Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Technical Report for Ego4D Long-Term Action Anticipation Challenge 2025

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims a three-stage framework with hand-object cues and a verb-noun co-occurrence matrix wins the Ego4D 2025 long-term action anticipation challenge with a 0.8493 action edit distance.

desk verdict A plausible first-place leaderboard result with an unsupported story: the two new components are never ablated, and the recognition accuracy in Table 2 is actually worse than the baseline. read the letter →

arxiv 2506.02550 v2 pith:MD2ULJV5 submitted 2025-06-03 cs.CV cs.AI

Qiaohui Chu , Haoyu Zhang , Yisen Feng , Meng Liu , Weili Guan , Yaowei Wang , Liqiang Nie This is my paper
classification cs.CVcs.AI
keywords Ego4Dlong-termactionanticipationegocentricvideounderstandingverb-nounpredictionco-occurrencematrixhand-objectinteractionlargelanguagemodelseditdistance
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is trying to establish that a three-stage framework for egocentric long-term action anticipation—feature extraction, action recognition, and large-language-model-based future-action prediction—achieves first place in the 2025 Ego4D long-term action anticipation challenge. The framework adds two innovations to the recognition stage: hand-object interaction cues from a segmentation model, and a verb-noun co-occurrence matrix that reweights the model's candidate predictions. On the Ego4D v2 test leaderboard, the method reaches an action-level edit distance of 0.8493, narrowly ahead of the runner-up at 0.8504. If correct, this shows that improving the intermediate action recognition stage is a direct route to better long-term anticipation, rather than only scaling the language model.

What carries the argument

The load-bearing piece is the verb-noun co-occurrence matrix, a normalized table of how often verb-noun pairs appear together in the training data; the recognizer multiplies the Transformer's top-five verb and noun probability distributions by this matrix to form the most plausible verb-noun pairs. A second supporting piece is the hand-object interaction branch: a segmentation model isolates hand-object regions, whose features are encoded separately and fused with whole-frame features by a lightweight MLP. These enriched features and reweighted predictions feed the anticipation stage, where a fine-tuned Llama2-7B receives the recognized actions as a textual prompt and generates the future 20 actions.

What would settle it

Run the identical pipeline on the Ego4D v2 test set with the hand-object branch removed and the co-occurrence matrix replaced by a uniform one; if the action edit distance stays at or near 0.8493, the claim that these additions are responsible for the gain is falsified.

Watch

Extended reading notes

Core claim

The central claim is that the proposed framework—the EgoVideo-V visual encoder, a SAM2-derived hand-object feature stream fused via an MLP, a Transformer recognizer that uses a verb-noun co-occurrence matrix to select top-five verb-noun pairs, and a fine-tuned Llama2-7B that turns those pairs into a 20-step future action sequence—produces the best reported action anticipation results on the Ego4D v2 test set. The paper reports verb, noun, and action edit distances of 0.6346, 0.5986, and 0.8493 respectively, ranking first on the public leaderboard. The authors argue that the co-occurrence matrix and hand-object cues improve recognition accuracy, and that this improved recognition propagates to the LLM stage, so the whole pipeline outperforms both the PALM and AntGPT baselines and the runner-up EgoVideo-based system.

Load-bearing premise

The paper never tests whether the two proposed additions—hand-object cues and the co-occurrence matrix—actually contribute anything, because its ablations only swap the visual encoder and the LLM, so the claim that these components drive the first-place result rests on an untested assumption.

Editorial extensions

If this is right

  • The best LTA performance on Ego4D v2 now comes from combining an egocentric visual encoder with an explicit hand-object prior and a verb-noun co-occurrence prior, not from the language model alone.
  • Fine-tuned Llama2-7B can produce a plausible 20-action future sequence from a textual history of recognized verb-noun pairs, and its output quality is bounded by the quality of the recognition stage.
  • Higher action recognition accuracy alone does not guarantee better anticipation: the paper's method has lower recognition accuracy than the fine-tuned EgoVideo baseline yet lower (better) action edit distance.
  • The co-occurrence matrix approach transfers from the QueryMamba system to a Transformer recognizer and improves verb-noun pair selection.

