Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Flash-VStream: Efficient Real-Time Understanding for Long Video Streams

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A 7B video-language model can answer hour-long video questions in under a second using a fixed 11,520-token memory.

desk verdict A solid memory architecture for real-time long-video QA with honest ablations, but the headline comparison against Qwen2-VL-online is confounded by instruction tuning, so the SOTA claims over-reach. read the letter →

arxiv 2506.23825 v2 pith:C5VS42FP submitted 2025-06-30 cs.CV

Haoji Zhang , Yiqin Wang , Yansong Tang , Yong Liu , Jiashi Feng , Xiaojie Jin This is my paper
classification cs.CV
keywords longvideounderstandingreal-timequestionansweringlanguagemodelmemorycompressionstreamingK-meansclusteringefficientinferenceEgoSchema
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

Flash-VStream is a video-language model built to answer questions about streams that can run for an hour, with the first token of the answer produced in under one second. The paper's central claim is that a fixed memory of 11,520 compressed visual tokens is enough: Flash-VStream reports 68.2 on the full EgoSchema benchmark, 4.2 points above the equal-cost Qwen2-VL-online baseline, and matches or beats much larger token budgets on four other video benchmarks. This is achieved by splitting the pipeline into two asynchronous processes, one that continuously compresses frames into a cluster-based memory and one that answers queries from that fixed memory. If the claim holds, long-video question answering becomes a live, interactive service rather than a batch job, at a fraction of the compute and GPU memory.

What carries the argument

The Flash Memory module is a fixed-size two-part memory of clustered frame features. CSM performs online K-means clustering over pooled low-resolution frame features, keeping $N_{\mathrm{CSM}}=60$ centroids of size $64$ tokens each; DAM then selects $N_{\mathrm{DAM}}=30$ high-resolution feature maps of size $256$ tokens each by nearest-Euclidean-distance retrieval against the largest CSM clusters, using a feature bank that can be offloaded to disk. The interleaved $3840+7680=11520$ tokens are positioned with AM-RoPE, which replaces a single position index with a triplet for time, height, and width. The module's job is to concentrate the LLM's attention budget on temporally informative content, so that the question handler only ever sees a fixed-size context.

What would settle it

Run Flash-VStream and uniform frame sampling at the same 11,520-token budget on a benchmark of heavily edited, text-dense videos; if uniform sampling matches or beats Flash-VStream there, the temporal-redundancy premise fails for those stream types.

Watch

Extended reading notes

Core claim

Flash-VStream's core claim is that long videos can be reduced, online, to a fixed-size 'Flash Memory' without losing the information needed to answer questions. The memory has two parts: a Context Synopsis Memory (CSM), which keeps $N_{\mathrm{CSM}}$ cluster centroids of low-resolution frame features, with the cluster sizes acting as a measure of information density along time; and a Detail Augmentation Memory (DAM), which stores high-resolution feature maps of the $N_{\mathrm{DAM}}$ frames nearest to the largest clusters. The two parts are interleaved in chronological order and fed to the language model with an Adaptive Multimodal RoPE that assigns each token a temporal position. With this design, the paper reports state-of-the-art accuracy on the full EgoSchema benchmark and, at an equal 11,520-token cost, steady gains over the online baseline on short-video and long-video benchmarks alike.

Load-bearing premise

The claim rests on the premise that natural video is mostly temporally redundant, so a fixed-size memory of cluster centroids plus a few key-frame features keeps everything needed to answer any question; the paper's own limitation section concedes this fails for text-heavy and rapidly edited videos.

Editorial extensions

If this is right

  • A video question can be answered during the stream, not only at the end, because the memory is updated continuously and the question handler reads it at any moment.
  • GPU memory and inference latency stop growing with video length: the frame handler encodes and clusters frames incrementally, while the LLM always sees the same 11,520-token context.
  • Equal-cost comparisons become meaningful: at the same token budget, Flash-VStream beats uniform sampling and the online baseline, so compression plus targeted retrieval is a better use of a fixed budget than naive subsampling.
  • The state-of-the-art EgoSchema result indicates that first-person, hour-long reasoning tasks are within reach of real-time deployment, which is a precondition for wearable assistants and robotics.

Reading between the lines

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

  • A natural extension, not tested in the paper, would be to apply the same cluster-and-retrieve memory to audio or multi-camera streams, where the same redundancy argument often holds.
  • The paper's own failure analysis suggests the advantage will shrink on heavily edited, text-dense content; a benchmark that controls edit rate would quantify how much of the EgoSchema gain is due to temporal redundancy rather than to the memory design.
  • Because DAM only keeps frames near the largest clusters, questions that hinge on rare, isolated frames (e.g., a single anomalous event) are the likely failure mode; storing a small long-tail partition in DAM would be a testable fix.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes Flash-VStream, a 7B video-LLM that processes long video streams online using a fixed 11,520-token "Flash Memory" comprising a Context Synopsis Memory (K-means cluster centroids of low-resolution frame features), a Detail Augmentation Memory (high-resolution features of key frames retrieved from the largest clusters), and an Adaptive Multimodal RoPE for temporal positioning. A two-process asynchronous framework separates frame encoding and memory update from question answering. The authors report 68.2 on the full EgoSchema benchmark and gains over a token-truncated Qwen2-VL-online baseline on MLVU, LVBench, MVBench, and Video-MME, with response latency under one second, supported by ablations of memory components, clustering and retrieval policies, and capacity allocation.

