment and refinement strategy could be tested on other few-shot video tasks such as temporal action detection or video object segmentation.
  • If the components prove robust, training costs for action recognition systems in domains with scarce labels would decrease.
  • Further gains may appear when the Mamba-based refiner is combined with additional modalities or larger backbone models.

Load-bearing premise

The TSA and STPR components will transfer effectively to unseen real-world video distributions beyond the five evaluated benchmarks.

What would settle it

A controlled test on a held-out video dataset containing action categories with different temporal structures and visual sparsity patterns, where the full STAR framework shows no accuracy gain over a baseline that lacks the TSA and STPR modules.

Figures

Figures reproduced from arXiv: 2605.13202 by Hongli Liu, Shengjie Zhao, Yu Wang.

Figure 1
Figure 1. Figure 1: Motivation. (a) Actions such as Golf Swing and Tennis Swing) differ only in a few decisive frames. (b) Average pooling compresses temporal structures, weakening semantic alignment. (c) Prior methods introduce seman￾tic guidance but lack explicit semantic-temporal interactions. (d) Our STAR explicitly models frame-level semantic-temporal relations to precisely capture key action dynamics. similarity matchin… view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of STAR. Support and query videos are first encoded into frame-level features. The STPR module then applies semantically guided temporal modeling to capture long-range dynamics, followed by the TSA module which aligns the enriched frames with semantic descriptors. Finally, these matching signals are fused for query prediction. Additional support samples are omitted for clarity. query a… view at source ↗
Figure 3
Figure 3. Figure 3: Structure of the Temporal State Space Module (TSSM), built upon [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of the Semantic Temporal Prototype Refiner (STPR) [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Top-1 accuracy (%) comparison of different methods under varying [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: N-way 1-shot performance of our STAR framework and baseline [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of frame-to-class attention maps from the TSA module.The left side shows sampled video frames, while the right displays the corresponding [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Generation of Temporal Class Representations via LLM prompting. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visualization of frame-level importance scores on the HMDB51 dataset. The line graphs illustrate the temporal distribution of importance scores [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: T-SNE visualization of five challenging action classes from the [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
read the original abstract

Few-shot action recognition (FSAR) requires models to generalize to novel action categories from only a handful of annotated samples. Despite progress with vision-language models, existing approaches still suffer from semantic-temporal misalignment, where static textual prompts fail to capture decisive visual cues that appear sparsely across sequences, and from inadequate modeling of multi-scale temporal dynamics, as short-term discriminative cues and long-range dependencies are often either oversmoothed or fragmented. To address these challenges, we propose Semantic Temporal Adaptive Representation Learning (STAR), a unified framework, consisting of a semantic-alignment component and a temporal-aware component, effectively bridging the semantic and temporal gaps and transferring the sequence modeling capability of Mamba into the FSAR. The semantic alignment module introduces a Temporal Semantic Attention (TSA) mechanism, which performs frame-level cross-modal alignment with textual cues, ensuring fine-grained semantic-temporal consistency. The temporal-aware module incorporates a Semantic Temporal Prototype Refiner (STPR) that integrates semantic-guided Mamba blocks with multi-frequency temporal sampling and bidirectional state-space refinement, yielding semantically aligned prototypes with enhanced discriminative fidelity and temporal consistency. Furthermore, temporally dependent class descriptors derived from large language models (LLMs) provide long-range semantic guidance. Extensive experiments on five FSAR benchmarks demonstrate the consistent superiority of STAR over state-of-the-art methods. For instance, STAR achieves up to 8.1% and 6.7% gains on the SSv2-Full and SSv2-Small datasets under the 1-shot setting, and 7.3% on HMDB51, validating its effectiveness under limited supervision. The code is available at https://github.com/HongliLiu1/STAR-main.

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 / 3 minor

Summary. The paper introduces STAR, a unified framework for few-shot action recognition that combines a Temporal Semantic Attention (TSA) module for frame-level cross-modal semantic alignment with a Semantic Temporal Prototype Refiner (STPR) that employs semantic-guided Mamba blocks, multi-frequency temporal sampling, and bidirectional state-space refinement, augmented by LLM-derived temporally dependent class descriptors. It reports consistent outperformance over prior methods on five FSAR benchmarks, with specific gains of up to 8.1% and 6.7% on SSv2-Full and SSv2-Small under the 1-shot protocol and 7.3% on HMDB51.