Reading between the lines

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

  • Because no ablation removes the hand-object branch or the co-occurrence matrix, the leaderboard gain could partly come from the choice of EgoVideo-V or the Llama2 fine-tuning recipe; an ablation isolating these two components would settle that.
  • The verb-noun co-occurrence matrix is a cheap, model-agnostic prior that could be dropped into other egocentric verb-noun recognition or anticipation systems regardless of the LLM stage.
  • One testable extension is to replace the SAM2 hand-object features with other segmentation or hand-detection cues to see whether the gain is specific to that segmentation model or generic to hand-object grounding.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. This manuscript is a challenge technical report describing a three-stage framework for Ego4D Long-Term Action Anticipation: (1) feature extraction with EgoVideo-V and SAM2-based hand-object interaction features, (2) action recognition with a Transformer plus a verb-noun co-occurrence matrix, and (3) long-term anticipation with a fine-tuned Llama2-7B that converts recognized verb-noun pairs into textual prompts. The paper reports first place on the Ego4D v2 LTA leaderboard at CVPR 2025, with an action Edit Distance of 0.8493 (verb ED 0.6346, noun ED 0.5986), and compares against PALM, AntGPT, EgoVideo, and six leaderboard teams. Section 3.2 provides an ablation across visual encoders and LLMs, and Section 3.3 gives qualitative case studies. The central claim is the first-place rank and its attribution to the proposed components, but the experimental evidence for that attribution is incomplete.

Significance. If the first-place result and its attribution are valid, the paper would demonstrate that a recognition-stage enhanced with hand-object cues and co-occurrence reranking, followed by a fine-tuned LLM, can achieve state-of-the-art long-term action anticipation on the Ego4D benchmark. The strength of the paper is that the final score is an externally validated leaderboard result, and the comparison against PALM and AntGPT is useful. However, the proposed components are never isolated, the recognition-stage accuracy reported for the main configuration is lower than the EgoVideo baseline while the LTA score is better, and no variance or significance analysis accompanies a 0.0011 ED margin. These issues mean the method's contribution is currently not established, even if the reported rank is correct.

major comments (3)
  1. [Section 3.2 / Table 2] The central attribution of the first-place result is unsupported. Table 2 is explicitly described as ablating visual encoders and inference LLMs, and neither the SAM2 hand-object fusion (Section 2.1) nor the verb-noun co-occurrence reranker (Section 2.2) is removed in any row. All "Ours" rows contain both components, so the comparison between any baseline and "Ours" changes the recognition stage, the LLM, and the two proposed additions simultaneously. Without at least one configuration that disables each proposed component separately and jointly, the statement in Section 3.2 that "combining improvements across all three stages" yields the result does not follow from the data; the first-place ranking could hold even if both additions contribute nothing. Please add ablations that drop the SAM2 fusion and the co-occurrence matrix.
  2. [Section 3.2 / Table 2] The recognition-accuracy discussion is contradicted by the table. The "Ours (EgoVideo-V)" row reports 22.15% action recognition accuracy, which is 5.49 points below the fine-tuned EgoVideo-V baseline's 27.64% and only 1.52 points above the EgoVLP sliding-window baseline of 20.63%. Calling this "slightly lower" is misleading, and the sentence that "the advanced EgoVideo-V achieves the highest action recognition accuracy, reaching 22.15%" is only true among the two Ours rows, not among all methods in the table. Since the LTA improvement (0.8504 to 0.8493) occurs despite a worse recognition stage, the gain cannot be attributed to improved recognition accuracy as Section 1 and Section 3.2 claim; the improvement must come from the LLM, the reranker, or benchmark variation. The text needs to be corrected and the cause identified through controlled experiments.
  3. [Section 3.1 / Table 1] The margin over the runner-up is 0.0011 in action Edit Distance (0.8493 vs. 0.8504), and the paper reports no variance estimate, confidence interval, or multiple-run analysis. The first-place rank is an externally reported leaderboard fact, but the claim that the specific proposed configuration is superior by a stable margin is not supported by any statistical evidence. If the leaderboard is based on a single fixed test-set evaluation, state this explicitly; otherwise, provide repeated-run statistics or otherwise justify that a 0.0011 difference is meaningful rather than noise.
minor comments (4)
  1. [Figure 1 caption] The caption describes the pipeline as a "two-stage pipeline," but the abstract, introduction, and methodology all describe a three-stage framework; the caption should be corrected.
  2. [Table 1] Table 1 has formatting errors that make it hard to read: the "BigMac (mtp)" row renders as "30.63400.6395 0.8649" without separators, and the team label "Ours (iLearn2.O)" is inconsistent with the naming used elsewhere. Please fix the table formatting and unify the team/method names.
  3. [Section 2.2] The verb-noun co-occurrence matrix is central to the method, but the paper does not describe how the matrix is estimated, normalized, or integrated with the Transformer probabilities; adding a formal description or a small equation would substantially improve reproducibility.
  4. [Abstract / Section 3.2] The abstract promises that code will be released, but no code, checkpoints, or detailed hyperparameters are included in the report; please clarify the intended release date and provide enough configuration detail for the experimental claims to be reproduced.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the central result is an external leaderboard outcome, and the unablated component claims are a support issue, not a circular reduction.