Significance. If the results hold, the efficiency contribution is substantial: a 7B model can answer questions about hour-long videos from a fixed 11,520-token memory, and the architectural ablations in Tab. 4 support the claim that CSM+DAM improves over uniform sampling under identical training. The paper is also transparent in its limitations section (F.1) and offers a useful critique of GPT-3.5-based open-ended VQA metrics (F.2), and it releases code. However, the headline "equal-cost" comparison is not yet a clean attribution of the gains to the memory design because of the instruction-tuning confound, and the test-set-based selection of R_CSM and Rpool weakens the strength of the reported numbers.

major comments (4)
  1. [§4.1, Tabs. 2–3] The "equal-cost" comparison is confounded by instruction tuning. Flash-VStream is initialized from Qwen2-VL-7b and then LoRA-tuned on a 9k subset of LLaVA-Video (Sec. 4.1, Tab. 7), whereas Qwen2-VL-online is the base Qwen2-VL-7b with its token budget truncated to 11520 and no additional training. The +4.2 EgoSchema, +3.4 MLVU, and +2.2 LVBench gaps therefore cannot be attributed to the Flash Memory mechanism alone. The paper's own ablation row ⑥ vs. row ① in Tab. 4 shows that replacing CSM+DAM with uniform sampling under identical training reduces average accuracy by only about 1.8 points (64.9 to 63.1), much smaller than the claimed gaps over the untrained baseline. The authors should either train Qwen2-VL-online under the same LoRA protocol or evaluate Flash-VStream without instruction tuning.
  2. [§4.4, Fig. 4] The hyperparameters R_CSM and Rpool are selected by grid search on EgoSchema and MVBench, and the same two benchmarks are then reported as final results in Tab. 2. This is a form of test-set selection, so the reported numbers are optimistically biased. The authors should use a validation split or a separate selection procedure, and should report the grid results on all five benchmarks rather than only on the two used for selection.
  3. [§1 vs. §F.1] The opening premise that "temporal redundancy is prevalent in all video types" (Sec. 1) and the unqualified state-of-the-art claim are in tension with the paper's own limitation section F.1, which concedes that text-intensive videos and videos with rapid scene changes break the fixed-memory assumption (Figs. 13–14). The real-time accuracy guarantee is therefore conditional on video style. The authors should temper the general claim or provide a quantitative analysis of which benchmark categories satisfy the assumption.
  4. [§4.1, Tabs. 2–3] All reported results appear to come from single runs without standard errors, multiple seeds, or significance tests. Given the small margins in several comparisons (e.g., +2.2 on LVBench, +1.8 on Video-MME without subtitles), these differences may be within run-to-run variance. The authors should report seeds, error bars, or at least explicitly acknowledge this limitation.
minor comments (6)
  1. [§4.3] The text states that Flash-VStream "performs slightly behind Qwen2-VL on MLVU," but Tab. 2 shows Flash-VStream at 66.3 versus Qwen2-VL* at 66.0. This contradiction should be fixed.
  2. [§3.4, Fig. 3] The label "After 0.7s Time=60min" in Fig. 3 is unclear; please clarify which snapshot it refers to and what 0.7s denotes.
  3. [Tabs. 2–3] For VideoLLM-Online, the entry "2fps" is not a token count; please use consistent units for NVtokens or explain the conversion.
  4. [§4.2, Fig. 6c] Please clarify whether the frame-handler wall time in Fig. 6c is per frame or accumulated over the whole input prefix, since the text discusses the value exceeding one second when input frames exceed 1000.
  5. [§3.4] AM-RoPE is listed as a contribution but is never ablated in the experiments. Please add an ablation isolating AM-RoPE or remove it from the contribution list.
  6. [§4.3] The claim that Qwen2-VL is "difficult to reproduce" under its official setting is not quantified; please state which reported numbers differ and by how much.

Circularity Check

0 steps flagged · score 2.0 of 10

No substantive circularity: Flash-VStream's accuracy and efficiency claims are measured on external benchmarks and attributed via controlled ablations; the two self-citations appear only in related-work positioning and carry none of the load.

