Pith. sign in

REVIEW 4 major objections 5 minor 30 references

StreamFlow: Dynamic Memory Flows for Streaming Video Understanding

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read StreamFlow claims that a frozen video-language model can match or beat retrained streaming models purely by managing visual memory.

desk verdict StreamFlow is a well-engineered memory recipe with careful measurement, but its headline numbers are in-sample because hyperparameters were tuned on the evaluation benchmarks, so the SOTA claim doesn't hold as stated. read the letter →

arxiv 2608.10949 v1 pith:3ZDYHU2V submitted 2026-08-11 cs.CV cs.CL

Muxin Fu , Yifan Zhang , Wentao Zhang , Fangming Guo , Qian Chen , Guibin Zhang , Shuicheng Yan , Bo An This is my paper · ORCID
classification cs.CVcs.CL
keywords streamingvideounderstandingvisualmemorymultimodallargelanguagemodelsattention-guidedretrievalconsolidationtemporalredundancyfilteringquestionansweringefficientinference
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

StreamFlow tries to establish that streaming video understanding does not require retraining the multimodal backbone: a frozen multimodal large language model (MLLM) can handle an unbounded video stream if the memory around it decides what to encode and when to recall. The paper combines a dynamics-aware mid-term memory, which drops static pixels before visual encoding, with a latent long-term memory, which consolidates older footage into a fixed number of visual latents. During answer generation, an attention-guided trigger re-injects retrieved latents when the model's attention to visual evidence weakens. On StreamingBench the system reaches 67.73% overall accuracy, and relative to a vanilla encoder it raises the mean visual attention score by 59.1% while cutting end-to-end latency by 50.4% and peak memory by 21.1%. If these numbers hold, memory policy can substitute for backbone surgery in real-time video reasoning.

What carries the argument

The load-bearing mechanism is the visual attention score $\mathrm{VAS}_t$, the layer- and head-averaged attention mass that the token being generated places on accessible visual tokens; a score below $\tau=0.10$ triggers retrieval. Retrieval scores each stored GOP by the maximum cosine similarity between a pooled text-query embedding and its frame embeddings, takes the top $K=4$, and compresses them through learned queries into $L=32$ latents inserted after the current prefix. The mid-term memory is built from patch-level temporal residuals $R^g_f(u)=X^g_f(u)-X^g_0(u)$ against each GOP's I-frame, keeping the top $\lceil \rho N\rceil$ patches per P-frame with $\rho=0.5$, while the long-term memory merges the most similar adjacent GOP pair when its capacity of 96 GOPs is full.

What would settle it

Compare correct-answer log-odds on the StreamingBench RTVU subset under VAS-triggered injection versus random or fixed-position injection matched in the number and length of injected latents; if the gap is zero, the attention trigger itself adds nothing over a larger static visual prefix.

Watch

Extended reading notes

Core claim

The paper's central claim is that the two known weaknesses of memory-based streaming video MLLMs—wasted encoding on redundant frames and rigid access to history—can be removed together by a continuous visual memory flow. StreamFlow selectively encodes only dynamic patches, consolidates old groups of pictures into latent tokens under a fixed capacity, and injects retrieved latents into the generated sequence exactly when the visual attention score falls below a threshold. The authors report that this design is sufficient to outperform prior streaming and long-video methods, and the controlled comparisons with the same backbone attribute the gains to the memory design rather than to model size.

Load-bearing premise

The load-bearing premise, introduced in Section 3.4, is that the visual attention score is a reliable online signal of when the model actually needs historical visual evidence; if a low VAS does not track grounding need, the injection policy adds tokens at the wrong times and the reported gains would reduce to the static benefit of a larger visual prefix.

Editorial extensions

If this is right

  • A frozen MLLM can keep its visual grounding through a long answer if historical evidence is re-injected when attention to visual tokens decays.
  • Temporal redundancy can be filtered in raw pixel space before ViT encoding without losing essential dynamics: at 50% P-frame retention, 61.08% of local motion is preserved and 95.28% of static patches rejected.
  • Long-video evidence can be compressed into a bounded latent memory: removing long-term memory costs 10.44 points on VideoMME-Long, so consolidation is what makes hour-long videos tractable.
  • Streaming efficiency does not require sacrificing accuracy: the same design reduces context length by 37.8%, KV-cache memory by 37.8%, and self-attention latency by 63.4% while improving accuracy.
  • The memory design transfers to offline long-video benchmarks, improving MLVU and VideoMME over streaming baselines that are adapted to the same backbone.