full rationale

This paper is a challenge technical report describing a three-stage pipeline for Ego4D LTA. Its central claim, first place with LTA Edit Distance 0.8493, is validated by an external public leaderboard, not by any derivation from the method's own assumptions. The method builds on prior external work (PALM, AntGPT, EgoVideo-V, SAM2, QueryMamba) rather than on self-citations that carry the argument. The intro cites several works by the same group, but none of those citations supplies a load-bearing premise or forbids alternatives, so they do not constitute circularity. The main weakness is that Table 2 does not ablate the SAM2 hand-object fusion or the co-occurrence reranker, and Section 3.2 attributes gains to these components despite the recognition accuracy being lower than the EgoVideo-V baseline. However, missing ablations and an unsupported attribution are concerns about evidence quality, not about a derivation being equivalent to its inputs. No equation is fitted to the predicted quantity, no fitted parameter is renamed as a prediction, and no uniqueness theorem is imported from the authors' own prior work. Therefore, under the stated rules, there is no specific circular step to report, and the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The framework leans on several unstated domain assumptions: pretrained encoders transfer to Ego4D, co-occurrence statistics from training data transfer to the test distribution, and an LLM fine-tuned on observed action text can produce plausible future sequences. None of these are proven in the report. The main fitted object is the verb-noun co-occurrence matrix, which is estimated from data but not described or ablated.

free parameters (3)
  • verb-noun co-occurrence matrix = estimated from Ego4D training data (details not reported)
    Used in Section 2.2 to rerank top-5 verb and noun predictions; its construction, normalization, and standalone contribution are not given.
  • top-k candidate count = 5
    The recognition stage selects the top five verbs and nouns before reranking; the choice is fixed and its sensitivity is not tested.
  • frames per segment = 4
    Adopted from AntGPT; the number of sampled frames is fixed and not ablated.
assumptions (4)
  • domain assumption EgoVideo-V visual features contain sufficient information for verb and noun recognition.
    The recognition stage in Section 2.1 is built on this pretrained encoder without a task-specific analysis.
  • domain assumption Verb-noun co-occurrence statistics from training data generalize to the test distribution.
    Section 2.2 uses them as a hard re-ranking prior; distribution shift could make the prior harmful.
  • domain assumption Fine-tuned Llama 2 can generate coherent chronological action sequences from observed action text.
    Section 2.3 assumes this capability from prior work, but fine-tuning details are omitted.
  • domain assumption SAM2 accurately localizes hand-object interactions relevant to future actions.
    Section 2.1 adds hand-object region features without an ablation showing they help.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Technical Report for Ego4D Long-Term Action Anticipation Challenge 2025." pith.science (2026). https://pith.science/paper/MD2ULJV5

@misc{pith2026250602550,
  author       = {Pith},
  title        = {Pith review of: Technical Report for Ego4D Long-Term Action Anticipation Challenge 2025},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MD2ULJV5}},
  note         = {Machine review of arXiv:2506.02550}
}
read the original abstract

In this report, we present a novel three-stage framework developed for the Ego4D Long-Term Action Anticipation (LTA) task. Inspired by recent advances in foundation models, our method consists of three stages: feature extraction, action recognition, and long-term action anticipation. First, visual features are extracted using a high-performance visual encoder. The features are then fed into a Transformer to predict verbs and nouns, with a verb-noun co-occurrence matrix incorporated to enhance recognition accuracy. Finally, the predicted verb-noun pairs are formatted as textual prompts and input into a fine-tuned large language model (LLM) to anticipate future action sequences. Our framework achieves first place in this challenge at CVPR 2025, establishing a new state-of-the-art in long-term action prediction. Our code will be released at https://github.com/CorrineQiu/Ego4D-LTA-Challenge-2025.

Figures

Figures reproduced from arXiv: 2506.02550 by the authors.

Figure 1
Figure 1. An illustration of our two-stage pipeline. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Successful and failed cases of Action Recognition [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An example from our action anticipation model. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. OSGNet @ Ego4D Episodic Memory Challenge 2025

    cs.CV 2025-06 conditional novelty 4.0 of 10

    OSGNet, an early-fusion grounding model, wins all three Ego4D Episodic Memory Challenge tracks by converting localization tasks into retrieval problems.