full rationale

Flash-VStream is empirically validated rather than derived, so no claim reduces to its inputs by construction. The headline gains (Tab. 2-3) are measured on external benchmarks, and the attribution of those gains to the Flash Memory design rests on controlled ablations in which the training protocol is held fixed: Tab. 4 row 6 (uniform sampling with identical LoRA tuning) is the correct counterfactual, and rows 1 versus 6 show the memory design adds +2.2 EgoSchema and +1.8 average points; Tab. 5-6 further discriminate K-means clustering and Feature-Centric retrieval against alternative policies. The equations in Sec. 3 (Eq. 3-4 CSM centroids; Eq. 5-6 DAM retrieval; Eq. 7-9 interleaving) define the mechanism and are then tested against baselines, not assumed correct, so no self-definitional reduction exists. The two self-citations ([63], [64]) appear only in related-work positioning and carry none of the load; the 'temporal redundancy is prevalent in all video types' premise is cited to external works [20,21,60-62], and no uniqueness theorem or ansatz is smuggled from the authors' prior work. Two caveats should be weighed as correctness risk rather than circularity. First, the equal-cost comparison in Tab. 2-3 pits LoRA-tuned Flash-VStream against an untuned Qwen2-VL-online baseline, so the headline +4.2 EgoSchema gap overstates the memory design's contribution; the paper's own controlled ablation (Tab. 4 rows 1 versus 6) gives the smaller +2.2 estimate, and row 8 reproduces the Qwen2-VL-online token setting. Second, the CSM capacity proportion and pool ratio were selected by grid search directly on the EgoSchema and MVBench test sets (Fig. 4), so those two reported numbers carry mild test-set selection bias; this is hyperparameter selection, not a fitted value that determines the result by construction. The stated limitation F.1 concedes that text-intensive and rapid-cut videos violate the fixed-memory assumption, which bounds the scope of the real-time accuracy claim but does not make the derivation circular. Overall the central claims are self-contained against external benchmarks and controlled comparisons; only minor, non-load-bearing self-citations prevent a score of 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 3 invented entities

The method rests on the empirical assumption of temporal redundancy, a fixed-budget memory whose capacity split is tuned on the evaluation benchmarks, and a hardware-specific real-time budget. No external physical or mathematical constants are introduced.

free parameters (5)
  • CSM capacity proportion R_CSM = 1/3
    Chosen by grid search on EgoSchema and MVBench (Fig. 4) to maximize accuracy under the fixed 11520-token budget.
  • Spatial pool ratio Rpool = 4
    Chosen by grid search on EgoSchema and MVBench (Fig. 4).
  • Token budget NVtokens = 11520
    Selected to satisfy the real-time latency threshold of under one second on an A100 GPU (Fig. 2).
  • Frame extraction rate = 1 fps
    Used in the online evaluation protocol; not justified beyond consistency with prior benchmarks.
  • K-means max iterations T = unspecified
    Appears as a parameter in Algorithm 1 but no value is reported, leaving a gap in exact reproduction.
assumptions (5)
  • domain assumption Temporal redundancy is prevalent in all video types.
    Invoked in Sec. 1 and Sec. 3 to justify compressing a video into a fixed-size memory. The paper's own Limitations F.1 admits text-intensive and rapidly edited videos violate this.
  • domain assumption K-means clustering with a fixed number of centroids effectively models information density distribution.
    Used in Sec. 3.2 to define CSM; empirically tested in Tab. 5 but not proven across all video genres.
  • domain assumption The feature bank can be offloaded to disk without affecting accuracy.
    Stated in Sec. 3.3; relies on I/O assumptions and ignores disk latency effects on the frame handler.
  • domain assumption A 7B model with at most 12000 video tokens achieves sub-second response latency on an A100 GPU.
    Measured in Sec. 4.2 on a specific GPU with FlashAttention-2; may not transfer to other hardware.
  • domain assumption Offline evaluation with questions posed at the end of pre-extracted frames approximates real-time streaming behavior.
    Used in Sec. 4.1; does not test interactive multi-turn questioning during the stream.
invented entities (3)
  • Context Synopsis Memory (CSM)
    purpose: Aggregates long-term temporal information via k-means cluster centroids of low-resolution frame features.
    Component's value is supported only by within-paper ablations; no external falsifiable handle is provided.
  • Detail Augmentation Memory (DAM)
    purpose: Retains high-resolution features of frames nearest to the largest cluster centroids to preserve spatial detail.
    Component's value is supported only by within-paper ablations; no external falsifiable handle is provided.
  • Adaptive Multimodal RoPE (AM-RoPE)
    purpose: Modifies position embeddings to handle compressed tokens at average temporal positions and pooled spatial coordinates.
    Proposed as an engineering adaptation; effectiveness is shown only through the main results, not an isolated study.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Flash-VStream: Efficient Real-Time Understanding for Long Video Streams." pith.science (2026). https://pith.science/paper/C5VS42FP

@misc{pith2026250623825,
  author       = {Pith},
  title        = {Pith review of: Flash-VStream: Efficient Real-Time Understanding for Long Video Streams},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C5VS42FP}},
  note         = {Machine review of arXiv:2506.23825}
}
read the original abstract

Benefiting from the advances in large language models and cross-modal alignment, existing multimodal large language models have achieved prominent performance in image and short video understanding. However, the understanding of long videos is still challenging, as their long-context nature results in significant computational and memory overhead. Most existing work treats long videos in the same way as short videos, which is inefficient for real-world applications and hard to generalize to even longer videos. To address these issues, we propose Flash-VStream, an efficient video language model capable of processing extremely long videos and responding to user queries in real time. Particularly, we design a Flash Memory module, containing a low-capacity context memory to aggregate long-context temporal information and model the distribution of information density, and a high-capacity augmentation memory to retrieve detailed spatial information based on this distribution. Compared to existing models, Flash-VStream achieves significant reductions in inference latency. Extensive experiments on long video benchmarks and comprehensive video benchmarks, i.e., EgoSchema, MLVU, LVBench, MVBench and Video-MME, demonstrate the state-of-the-art performance and outstanding efficiency of our method. Code is available at https://github.com/IVGSZ/Flash-VStream.