Significance. If the empirical gains prove robust, the work offers a practical advance in FSAR by explicitly targeting semantic-temporal misalignment through integrated vision-language and state-space modeling components. The public code release provides a direct path for independent verification and extension, which strengthens the contribution relative to purely architectural proposals.

major comments (2)
  1. [§4 (Experiments), Table 2 and Table 3] §4 (Experiments), Table 2 and Table 3: the reported improvements (e.g., 8.1% on SSv2-Full 1-shot) are given as single-point estimates without error bars, standard deviations across random seeds, or statistical significance tests; this leaves open whether the gains attributable to TSA and STPR exceed run-to-run variance and therefore weakens the central performance claim.
  2. [§3.2 (STPR description)] §3.2 (STPR description): the multi-frequency temporal sampling and bidirectional Mamba refinement are presented as key innovations, yet the ablation in Table 4 does not isolate the contribution of the frequency sampling schedule from the underlying Mamba blocks and LLM descriptors, making it impossible to attribute the reported prototype quality gains specifically to the proposed components.
minor comments (3)
  1. [Abstract] Abstract: the phrase 'transferring the sequence modeling capability of Mamba into the FSAR' is imprecise; FSAR is already defined earlier in the sentence, but the transfer mechanism should be stated more explicitly.
  2. [§3.1 (TSA)] §3.1 (TSA): the cross-modal attention formulation would benefit from an explicit equation showing how frame-level visual features are aligned with the LLM-derived textual cues, rather than relying solely on the prose description.
  3. [Figure 2] Figure 2: the diagram of the overall pipeline is clear, but the legend for the Mamba block colors and the bidirectional arrows is missing, reducing readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment and recommendation for minor revision. The comments highlight important aspects of empirical robustness and component isolation that we will address in the revised manuscript.

read point-by-point responses

  1. Referee: [§4 (Experiments), Table 2 and Table 3] §4 (Experiments), Table 2 and Table 3: the reported improvements (e.g., 8.1% on SSv2-Full 1-shot) are given as single-point estimates without error bars, standard deviations across random seeds, or statistical significance tests; this leaves open whether the gains attributable to TSA and STPR exceed run-to-run variance and therefore weakens the central performance claim.

    Authors: We agree that single-point estimates limit the strength of the performance claims. In the revision we will rerun the primary 1-shot and 5-shot experiments on SSv2-Full, SSv2-Small and HMDB51 using five independent random seeds, reporting mean accuracy together with standard deviation for all methods in Tables 2 and 3. We will also include a brief note on whether the observed margins exceed the reported standard deviations. revision: yes

  2. Referee: [§3.2 (STPR description)] §3.2 (STPR description): the multi-frequency temporal sampling and bidirectional Mamba refinement are presented as key innovations, yet the ablation in Table 4 does not isolate the contribution of the frequency sampling schedule from the underlying Mamba blocks and LLM descriptors, making it impossible to attribute the reported prototype quality gains specifically to the proposed components.

    Authors: We acknowledge that the current Table 4 ablation does not disentangle the multi-frequency sampling schedule from the Mamba blocks and LLM descriptors. In the revision we will add two new rows to Table 4: (i) STPR without multi-frequency sampling (uniform sampling only) and (ii) STPR with frequency sampling but without bidirectional refinement. These controlled variants will allow readers to attribute performance differences more precisely to each sub-component. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper proposes an empirical engineering framework (STAR) with two new modules—TSA for frame-level cross-modal alignment and STPR for semantic-guided Mamba-based temporal refinement—plus LLM-derived descriptors. Claims rest on reported benchmark gains (e.g., +8.1% on SSv2-Full 1-shot) rather than any closed mathematical derivation. No equations, fitted parameters, or self-citations are shown that reduce the central results to inputs by construction. External components (Mamba, LLMs) and released code provide independent verification paths, so the work is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach relies on standard deep-learning assumptions about data distributions and pre-trained model transfer; no new free parameters or invented physical entities are introduced in the abstract.