Reference graph

Works this paper leans on

24 extracted references · 4 canonical work pages · cited by 1 Pith paper

  1. [1]

    Videollm: Modeling video sequence with large language models.arXiv preprint arXiv:2305.13292, 2023

    Guo Chen, Yin-Dong Zheng, Jiahao Wang, Jilan Xu, Yifei Huang, Junting Pan, Yi Wang, Yali Wang, Yu Qiao, Tong Lu, et al. Videollm: Modeling video sequence with large language models.arXiv preprint arXiv:2305.13292, 2023. 3

    arXiv 2023

  2. [2]

    Objectnlq@ ego4d episodic memory challenge 2024.arXiv preprint arXiv:2406.15778, 2024

    Yisen Feng, Haoyu Zhang, Yuquan Xie, Zaijing Li, Meng Liu, and Liqiang Nie. Objectnlq@ ego4d episodic memory challenge 2024.arXiv preprint arXiv:2406.15778, 2024. 1

    arXiv 2024

  3. [3]

    Object-shot enhanced grounding network for egocentric video.arXiv preprint arXiv:2505.04270, 2025

    Yisen Feng, Haoyu Zhang, Meng Liu, Weili Guan, and Liqiang Nie. Object-shot enhanced grounding network for egocentric video.arXiv preprint arXiv:2505.04270, 2025. 1

    arXiv 2025

  4. [4]

    Ego4d: Around the world in 3,000 hours of egocentric video

    Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 18995–19012, 2022. 1

    2022

  5. [5]

    Bi-directional het- erogeneous graph hashing towards efficient outfit recom- mendation

    Weili Guan, Xuemeng Song, Haoyu Zhang, Meng Liu, Chung-Hsing Yeh, and Xiaojun Chang. Bi-directional het- erogeneous graph hashing towards efficient outfit recom- mendation. InProceedings of the 30th ACM international conference on multimedia, pages 268–276, 2022. 1

    2022

  6. [6]

    Palm: Predicting actions through language models@ ego4d long-term action anticipation challenge 2023.arXiv preprint arXiv:2306.16545, 2023

    Daoji Huang, Otmar Hilliges, Luc Van Gool, and Xi Wang. Palm: Predicting actions through language models@ ego4d long-term action anticipation challenge 2023.arXiv preprint arXiv:2306.16545, 2023. 1, 3

    arXiv 2023

  7. [7]

    Palm: Predicting actions through language models

    Sanghwan Kim, Daoji Huang, Yongqin Xian, Otmar Hilliges, Luc Van Gool, and Xi Wang. Palm: Predicting actions through language models. InEuropean Conference on Computer Vision, pages 140–158. Springer, 2024. 3, 4

    work page 2024

  8. [8]

    Egocentric video-language pretraining.Advances in Neural Information Processing Sys- tems, 35:7575–7586, 2022

    Kevin Qinghong Lin, Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Z Xu, Difei Gao, Rong-Cheng Tu, Wen- zhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining.Advances in Neural Information Processing Sys- tems, 35:7575–7586, 2022. 2, 4

    work page 2022

Show all 24 references
  1. [9]

    Attentive moment retrieval in videos

    Meng Liu, Xiang Wang, Liqiang Nie, Xiangnan He, Bao- quan Chen, and Tat-Seng Chua. Attentive moment retrieval in videos. InThe 41st international ACM SIGIR conference on research & development in information retrieval, pages 15–24, 2018. 1

    2018

  2. [10]

    Cross-modal moment localiza- tion in videos

    Meng Liu, Xiang Wang, Liqiang Nie, Qi Tian, Baoquan Chen, and Tat-Seng Chua. Cross-modal moment localiza- tion in videos. InProceedings of the 26th ACM international conference on Multimedia, pages 843–851, 2018. 1

    2018

  3. [11]

    Egovideo: Exploring egocentric founda- tion model and downstream adaptation.arXiv preprint arXiv:2406.18070, 2024

    Baoqi Pei, Guo Chen, Jilan Xu, Yuping He, Yicheng Liu, Kanghua Pan, Yifei Huang, Yali Wang, Tong Lu, Limin Wang, et al. Egovideo: Exploring egocentric founda- tion model and downstream adaptation.arXiv preprint arXiv:2406.18070, 2024. 2, 3, 4

    2024 arXiv

  4. [12]

    Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024. 2

    2024 arXiv

  5. [13]

    Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023. 1, 4

    2023 arXiv

  6. [14]