Figures

Figures reproduced from arXiv: 2506.23825 by the authors.

Figure 1
Figure 1. Comparison with previous methods. Flash-VStream can understand long videos accurately in an online manner. Here “C” denotes critical clues for questions. ployment, particularly on resource-constrained edge devices. Additionally, excessive computational requirements increase inference latency, directly affecting applications requiring real-time human-computer interactions. In this context, we define a video language … view at source ↗
Figure 2
Figure 2. Response latency / Accuracy on EgoSchema v.s. In￾ference cost. Inputting more than 12000 tokens will not meet real-time requirements. Flash-VStream can respond to user queries in real time while maintaining outstanding performance. (MLLMs) [11, 27–29, 32, 33, 54], many works begin to build Video Language Models (VLMs) based on them. LLaMA￾VID [31] represents single-frame features with only 2 tokens. Chat-UniVi [25] … view at source ↗
Figure 3
Figure 3. Overview of Flash-VStream two-process framework. The frame handler process continuously encodes new frames. The question handler process asynchronously responds to human inquiries in real-time. Flash Memory is composed of interleaved Context Synopsis Memory and Detail Augmentation Memory, organized in chronological order. CSM is updated by clustering low resolution feature maps on an inter-frame level. DAM is update… view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Impact of Pool Ratio Rpool and CSM Capacity Propor￾tion RCSM. The upper left grid is blank since its setting is invalid. feature maps MDAM with size S DAM = NDAM × h × w × d. M˜ CSM = sort(MCSM, key ={|Sk|}NCSM k=1 ) (5) MDAM = {e H f(k) } NDAM k=1 , f(k) = argmin i D(…
Figure 5
Figure 5. Figure 5: Memory Distribution Visualization and Case Study. The left side presents a PCA visualization of the Flash Memory distribution in the feature space. Each point in it stands for a feature map of a single frame or a slice of memory. The CSM and DAM appropriately represent…
Figure 6
Figure 6. Figure 6: (a) Response latency comparison. (b) Max GPU memory comparison. (c) Execution wall time analysis. Response latency refers to the wall time between inputting a question and outputting the first token of the answer. Max GPU memory indicates the peak GPU memory usage duri…
Figure 7
Figure 7. Figure 7: Case Study. This figure presents a case study on documentary video about the 10 longest rivers in Asia, highlighting their lengths, sources, and outflows. The study includes a question regarding the possibility of seeing snow on the banks of these rivers, with multiple…
Figure 8
Figure 8. Figure 8: Case Study. This figure presents a case study involving a cartoon video depicting a mother bird bringing a fish to a fox. The study includes a question about the reason behind this action, with multiple-choice options provided. The ground truth answer is indicated, alo…
Figure 9
Figure 9. Figure 9: Case Study. This figure presents a case study involving an advertising video, depicting various scenes including people by the pool, on the beach, and along a coastal hillside. The study includes a question about the number of people on the staircase at the end of the …
Figure 10
Figure 10. Figure 10: Case Study. This figure presents a case study involving a sports documentary video of badminton tournaments, depicting various matches and players. The study includes a question about the location of the first match, with multiple-choice options provided. The ground t…
Figure 11
Figure 11. Figure 11: Case Study. This figure presents a case study involving a tutorial video depicting various magic tricks. The study includes a question about the order of events in the video, with multiple-choice options provided. The ground truth answer is indicated, along with the p…
Figure 12
Figure 12. Figure 12: Case Study. This figure presents a case study involving a sports video from a high jump competition, depicting various athletes and their performances. The video frames capture moments of intense competition, showcasing the athletes’ skills and determination as they s…
Figure 13
Figure 13. Figure 13: Fail Case Analysis. This figure presents a case study involving a video on space debris and proposed cleanup methods. The video frames illustrate various statistics and methods related to space debris, highlighting the challenges and potential solutions for mitigating…
Figure 14
Figure 14. Figure 14: Fail Case Analysis. This figure presents a case study involving a video showing various animals and their behaviors. The video frames capture different moments of animal interactions and activities, highlighting the diverse behaviors exhibited by the animals. The stud…

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Delayed Bidirectional Alignment via Disentangled Audio Semantics for Audio-Visual Segmentation

    cs.CV 2025-12 conditional novelty 5.0 of 10

    A delayed bidirectional audio-visual alignment framework with bank-grounded disentangled audio queries achieves state-of-the-art segmentation results on AVS-Objects and VPO benchmarks.