Reading between the lines

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

  • The VAS trigger is a general pattern: any autoregressive MLLM that watches its own attention could decide when to fetch external evidence, so the mechanism may transfer to retrieval-augmented generation and agentic settings, though the paper only evaluates video QA.
  • The matched-versus-shuffled memory experiment isolates relevance, but a stronger test would swap the trigger for random positions; until then the 0.27 log-odds gain is the measured size of the grounding effect.
  • The reported latency excludes GOP waiting time and frame-arrival delay, so in a live camera setting the 50.4% saving would shrink unless frame collection is overlapped with encoding.
  • Because filtering happens on raw pixel residuals, very small objects that move slowly could be dropped at low retention ratios; testing task-level accuracy at $\rho=0.125$ on fine-grained recognition would reveal the boundary.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes StreamFlow, a memory framework that equips a frozen multimodal LLM with streaming video understanding. The method has three parts: a dynamics-aware mid-term memory that selects temporally changing patches in raw pixel space before visual encoding; a latent long-term memory that consolidates older groups of pictures into a fixed number of visual latents; and an attention-guided injection policy that retrieves and inserts long-term latents during generation when the model's visual attention score falls below a threshold. The authors report 67.73% overall accuracy on StreamingBench, along with strong results on MVBench, MLVU, and VideoMME, and claim a 50.4% latency reduction and 21.1% peak-memory reduction relative to a vanilla encoder. A controlled comparison with selected baselines reimplemented on the same Qwen3.5-9B backbone is reported in Appendix C.1, where the StreamingBench margin over StreamingTOM is 1.17 points.

Significance. If the reported numbers hold, the paper would offer a lightweight and modular way to give frozen MLLMs streaming capability, with two genuinely interesting ideas: filtering temporal redundancy before visual encoding, and dynamically reactivating long-term visual evidence during generation. Strengths of the manuscript include the controlled same-backbone comparison in Table 5, the explicit measurement boundaries for the efficiency results in Appendix C.4, the counterfactual matched-versus-shuffled memory experiment in Appendix B.2, and the multi-way ablations in Figure 5. However, the headline state-of-the-art claim is not yet established: the main comparison in Table 1 is between a 9B model at 1 fps and 7B baselines running at 0.2–1 fps, and the core hyperparameters (P-frame retention ratio, latent length, insertion policy) are selected using the evaluation benchmarks themselves without a held-out split or code release. The mechanistic claim about visual grounding is also partly circular and only weakly supported by the counterfactual evidence (Spearman rho = 0.24).

major comments (4)
  1. [Section 4.4 and Appendix C.3, Table 4] The headline accuracy of 67.73% on StreamingBench is an in-sample estimate because key hyperparameters are selected using the evaluation benchmarks themselves. Figure 5(a) chooses the P-frame retention ratio rho = 0.5 using RTVU accuracy, Figure 8 chooses latent length L = 32 using MLVU accuracy, and Figure 5(b) chooses VAS-guided insertion using RTVU accuracy; Table 4 then fixes these values for all reported results. Without a held-out split or repeated nested evaluation, the 1.17-point same-backbone margin over StreamingTOM in Table 5 could shrink or invert on fresh data. Since no code or data release is provided, the selection process cannot be audited. This issue is load-bearing because every downstream claim (grounding, efficiency, offline generalization) inherits the credibility of the headline accuracy.
  2. [Table 1 and Table 5] The state-of-the-art statement in the abstract is weakened by mismatched backbones and frame rates. In Table 1, StreamFlow-9B runs at 1 fps and 9B parameters, while the strongest reported baselines are 7B models at 0.5 or 0.2 fps. Under the controlled same-backbone comparison in Table 5, the StreamingBench advantage over StreamingTOM is only 1.17 points, and the VideoMME advantage over ReKV is 1.63 points; these margins are small relative to the variability one would expect from hyperparameter sensitivity and benchmark noise. The authors should either present the headline SOTA claim only for the controlled comparison, or provide confidence intervals and multiple-seed results for the matched-backbone experiments.
  3. [Section 3.4 and Appendix B.2] The claim that attention-guided injection improves visual grounding is partly circular. VAS is defined as the attention mass assigned to accessible visual tokens (Eq. 8), and the injection policy inserts new visual latents when VAS falls below a threshold; therefore, a rise in VAS after injection is expected by construction. The counterfactual evidence in Appendix B.2 is suggestive but weak: matched versus shuffled long-term memory changes correct-answer log-odds by only 0.27 with Spearman rho = 0.24, and disabling insertion costs only 1.37 accuracy points on RTVU (Figure 5b). Please provide a stronger validation that VAS tracks task-relevant grounding need, for example by comparing VAS-guided insertion against an oracle that inserts at question-relevant timestamps, or by controlling for the effect of simply adding more visual tokens.
  4. [Appendix C.4 and Table 3] The 50.4% end-to-end latency reduction is measured under a boundary that the authors themselves describe as 'offline benchmark execution rather than per-frame service latency' and that excludes frame-arrival time, disk video decoding, model loading, and compilation. For a streaming video method, the wall-clock latency in a real deployment would also include the time to collect frames at 1 fps and the GOP waiting delay. The paper should either report a service-latency model that includes frame arrival, or explicitly revise the abstract and conclusion to say that the latency reduction refers to offline processing time after the video records are resident in memory.
minor comments (5)
  1. [Abstract and Table 1] The abstract claims 'surpassing the previous best result by 4.63%' on StreamingBench, but Table 1 lists StreamingBench results for only two methods (LiveVLM and StreamFlow) among the baselines; the other entries are unreported. Please clarify how the 4.63% margin is computed and which baselines actually report StreamingBench numbers.
  2. [Figure 5(b)] The x-axis label 'No insert.' is truncated; it should read 'No insertion'.
  3. [Equation (6) and Section 3.3] Memory consolidation merges the most similar adjacent pair based only on I-frame token similarity. Since P-frames can carry crucial transitional evidence, please discuss or ablate the case where two GOPs have similar I-frames but distinct P-frame content.
  4. [Table 3 and Appendix C.4] The visual front end reduces Pre-ViT patches by 37.3% but visual-front-end latency by 50.0%; given the measurement boundary described in Appendix C.4, it would be helpful to explain why the latency reduction exceeds the patch reduction, since residual scoring and sparse packing add overhead.
  5. [Appendix C.4] The VAS computation is said to be accumulated over all layers and heads at every decoding step; please report the fraction of total task latency attributable to this online VAS computation, since it is part of the claimed efficiency gain.