axioms (1)
  • domain assumption Standard assumptions in supervised deep learning for video classification hold, including that training and test distributions share similar statistics.
    Implicit in nearly all few-shot vision papers and required for the reported generalization claims.

pith-pipeline@v0.9.0 · 5605 in / 1125 out tokens · 34152 ms · 2026-05-14T20:14:26.774284+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

72 extracted references · 72 canonical work pages · 4 internal anchors

  1. [1]

    Tsm: Temporal shift module for efficient video understanding,

    J. Lin, C. Gan, and S. Han, “Tsm: Temporal shift module for efficient video understanding,” in2019 IEEE/CVF International Conference on Computer Vision, 2019, pp. 7082–7092

    work page 2019

  2. [2]

    Learn- ing spatiotemporal features with 3d convolutional networks,

    D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri, “Learn- ing spatiotemporal features with 3d convolutional networks,” inIEEE International Conference on Computer Vision, 2015, pp. 4489–4497

    work page 2015

  3. [3]

    Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset

    J. Carreira and A. Zisserman, “Quo vadis, action recognition? A new model and the kinetics dataset,”CoRR, vol. abs/1705.07750, 2017. [Online]. Available: http://arxiv.org/abs/1705.07750

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  4. [4]

    Hu- man action recognition from various data modalities: A review,

    Z. Sun, Q. Ke, H. Rahmani, M. Bennamoun, G. Wang, and J. Liu, “Hu- man action recognition from various data modalities: A review,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 3, pp. 3200–3225, 2023

    work page 2023

  5. [5]

    Vivit: A video vision transformer,

    A. Arnab, M. Dehghani, G. Heigold, C. Sun, M. Lu ˇci´c, and C. Schmid, “Vivit: A video vision transformer,” inIEEE/CVF International Confer- ence on Computer Vision, 2021, pp. 6836–6846

    work page 2021

  6. [6]

    Compound memory networks for few-shot video classification,

    L. Zhu and Y . Yang, “Compound memory networks for few-shot video classification,” inEuropean conference on computer vision, 2018, pp. 751–766

    work page 2018

  7. [7]

    Few-shot video classification via temporal alignment,

    K. Cao, J. Ji, Z. Cao, C.-Y . Chang, and J. C. Niebles, “Few-shot video classification via temporal alignment,” in2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 10 615–10 624

    work page 2020

  8. [8]

    Hybrid relation guided set matching for few-shot action recognition,

    X. Wang, S. Zhang, Z. Qing, M. Tang, Z. Zuo, C. Gao, R. Jin, and N. Sang, “Hybrid relation guided set matching for few-shot action recognition,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 19 948–19 957

    work page 2022

  9. [9]

    Prototypical networks for few-shot learning,

    J. Snell, K. Swersky, and R. Zemel, “Prototypical networks for few-shot learning,”Advances in Neural Information Processing Systems, vol. 30, 2017

    work page 2017

  10. [10]

    Matching compound prototypes for few-shot action recognition,

    Y . Huang, L. Yang, G. Chen, H. Zhang, F. Lu, and Y . Sato, “Matching compound prototypes for few-shot action recognition,”International Journal of Computer Vision, vol. 132, no. 9, pp. 3977–4002, 2024

    work page 2024

  11. [11]

    Tarn: Temporal attentive relation network for few-shot and zero-shot action recognition,

    M. Bishay, G. Zoumpourlis, and I. Patras, “Tarn: Temporal attentive relation network for few-shot and zero-shot action recognition,” in British Machine Vision Conference, 2019, p. 154

    work page 2019

  12. [12]

    Temporal alignment- free video matching for few-shot action recognition,

    S. Lee, W. Moon, H. S. Seong, and J.-P. Heo, “Temporal alignment- free video matching for few-shot action recognition,”arXiv preprint arXiv:2504.05956, 2025

    work page arXiv 2025

  13. [13]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational Conference on Machine Learning. PmLR, 2021, pp. 8748–8763

    work page 2021

  14. [14]

    Clip-guided prototype modulating for few-shot action recognition,

    X. Wang, S. Zhang, J. Cen, C. Gao, Y . Zhang, D. Zhao, and N. Sang, “Clip-guided prototype modulating for few-shot action recognition,” International Journal of Computer Vision, vol. 132, no. 6, pp. 1899– 1912, 2024

    work page 1912

  15. [15]