  2. Empowering Nanoscale Connectivity through Molecular Communication: A Case Study of Virus Infection

    cs.NI 2025-08 unverdicted novelty 4.0 of 10

    A position paper proposing molecular communication as the link layer for epidemic-control bio-nano networks, with an ORF3a-based mutation identification simulation; the provided manuscript body does not match this abstract.

Reference graph

Works this paper leans on

85 extracted references · 57 canonical work pages · cited by 2 Pith papers

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1

    arXiv 2023

  2. [2]

    Self-calibrated clip for training-free open-vocabulary segmentation

    Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 3

    arXiv 2024

  3. [3]

    Memory consolidation enables long-context video understanding

    Ivana Balazevic, Yuge Shi, Pinelopi Papalampidi, Rahma Chaabouni, Skanda Koppula, and Olivier J Hénaff. Memory consolidation enables long-context video understanding. In ICML, pages 2527–2542, 2024. 3, 4

    2024

  4. [4]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, pages 33, 1877–1901,

    1901

  5. [5]

    A Memory-Network Based Solution for Multivariate Time-Series Forecasting

    Yen-Yu Chang, Fan-Yun Sun, Yueh-Hua Wu, and Shou-De Lin. A memory-network based solution for multivariate time- series forecasting. arXiv preprint arXiv:1809.02105, 2018. 3

    work page Pith review arXiv 2018

  6. [6]

    Distributed deep learning model for intelligent video surveillance systems with edge computing

    Jianguo Chen, Kenli Li, Qingying Deng, Keqin Li, and S Yu Philip. Distributed deep learning model for intelligent video surveillance systems with edge computing. IEEE Transac- tions on Industrial Informatics, 2019. 1

    2019

  7. [7]

    Videollm-online: Online video large language model for streaming video

    Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In CVPR, pages 18407–18418, 2024. 3, 5, 6

    2024

  8. [8]

    Sharegpt4video: Improving video understanding and generation with better captions

    Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Zhenyu Tang, Li Yuan, et al. Sharegpt4video: Improving video understanding and generation with better captions. NeurIPS, 37:19472– 19495, 2025. 1, 5, 6

    2025

Show all 85 references
  1. [9]

    Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model

    Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, pages 640–658. Springer, 2022. 3

    2022

  2. [10]

    Videollama 2: Advancing spatial- temporal modeling and audio understanding in video-llms

    Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial- temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 5, 6

    2024 arXiv

  3. [11]

    Instructblip: Towards general-purpose vision-language models with instruction tuning

    Wenliang Dai, Junnan Li, DONGXU LI, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructblip: Towards general-purpose vision-language models with instruction tuning. NeurIPS, 36,

  4. [12]

    Flashattention-2: Faster attention with better paral- lelism and work partitioning

    Tri Dao. Flashattention-2: Faster attention with better paral- lelism and work partitioning. In ICLR, 2024. 6, 13

    2024

  5. [13]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

    2021

  6. [14]

    The llama 3 herd of models

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Ab- hishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783,

    arXiv

  7. [15]

    A density-based algorithm for discovering clusters in large spatial databases with noise

    Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In KDD, pages 226–231,

  8. [16]

    Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis

    Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. In CVPR, pages 24108–24118, 2025. 2, 5

    2025

  9. [17]

    Temporal sentence grounding in streaming videos

    Tian Gan, Xiao Wang, Yan Sun, Jianlong Wu, Qingpei Guo, and Liqiang Nie. Temporal sentence grounding in streaming videos. In ACM MM, pages 4637–4646, 2023. 1

    2023

  10. [18]

    Mist: Multi-modal iterative spatial- temporal transformer for long-form video question answering

    Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yi Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial- temporal transformer for long-form video question answering. In CVPR, pages 14773–14783, 2023. 2

    2023

  11. [19]

    Clip- adapter: Better vision-language models with feature adapters

    Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip- adapter: Better vision-language models with feature adapters. IJCV, 132(2):581–595, 2024. 3

    2024

  12. [20]

    Frameexit: Conditional early exiting for efficient video recognition

    Amir Ghodrati, Babak Ehteshami Bejnordi, and Amirhossein Habibian. Frameexit: Conditional early exiting for efficient video recognition. In CVPR, pages 15608–15618, 2021. 1

    2021

  13. [21]

    Dynamic neural networks: A survey

    Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. TPAMI, 44(11):7436–7456, 2021. 1

    2021

  14. [22]

    A twofold siamese network for real-time object tracking

    Anfeng He, Chong Luo, Xinmei Tian, and Wenjun Zeng. A twofold siamese network for real-time object tracking. In CVPR, pages 4834–4843, 2018. 3

    2018

  15. [23]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR,

  16. [24]

    Movienet: A holistic dataset for movie understanding

    Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, pages 709–727, 2020. 3

    2020

  17. [25]

    Chat-univi: Unified visual representation em- powers large language models with image and video under- standing

    Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation em- powers large language models with image and video under- standing. In CVPR, pages 13700–13710, 2024. 2, 4, 5, 6, 13, 14

    2024

  18. [26]

    Seed-bench: Benchmarking multimodal large language models

    Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In CVPR, pages 13299– 13308, 2024. 14

    2024

  19. [27]

    Llava-onevision: Easy visual task transfer

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chun- yuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 5, 13, 14

    2024 arXiv

  20. [28]

    Blip: Bootstrapping language-image pre-training for unified vision- language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision- language understanding and generation. In ICML, pages 12888–12900, 2022. 9

    2022

  21. [29]

    Blip- 2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip- 2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742, 2023. 2, 3

    2023

  22. [30]

    Mvbench: A comprehensive multi-modal video understand- ing benchmark

    Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understand- ing benchmark. In CVPR, pages 22195–22206, 2024. 2, 5, 6

    2024

  23. [31]

    Llama-vid: An image is worth 2 tokens in large language models

    Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In ECCV, pages 323–340. Springer, 2025. 2, 5, 6, 13, 14

    2025

  24. [32]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36:34892–34916, 2023. 2

    2023

  25. [33]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296–26306, 2024. 2, 3

    2024

  26. [34]

    Kangaroo: A powerful video-language model supporting long-context video input

    Jiajun Liu, Yibing Wang, Hanghang Ma, Xiaoping Wu, Xi- aoqi Ma, Xiaoming Wei, Jianbin Jiao, Enhua Wu, and Jie Hu. Kangaroo: A powerful video-language model supporting long-context video input. arXiv preprint arXiv:2408.15542,

    arXiv

  27. [35]

    Learning quality-aware dynamic mem- ory for video object segmentation

    Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, and Yujiu Yang. Learning quality-aware dynamic mem- ory for video object segmentation. In ECCV, pages 468–486,

  28. [36]

    Universal segmentation at arbitrary granularity with language instruction

    Yong Liu, Cairong Zhang, Yitong Wang, Jiahao Wang, Yujiu Yang, and Yansong Tang. Universal segmentation at arbitrary granularity with language instruction. In CVPR, pages 3459– 3469, 2024. 3

    2024

  29. [37]

    Oryx mllm: On-demand spatial- temporal understanding at arbitrary resolution

    Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand spatial- temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 5, 6

    2024 arXiv

  30. [38]

    Soc: Semantic- assisted object cluster for referring video object segmentation

    Zhuoyan Luo, Yicheng Xiao, Yong Liu, Shuyan Li, Yitong Wang, Yansong Tang, Xiu Li, and Yujiu Yang. Soc: Semantic- assisted object cluster for referring video object segmentation. NeurIPS, pages 26425–26437, 2023. 3

    2023

  31. [39]

    Multi-task deep learning for real-time 3d human pose estimation and action recognition

    Diogo C Luvizon, David Picard, and Hedi Tabia. Multi-task deep learning for real-time 3d human pose estimation and action recognition. IEEE TPAMI, 43(8):2752–2764, 2020. 3

    2020

  32. [40]

    Vista-llama: Reducing hallucination in video language models via equal distance to visual tokens

    Fan Ma, Xiaojie Jin, Heng Wang, Yuchen Xian, Jiashi Feng, and Yi Yang. Vista-llama: Reducing hallucination in video language models via equal distance to visual tokens. InCVPR, pages 13151–13160, 2024. 1, 2

    2024

  33. [41]

    Video-chatgpt: Towards detailed video under- standing via large vision and language models

    Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Khan. Video-chatgpt: Towards detailed video under- standing via large vision and language models. In ACL, pages 12585–12602, 2024. 14

    2024

  34. [42]

    Some methods for classification and anal- ysis of multivariate observations

    J MacQueen. Some methods for classification and anal- ysis of multivariate observations. In Proceedings of 5-th Berkeley Symposium on Mathematical Statistics and Proba- bility/University of California Press, 1967. 4, 7

    1967

  35. [43]

    Egoschema: A diagnostic benchmark for very long- form video language understanding

    Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long- form video language understanding. NeurIPS, 36:46212– 46244, 2023. 2, 5

    2023

  36. [44]

    Deepres: A deep learning-based video summarization strategy for resource- constrained industrial surveillance scenarios

    Khan Muhammad, Tanveer Hussain, Javier Del Ser, Vasile Palade, and Victor Hugo C De Albuquerque. Deepres: A deep learning-based video summarization strategy for resource- constrained industrial surveillance scenarios. IEEE Trans- actions on Industrial Informatics , 16(9):5938–...