    Time: Temporal-sensitive multi-dimensional instruc- tion tuning and benchmarking for video-llms.arXiv preprint arXiv:2503.09994, 2025

    Yunxiao Wang, Meng Liu, Rui Shao, Haoyu Zhang, Bin Wen, Fan Yang, Tingting Gao, Di Zhang, and Liqiang Nie. Time: Temporal-sensitive multi-dimensional instruc- tion tuning and benchmarking for video-llms.arXiv preprint arXiv:2503.09994, 2025. 1

    2025 arXiv

  7. [15]

    Multimodal dialog system: Rela- tional graph-based context-aware question understanding

    Haoyu Zhang, Meng Liu, Zan Gao, Xiaoqiang Lei, Yinglong Wang, and Liqiang Nie. Multimodal dialog system: Rela- tional graph-based context-aware question understanding. In Proceedings of the 29th ACM international conference on multimedia, pages 695–703, 2021. 1

    2021

  8. [16]

    Attribute-guided collab- orative learning for partial person re-identification.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):14144–14160, 2023

    Haoyu Zhang, Meng Liu, Yuhong Li, Ming Yan, Zan Gao, Xiaojun Chang, and Liqiang Nie. Attribute-guided collab- orative learning for partial person re-identification.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):14144–14160, 2023. 1

    2023

  9. [17]

    Uncovering hidden connections: Iterative tracking and reasoning for video-grounded dialog

    Haoyu Zhang, Meng Liu, Yaowei Wang, Da Cao, Weili Guan, and Liqiang Nie. Uncovering hidden connections: Iterative tracking and reasoning for video-grounded dialog. arXiv preprint arXiv:2310.07259, 2023. 1

    2023 arXiv

  10. [18]

    Multi-factor adaptive vision selec- tion for egocentric video question answering

    Haoyu Zhang, Meng Liu, Zixin Liu, Xuemeng Song, Yaowei Wang, and Liqiang Nie. Multi-factor adaptive vision selec- tion for egocentric video question answering. InProceedings of the 41st International Conference on Machine Learning, pages 59310–59328. PMLR, 2024. 1

    2024

  11. [19]

    Hcqa@ ego4d egoschema challenge 2024.arXiv preprint arXiv:2406.15771, 2024

    Haoyu Zhang, Yuquan Xie, Yisen Feng, Zaijing Li, Meng Liu, and Liqiang Nie. Hcqa@ ego4d egoschema challenge 2024.arXiv preprint arXiv:2406.15771, 2024. 1

    2024 arXiv

  12. [20]

    Exo2ego: Exocentric knowledge guided mllm for egocentric video understanding.arXiv preprint arXiv:2503.09143, 2025

    Haoyu Zhang, Qiaohui Chu, Meng Liu, Yunxiao Wang, Bin Wen, Fan Yang, Tingting Gao, Di Zhang, Yaowei Wang, and Liqiang Nie. Exo2ego: Exocentric knowledge guided mllm for egocentric video understanding.arXiv preprint arXiv:2503.09143, 2025. 1

    2025

  13. [21]

    Hcqa- 1.5@ ego4d egoschema challenge 2025.arXiv preprint arXiv:2505.20644, 2025

    Haoyu Zhang, Yisen Feng, Qiaohui Chu, Meng Liu, Weili Guan, Yaowei Wang, and Liqiang Nie. Hcqa- 1.5@ ego4d egoschema challenge 2025.arXiv preprint arXiv:2505.20644, 2025. 1

    2025 arXiv

  14. [22]

    Antgpt: Can large language models help long- term action anticipation from videos?arXiv preprint arXiv:2307.16368, 2023

    Qi Zhao, Shijie Wang, Ce Zhang, Changcheng Fu, Minh Quan Do, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. Antgpt: Can large language models help long- term action anticipation from videos?arXiv preprint arXiv:2307.16368, 2023. 2, 3

    2023 arXiv

  15. [23]

    Judging llm-as-a-judge with mt-bench and chatbot arena.Advances in Neural Information Processing Systems, 36:46595–46623, 2023

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena.Advances in Neural Information Processing Systems, 36:46595–46623, 2023. 2, 4

    2023

  16. [24]

    Zeyun Zhong, Manuel Martin, Frederik Diederichs, and Juergen Beyerer. Querymamba: A mamba-based encoder- decoder architecture with a statistical verb-noun interac- tion module for video action forecasting@ ego4d long- term action anticipation challenge 2024.arXiv preprint arXi...

    2024 arXiv

Pith tools

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