    Temporal-attentive covariance pooling networks for video recognition,

    Z. Gao, Q. Wang, B. Zhang, Q. Hu, and P. Li, “Temporal-attentive covariance pooling networks for video recognition,”Advances in Neural Information Processing Systems, vol. 34, pp. 13 587–13 598, 2021

    work page 2021

  16. [16]

    Compound prototype matching for few- shot action recognition,

    Y . Huang, L. Yang, and Y . Sato, “Compound prototype matching for few- shot action recognition,” inEuropean Conference on Computer Vision, 2022, pp. 351–368

    work page 2022

  17. [17]

    Slowfocus: Enhancing fine-grained temporal understanding in video llm,

    M. Nie, D. Ding, C. Wang, Y . Guo, J. Han, H. Xu, and L. Zhang, “Slowfocus: Enhancing fine-grained temporal understanding in video llm,” inAdvances in Neural Information Processing Systems, 2024

    work page 2024

  18. [18]

    Mvp-shot: Multi-velocity progressive-alignment framework for few-shot action recognition,

    H. Qu, R. Yan, X. Shu, H. Gao, P. Huang, and G.-S. Xie, “Mvp-shot: Multi-velocity progressive-alignment framework for few-shot action recognition,”arXiv preprint arXiv:2405.02077, 2024

    work page arXiv 2024

  19. [19]

    Frame order matters: A temporal sequence-aware model for few-shot action recognition,

    B. Li, M. Liu, G. Wang, and Y . Yu, “Frame order matters: A temporal sequence-aware model for few-shot action recognition,”arXiv preprint arXiv:2408.12475, 2024

    work page arXiv 2024

  20. [20]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” inFirst Conference on Language Modeling, 2024

    work page 2024

  21. [21]

    Reallocating and evolving general knowledge for few-shot learning,

    Y . Su, X. Liu, Z. Huang, J. He, R. Hong, and M. Wang, “Reallocating and evolving general knowledge for few-shot learning,”IEEE Transac- tions on Circuits and Systems for Video Technology, 2024

    work page 2024

  22. [22]

    Few-shot cross-domain object detection with instance-level prototype-based meta-learning,

    L. Zhang, B. Zhang, B. Shi, J. Fan, and T. Chen, “Few-shot cross-domain object detection with instance-level prototype-based meta-learning,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 10, pp. 9078–9089, 2024

    work page 2024

  23. [23]

    Matching networks for one shot learning,

    O. Vinyals, C. Blundell, T. Lillicrap, K. Kavukcuoglu, and D. Wierstra, “Matching networks for one shot learning,” inAdvances in Neural Information Processing Systems, vol. 29, 2016, pp. 3637–3645

    work page 2016

  24. [24]

    Learning to compare: Relation network for few-shot learning,

    F. Sung, Y . Yang, L. Zhang, T. Xiang, P. H. Torr, and T. M. Hospedales, “Learning to compare: Relation network for few-shot learning,” inIEEE conference on computer vision and pattern recognition, 2018, pp. 1199– 1208

    work page 2018

  25. [25]

    Unify the views: View-consistent prototype learning for few-shot segmentation,

    H. Liu, Y . Wang, and S. Zhao, “Unify the views: View-consistent prototype learning for few-shot segmentation,” 2026. [Online]. Available: https://arxiv.org/abs/2603.05952

    work page arXiv 2026

  26. [26]

    Few-shot metric learning: Online adaptation of embedding for retrieval,

    D. Jung, D. Kang, S. Kwak, and M. Cho, “Few-shot metric learning: Online adaptation of embedding for retrieval,” inProceedings of the Asian Conference on Computer Vision, 2022, pp. 1875–1891

    work page 2022

  27. [27]

    Revisiting few-shot learning from a causal perspective,

    G. Lin, Y . Xu, H. Lai, and J. Yin, “Revisiting few-shot learning from a causal perspective,”IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 11, pp. 6908–6919, 2024

    work page 2024

  28. [28]

    Learning attention-guided pyrami- dal features for few-shot fine-grained recognition,

    H. Tang, C. Yuan, Z. Li, and J. Tang, “Learning attention-guided pyrami- dal features for few-shot fine-grained recognition,”Pattern Recognition, vol. 130, p. 108792, 2022

    work page 2022

  29. [29]