Circularity Check

2 steps flagged · score 6.0 of 10

The 59.1% VAS grounding gain is the controller's own setpoint (self-definitional), and key hyperparameters are tuned on the evaluation benchmarks before the same benchmarks are reported as SOTA; the headline accuracy is still an external measurement with a controlled same-backbone comparison.

  1. self definitional [Section 3.4, Eq. (8); Section 4.3, Figure 4(a); Appendix B.1]
    "VAS is defined as the visual attention mass averaged across all layers and heads: VAS_t = (1/N_l N_h) Σ_l Σ_h Σ_{c∈V_t} A^{l,h}_{t,c}. ... A low VAS_t indicates that linguistic context is dominating visual evidence, prompting StreamFlow to reactivate long-term memory for subsequent generation. ... In contrast, StreamFlow maintains higher visual attention throughout decoding, raising mean VAS from 0.066 under Vanilla to 0.105, a relative gain of 59.1%."

    The metric used to claim improved grounding is the same quantity the controller is built to act on. Eq. (8) counts every accessible visual token in V_t, and Appendix B.1 states that V_t includes 'any long-term memory latents inserted earlier in generation'; the trigger is VAS_t < τ. So injecting latents when VAS drops and then measuring higher VAS is the controller tracking its own setpoint, not independent evidence of grounding. The matched-vs-shuffled counterfactual (Figure 4b) is the genuine independent test, but its reported effect is modest (log-odds gain 0.27, Spearman ρ = 0.24), so the headline 59.1% VAS gain remains largely self-definitional.

  2. fitted input called prediction [Section 4.4, Figure 5(a); Appendix C.3, Figure 8; Table 4]
    "On RTVU, Figure 5(a) analyzes the retention ratio ρ. Increasing the ratio from 0.25 to 0.50 improves accuracy from 78.75% to 81.55% ... motivating the default ratio of 0.50 as a balance between accuracy and visual-token cost. ... As shown in Figure 8, accuracy increases monotonically from 71.16% at L=4 to 75.83% at L=64. ... Consequently, the default L=32 improves over L=4 by 4.18 points."

    The hyperparameters ρ and L are selected by accuracy on RTVU and MLVU—the same evaluation benchmarks whose results are later reported as the method's performance in Tables 1 and 4. No held-out split or audit trail is reported, so the headline 67.73% StreamingBench and 75.34% MLVU numbers are in-sample estimates after selection on those labels. This is a fitted-input-called-prediction pattern: the reported 'prediction' of SOTA accuracy is partly produced by optimizing the reported quantity itself, rather than by an independent confirmation of the design.

full rationale

StreamFlow's central accuracy claims are external benchmark measurements, and the controlled Qwen3.5-9B comparison in Appendix C.1 gives the method independent content; no load-bearing self-citation chain was found (the cited Vismem and Persistent-Visual-Memory works appear only as background support). The partial circularity comes from two places. First, the claimed 59.1% VAS improvement is the controller's own setpoint: VAS (Eq. 8) is defined over all accessible visual tokens, injection is triggered when VAS falls below τ, and the post-insertion VAS then reflects the added tokens. The matched-vs-shuffled counterfactual is the right independent check, but its effect is modest, so the headline VAS gain is mostly self-definitional. Second, the default ρ and L are chosen on RTVU and MLVU accuracy (Figures 5a and 8) and then fixed in Table 4 for all reported numbers; with no held-out split, the 67.73% StreamingBench and 75.34% MLVU results are partly in-sample. These issues reduce the strength of the mechanistic and SOTA claims but do not make the entire derivation vacuous.