    2019

  37. [45]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Car- roll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. NeurIPS, pages 27730–27744, 2022. 2

    2022

  38. [46]

    Streaming long video understanding with large language models

    Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuangrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. NeurIPS, 37: 119336–119360, 2025. 5, 6

    2025

  39. [47]

    Timechat: A time-sensitive multimodal large language model for long video understanding

    Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, pages 14313–14323,

  40. [48]

    Robovqa: Multimodal long-horizon reasoning for robotics

    Pierre Sermanet, Tianli Ding, Jeffrey Zhao, Fei Xia, De- bidatta Dwibedi, Keerthana Gopalakrishnan, Christine Chan, Gabriel Dulac-Arnold, Sharath Maddineni, Nikhil J Joshi, et al. Robovqa: Multimodal long-horizon reasoning for robotics. In ICRA, pages 645–652. IEEE, 2024. 1

    2024

  41. [49]

    Moviechat: From dense token to sparse memory for long video understanding

    Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. InCVPR, pages 18221–18232, 2024. 2, 4, 5, 6, 7, 13, 14

    2024

  42. [50]

    Roformer: Enhanced transformer with rotary position embedding

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063,

  43. [51]

    Tracking as online decision-making: Learning a policy from streaming videos with reinforcement learning

    James Supancic III and Deva Ramanan. Tracking as online decision-making: Learning a policy from streaming videos with reinforcement learning. In ICCV, pages 322–331, 2017. 1

    2017

  44. [52]

    Dynamic memory based attention network for sequential recommendation

    Qiaoyu Tan, Jianwei Zhang, Ninghao Liu, Xiao Huang, Hongxia Yang, Jingren Zhou, and Xia Hu. Dynamic memory based attention network for sequential recommendation. In AAAI, pages 4384–4392, 2021. 3

    2021

  45. [53]

    Gemini 1.5: Unlocking mul- timodal understanding across millions of tokens of context

    Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking mul- timodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1

    2024 arXiv

  46. [54]

    Kimi-vl technical report

    Kimi Team, Angang Du, Bohong Yin, Bowei Xing, Bowen Qu, Bowen Wang, Cheng Chen, Chenlin Zhang, Chenzhuang Du, Chu Wei, et al. Kimi-vl technical report. arXiv preprint arXiv:2504.07491, 2025. 2

    2025 arXiv

  47. [55]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Mar- tinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Roz- ière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation lang...

    2023 arXiv

  48. [56]

    Llama 2: Open foundation and fine-tuned chat models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Am- jad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya 10 Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 4

    2023 arXiv

  49. [57]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 4, 5, 6, 13, 14

    2024 arXiv

  50. [58]

    Lvbench: An extreme long video understanding benchmark

    Weihan Wang, Zehai He, Wenyi Hong, Yean Cheng, Xiaohan Zhang, Ji Qi, Xiaotao Gu, Shiyu Huang, Bin Xu, Yuxiao Dong, et al. Lvbench: An extreme long video understanding benchmark. arXiv preprint arXiv:2406.08035, 2024. 2, 5

    2024 arXiv

  51. [59]

    Retake: Reducing temporal and knowledge redundancy for long video understanding

    Xiao Wang, Qingyi Si, Jianlong Wu, Shiyu Zhu, Li Cao, and Liqiang Nie. Retake: Reducing temporal and knowledge redundancy for long video understanding. arXiv preprint arXiv:2412.20504, 2024. 2

    2024 arXiv

  52. [60]

    Adaptive focus for efficient video recognition

    Yulin Wang, Zhaoxi Chen, Haojun Jiang, Shiji Song, Yizeng Han, and Gao Huang. Adaptive focus for efficient video recognition. In ICCV, pages 16249–16258, 2021. 1

    2021

  53. [61]

    Adafocus v2: End-to-end training of spatial dynamic net- works for video recognition

    Yulin Wang, Yang Yue, Yuanze Lin, Haojun Jiang, Zihang Lai, Victor Kulikov, Nikita Orlov, Humphrey Shi, and Gao Huang. Adafocus v2: End-to-end training of spatial dynamic net- works for video recognition. In CVPR, pages 20030–20040. IEEE, 2022

    2022

  54. [62]

    Adafocusv3: On unified spatial-temporal dynamic video recognition

    Yulin Wang, Yang Yue, Xinhong Xu, Ali Hassani, Victor Kulikov, Nikita Orlov, Shiji Song, Humphrey Shi, and Gao Huang. Adafocusv3: On unified spatial-temporal dynamic video recognition. In ECCV, pages 226–243. Springer, 2022. 1

    2022

  55. [63]

    Hierarchical memory for long video qa

    Yiqin Wang, Haoji Zhang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Hierarchical memory for long video qa. arXiv preprint arXiv:2407.00603, 2024. 2

    2024 arXiv

  56. [64]

    Ponder & press: Advancing visual gui agent towards general computer control

    Yiqin Wang, Haoji Zhang, Jingqi Tian, and Yansong Tang. Ponder & press: Advancing visual gui agent towards general computer control. arXiv preprint arXiv:2412.01268, 2024. 1

    2024 arXiv

  57. [65]

    Uni-adafocus: Spatial-temporal dynamic computation for video recognition

    Yulin Wang, Haoji Zhang, Yang Yue, Shiji Song, Chao Deng, Junlan Feng, and Gao Huang. Uni-adafocus: Spatial-temporal dynamic computation for video recognition. TPAMI, 2024. 3

    2024

  58. [66]