    Blockmix: meta regularization and self-calibrated inference for metric-based meta-learning,

    H. Tang, Z. Li, Z. Peng, and J. Tang, “Blockmix: meta regularization and self-calibrated inference for metric-based meta-learning,” inProceedings of the 28th ACM international conference on multimedia, 2020, pp. 610– 618

    work page 2020

  30. [30]

    Model-agnostic meta-learning for fast adaptation of deep networks,

    C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” inInternational conference on machine learning. PMLR, 2017, pp. 1126–1135

    work page 2017

  31. [31]

    Meta-exploiting frequency prior for cross-domain few-shot learning,

    F. Zhou, P. Wang, L. Zhang, Z. Chen, W. Wei, C. Ding, G. Lin, and Y . Zhang, “Meta-exploiting frequency prior for cross-domain few-shot learning,”Advances in Neural Information Processing Systems, vol. 37, pp. 116 783–116 814, 2024

    work page 2024

  32. [32]

    On the stability and generalization of meta- learning,

    Y . Wang and R. Arora, “On the stability and generalization of meta- learning,”Advances in Neural Information Processing Systems, vol. 37, pp. 83 665–83 710, 2024

    work page 2024

  33. [33]

    Few-shot action recognition via multi-view representation learning,

    X. Wang, Y . Lu, W. Yu, Y . Pang, and H. Wang, “Few-shot action recognition via multi-view representation learning,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 9, pp. 8522– 8535, 2024

    work page 2024

  34. [34]

    Task- adapter: Task-specific adaptation of image models for few-shot action recognition,

    C. Cao, Y . Zhang, Y . Yu, Q. Lv, L. Min, and Y . Zhang, “Task- adapter: Task-specific adaptation of image models for few-shot action recognition,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 9038–9047

    work page 2024

  35. [35]

    Cross-modal contrastive pre-training for few-shot skeleton action recognition,

    M. Lu, S. Yang, X. Lu, and J. Liu, “Cross-modal contrastive pre-training for few-shot skeleton action recognition,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 10, pp. 9798–9807, 2024

    work page 2024

  36. [36]

    Dual-recommendation disentanglement network for view fuzz in action recognition,

    W. Liu, X. Zhong, Z. Zhou, K. Jiang, Z. Wang, and C.-W. Lin, “Dual-recommendation disentanglement network for view fuzz in action recognition,”IEEE Transactions on Image Processing, vol. 32, pp. 2719– 2733, 2023

    work page 2023

  37. [37]

    Motion- consistent representation learning for uav-based action recognition,

    W. Liu, X. Zhong, Y . Dai, X. Jia, Z. Wang, and S. Satoh, “Motion- consistent representation learning for uav-based action recognition,” IEEE Transactions on Intelligent Transportation Systems, 2025

    work page 2025

  38. [38]

    Depth guided adaptive meta-fusion network for few-shot video recognition,

    Y . Fu, L. Zhang, J. Wang, Y . Fu, and Y .-G. Jiang, “Depth guided adaptive meta-fusion network for few-shot video recognition,” inACM International Conference on Multimedia, 2020, p. 1142–1151

    work page 2020

  39. [39]

    Temporal-relational crosstransformers for few-shot action recognition,

    T. Perrett, A. Masullo, T. Burghardt, M. Mirmehdi, and D. Damen, “Temporal-relational crosstransformers for few-shot action recognition,” inComputer Vision and Pattern Recognition, 2021, pp. 475–484

    work page 2021

  40. [40]

    Semantic-guided relation propagation network for few-shot action recognition,

    X. Wang, W. Ye, Z. Qi, X. Zhao, G. Wang, Y . Shan, and H. Wang, “Semantic-guided relation propagation network for few-shot action recognition,” inACM International Conference on Multimedia, 2021, pp. 816–825

    work page 2021

  41. [41]

    Molo: Motion-augmented long-short contrastive learning for few-shot action recognition,

    X. Wang, S. Zhang, Z. Qing, C. Gao, Y . Zhang, D. Zhao, and N. Sang, “Molo: Motion-augmented long-short contrastive learning for few-shot action recognition,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 18 011–18 021

    work page 2023

  42. [42]

    M3net: multi- view encoding, matching, and fusion for few-shot fine-grained action recognition,

    H. Tang, J. Liu, S. Yan, R. Yan, Z. Li, and J. Tang, “M3net: multi- view encoding, matching, and fusion for few-shot fine-grained action recognition,” inProceedings of the 31st ACM international conference on multimedia, 2023, pp. 1719–1728

    work page 2023

  43. [43]

    Hierarchy- aware interactive prompt learning for few-shot classification,

    X. Yin, J. Wu, W. Yang, X. Zhou, S. Zhang, and T. Zhang, “Hierarchy- aware interactive prompt learning for few-shot classification,”IEEE Transactions on Circuits and Systems for Video Technology, 2024

    work page 2024

  44. [44]

    Rethinking few-shot adaptation of vision-language models in two stages,

    M. Farina, M. Mancini, G. Iacca, and E. Ricci, “Rethinking few-shot adaptation of vision-language models in two stages,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 29 989–29 998

    work page 2025

  45. [45]

    Connecting giants: synergistic knowledge transfer of large multimodal models for few-shot learning,

    H. Tang, S. He, and J. Qin, “Connecting giants: synergistic knowledge transfer of large multimodal models for few-shot learning,” inProceed- ings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, ser. IJCAI ’25, 2025, pp. 6227–6235. IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY ,, VOL. 14, NO. 8, AUGUST 2021 14

    work page 2025

  46. [46]

    Cross-modal proxy evolving for ood detection with vision-language models,

    H. Tang, Y . Liu, S. Yan, F. Shen, S. He, and J. Qin, “Cross-modal proxy evolving for ood detection with vision-language models,”Proceedings of the AAAI Conference on Artificial Intelligence, no. 18, pp. 15 770– 15 778, 2026

    work page 2026

  47. [47]

    Multi-speed global contextual subspace matching for few-shot action recognition,

    T. Yu, P. Chen, Y . Dang, R. Huan, and R. Liang, “Multi-speed global contextual subspace matching for few-shot action recognition,” inACM International Conference on Multimedia, 2023, pp. 2344–2352

    work page 2023

  48. [48]

    On the importance of spatial relations for few-shot action recognition,

    Y . Zhang, Y . Fu, X. Ma, L. Qi, J. Chen, Z. Wu, and Y .-G. Jiang, “On the importance of spatial relations for few-shot action recognition,” in ACM International Conference on Multimedia, 2023, pp. 2243–2251

    work page 2023

  49. [49]

    Combining recurrent, convolutional, and continuous-time models with linear state space layers,

    A. Gu, I. Johnson, K. Goel, K. Saab, T. Dao, A. Rudra, and C. R ´e, “Combining recurrent, convolutional, and continuous-time models with linear state space layers,”Advances in Neural Information Processing Systems, vol. 34, pp. 572–585, 2021

    work page 2021

  50. [50]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, and C. Re, “Efficiently modeling long sequences with structured state spaces,” inInternational Conference on Learning Representations, 2022

    work page 2022

  51. [51]

    Selective structured state-spaces for long-form video understanding,

    J. Wang, W. Zhu, P. Wang, X. Yu, L. Liu, M. Omar, and R. Hamid, “Selective structured state-spaces for long-form video understanding,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 6387–6397

    work page 2023

  52. [52]

    Transformers are ssms: generalized models and effi- cient algorithms through structured state space duality,

    T. Dao and A. Gu, “Transformers are ssms: generalized models and effi- cient algorithms through structured state space duality,” inInternational Conference on Machine Learning, 2024, pp. 10 041–10 071

    work page 2024

  53. [53]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” inProceedings of the 31st International Conference on Neural Information Processing Systems, 2017, pp. 6000–6010

    work page 2017

  54. [54]

    Mambaout: Do we really need mamba for vision?