    Iterprime: Zero-shot referring image segmentation with iterative grad-cam refinement and primary word emphasis

    Yuji Wang, Jingchen Ni, Yong Liu, Chun Yuan, and Yansong Tang. Iterprime: Zero-shot referring image segmentation with iterative grad-cam refinement and primary word emphasis. In AAAI, pages 8159–8168, 2025. 3

    2025

  59. [67]

    Sam2-love: Segment anything model 2 in language- aided audio-visual scenes

    Yuji Wang, Haoran Xu, Yong Liu, Jiaze Li, and Yansong Tang. Sam2-love: Segment anything model 2 in language- aided audio-visual scenes. In CVPR, pages 28932–28941,

  60. [68]

    Towards real-time multi-object tracking

    Zhongdao Wang, Liang Zheng, Yixuan Liu, Yali Li, and Shengjin Wang. Towards real-time multi-object tracking. In ECCV, pages 107–122, 2020. 3

    2020

  61. [69]

    Videollm-mod: Efficient video- language streaming with mixture-of-depths vision compu- tation

    Shiwei Wu, Joya Chen, Kevin Qinghong Lin, Qimeng Wang, Yan Gao, Qianli Xu, Tong Xu, Yao Hu, Enhong Chen, and Mike Zheng Shou. Videollm-mod: Efficient video- language streaming with mixture-of-depths vision compu- tation. NeurIPS, 37:109922–109947, 2024. 5

    2024

  62. [70]

    Next-qa: Next phase of question-answering to explaining temporal actions

    Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In CVPR, pages 9777–9786, 2021. 14

    2021

  63. [71]

    Longvila: Scaling long-context visual language models for long videos

    Fuzhao Xue, Yukang Chen, Dacheng Li, Qinghao Hu, Ligeng Zhu, Xiuyu Li, Yunhao Fang, Haotian Tang, Shang Yang, Zhi- jian Liu, et al. Longvila: Scaling long-context visual language models for long videos. arXiv preprint arXiv:2408.10188,

    arXiv

  64. [72]

    Fine-grained video captioning via graph-based multi- granularity interaction learning

    Yichao Yan, Ning Zhuang, Bingbing Ni, Jian Zhang, Minghao Xu, Qiang Zhang, Zheng Zhang, Shuo Cheng, Qi Tian, Yi Xu, et al. Fine-grained video captioning via graph-based multi- granularity interaction learning. TPAMI, 44(2):666–683, 2019. 3

    2019

  65. [73]

    Qwen2 technical report

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 2, 3

    2024 arXiv

  66. [74]

    Language-aware vision transformer for referring segmentation

    Zhao Yang, Jiaqi Wang, Xubing Ye, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Language-aware vision transformer for referring segmentation. TPAMI, 2024. 3

    2024

  67. [75]

    Atp-llava: Adaptive token pruning for large vision language models

    Xubing Ye, Yukang Gan, Yixiao Ge, Xiao-Ping Zhang, and Yansong Tang. Atp-llava: Adaptive token pruning for large vision language models. In CVPR, pages 24972–24982, 2025. 2

    2025

  68. [76]

    V oco-llama: Towards vision compression with large language models

    Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, and Yansong Tang. V oco-llama: Towards vision compression with large language models. In CVPR, pages 29836–29846,

  69. [77]

    Self-chained image-language model for video localization and question answering

    Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. Advances in Neural Information Processing Systems, 36:76749–76771, 2023. 2

    2023

  70. [78]

    Activitynet-qa: A dataset for understanding complex web videos via question answering

    Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, pages 9127–9134, 2019. 14

    2019

  71. [79]

    Real-time action recognition with enhanced motion vector cnns

    Bowen Zhang, Limin Wang, Zhe Wang, Yu Qiao, and Hanli Wang. Real-time action recognition with enhanced motion vector cnns. In CVPR, pages 2718–2726, 2016. 3

    2016

  72. [80]

    Long context transfer from language to vision

    Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 2, 5, 6

    2024 arXiv

  73. [81]

    Video instruction tuning with synthetic data

    Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 5, 12

    2024 arXiv

  74. [82]

    Mlvu: A comprehensive benchmark for multi-task long video understanding

    Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2, 5

    2024 arXiv

  75. [83]

    Streaming dense video captioning

    Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In CVPR, pages 18243–18252, 2024. 3

    2024

  76. [84]

    In- starevive: One-step image enhancement via dynamic score matching

    Yixuan Zhu, Haolin Wang, Ao Li, Wenliang Zhao, Yansong Tang, Jingxuan Niu, Lei Chen, Jie Zhou, and Jiwen Lu. In- starevive: One-step image enhancement via dynamic score matching. arXiv preprint arXiv:2504.15513, 2025. 3 11 Flash-VStream: Efficient Real-Time Understanding for L...

    2025 arXiv

  77. [1000]

    Although this may result in delayed updates of visual information, it would not affect the response latency. Overall, the results highlight the efficiency advantages of Flash-VStream in terms of both response latency and GPU memory consumption, making it a competitive choice f...

    1920

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.