    W. Yu and X. Wang, “Mambaout: Do we really need mamba for vision?” arXiv preprint arXiv:2405.07992, 2024

    work page arXiv 2024

  55. [55]

    Vision mamba: efficient visual representation learning with bidirectional state space model,

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang, “Vision mamba: efficient visual representation learning with bidirectional state space model,” inInternational Conference on Machine Learning, 2024, pp. 62 429–62 442

    work page 2024

  56. [56]

    Vmamba: Visual state space model,

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, J. Jiao, and Y . Liu, “Vmamba: Visual state space model,”Advances in Neural Information Processing Systems, vol. 37, pp. 103 031–103 063, 2024

    work page 2024

  57. [57]

    Videomamba: State space model for efficient video understanding,

    K. Li, X. Li, Y . Wang, Y . He, Y . Wang, L. Wang, and Y . Qiao, “Videomamba: State space model for efficient video understanding,” in European Conference on Computer Vision, 2024, pp. 237–255

    work page 2024

  58. [58]

    Mamba-adaptor: State space model adaptor for visual recognition,

    F. Xie, J. Nie, Y . Tang, W. Zhang, and H. Zhao, “Mamba-adaptor: State space model adaptor for visual recognition,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 20 124–20 134

    work page 2025

  59. [59]

    Mambavision: A hybrid mamba- transformer vision backbone,

    A. Hatamizadeh and J. Kautz, “Mambavision: A hybrid mamba- transformer vision backbone,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 25 261–25 270

    work page 2025

  60. [60]

    Groupmamba: Efficient group-based visual state space model,

    A. Shaker, S. T. Wasim, S. Khan, J. Gall, and F. S. Khan, “Groupmamba: Efficient group-based visual state space model,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 14 912– 14 922

    work page 2025

  61. [61]

    Spatial channel attention for deep convolutional neural networks,

    T. Liu, R. Luo, L. Xu, D. Feng, L. Cao, S. Liu, and J. Guo, “Spatial channel attention for deep convolutional neural networks,”Mathematics, vol. 10, no. 10, p. 1750, 2022

    work page 2022

  62. [62]

    Hmdb: A large video database for human motion recognition,

    H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre, “Hmdb: A large video database for human motion recognition,” in2011 Interna- tional Conference on Computer Vision, 2011, pp. 2556–2563

    work page 2011

  63. [63]

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

    K. Soomro, A. R. Zamir, and M. Shah, “Ucf101: A dataset of 101 human actions classes from videos in the wild,”arXiv preprint arXiv:1212.0402, 2012

    work page internal anchor Pith review Pith/arXiv arXiv 2012

  64. [64]

    Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset ,

    J. Carreira and A. Zisserman, “ Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset ,” inComputer Vision and Pattern Recognition, 2017, pp. 4724–4733

    work page 2017

  65. [65]

    The” something something

    R. Goyal, S. Ebrahimi Kahou, V . Michalski, J. Materzynska, S. West- phal, H. Kim, V . Haenel, I. Fruend, P. Yianilos, M. Mueller-Freitaget al., “The” something something” video database for learning and evaluating visual common sense,” inIEEE International Conference on Computer Vision, 2017, pp. 5842–5850

    work page 2017

  66. [66]

    Few-shot action recognition with permutation-invariant attention,

    H. Zhang, L. Zhang, X. Qi, H. Li, P. H. Torr, and P. Koniusz, “Few-shot action recognition with permutation-invariant attention,” inEuropean Conference on Computer Vision, 2020, pp. 525–542

    work page 2020

  67. [67]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inComputer Vision and Pattern Recognition, 2016, pp. 770–778

    work page 2016

  68. [68]

    An image is worth 16x16 words: Trans- formers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Trans- formers for image recognition at scale,” inInternational Conference on Learning Representations, 2021

    work page 2021

  69. [69]

    GPT-4 Technical Report

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  70. [70]

    Adam: A Method for Stochastic Optimization

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  71. [71]

    Hello gpt-4o,

    OpenAI, “Hello gpt-4o,” https://openai.com/index/hello-gpt-4o, 2024, accessed: 2025-08-05

    work page 2024

  72. [72]

    Gemini 2.5: Our most intelligent ai model,

    Google DeepMind, “Gemini 2.5: Our most intelligent ai model,” https://blog.google/technology/google-deepmind/ gemini-model-thinking-updates-march-2025/, 2025, accessed: 2025- 08-05. Hongli LiuHongli Liu is currently pursuing the Ph.D. degree with the School of Computer Science, Tongji University, Shanghai, China. Her main re- search interests include comp...

    work page 2025