Assumptions & free parameters 7 free parameters · 4 assumptions · 3 invented entities

The framework depends on several hand-set hyperparameters, most consequentially rho, L, and tau; none are derived from theory, and sensitivity analyses are run on the same benchmarks that produce the headline numbers. The three architectural components are new modules with no external falsifiable handle beyond the paper's own evaluations, so the empirical claims carry the evidentiary weight.

free parameters (7)
  • P-frame retention ratio rho = 0.5
    Selected from the RTVU accuracy curve in Figure 5(a); increases from 0.25 to 0.50 add 2.80 points, larger ratios only 0.85 and 0.55 points.
  • VAS threshold tau = 0.1
    Hand-set trigger for long-term memory injection; no sensitivity ablation is reported, and it directly controls how often memory is re-injected.
  • Latent length per injection L = 32
    Tuned on MLVU accuracy in Figure 8; L=32 captures 89.5 percent of the improvement from L=4 to L=64 while halving token count.
  • Retrieved GOPs per injection K = 4
    Hand-set retrieval budget; no sensitivity analysis is reported.
  • Long-term memory capacity C = 96 GOPs
    Hand-set fixed budget for consolidation; no sensitivity analysis is reported.
  • GOP length and mid-term capacity = T=4 frames, 32 frames
    Hand-set temporal grouping and recent-window size; no sensitivity analysis is reported.
  • Injection scheduling limits = 5 max injections, 8-step min interval
    Hand-set bounds on dynamic memory overhead; no sensitivity analysis is reported.
assumptions (4)
  • domain assumption Raw-pixel temporal residuals are a faithful proxy for task-relevant new visual evidence.
    Invoked in Equations (1)-(3) and Figure 3; if residual scoring discards semantically important static content, the mid-term memory is lossy before encoding.
  • domain assumption Visual attention score tracks the model's need for visual grounding.
    Used as the online injection trigger in Section 3.4; the paper's own support is modest (matched-shuffled log-odds gain 0.27, Spearman rho=0.24, insertion ablation 1.37 points).
  • domain assumption Similarity-guided merging of adjacent GOPs preserves question-relevant historical information.
    Memory consolidation in Equations (6)-(7) merges the most similar adjacent pair; if the merge removes distinctive evidence, long-term recall is degraded.
  • domain assumption The frozen visual encoder and projector remain adequate for encoding and retrieving compressed latents.
    The answering backbone is frozen and only a LoRA compressor is trained; if the frozen encoder cannot represent consolidated latents well, the framework depends on this untested generality.
invented entities (3)
  • Latent long-term memory with GOP-level visual latents and neural compressor
    purpose: Bounded storage of historical visual evidence beyond the mid-term window, with similarity-guided merging.
    Benefit is shown through ablations and benchmark scores on the same data used to select hyperparameters; no falsifiable prediction outside the paper's evaluation set.
  • Dynamics-aware mid-term memory
    purpose: Selective raw-pixel patch retention before visual encoding, using reference-anchored residual scores.
    Evaluated through retention-policy comparisons and efficiency metrics within the paper; no external handle.
  • Attention-guided memory injection policy
    purpose: Trigger retrieval and insertion of long-term latents when VAS falls below threshold tau.
    The VAS increase is partly definitional; the counterfactual grounding analysis is internal to the paper and has no external predictive test.

how reviews work

0 comments
Cite this review

Pith. "Pith review of StreamFlow: Dynamic Memory Flows for Streaming Video Understanding." pith.science (2026). https://pith.science/paper/3ZDYHU2V

@misc{pith2026260810949,
  author       = {Pith},
  title        = {Pith review of: StreamFlow: Dynamic Memory Flows for Streaming Video Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3ZDYHU2V}},
  note         = {Machine review of arXiv:2608.10949}
}
read the original abstract

Streaming video understanding requires multimodal large language models (MLLMs) to preserve relevant evidence from continuously evolving streams under strict causality and bounded memory. Yet existing paradigms remain limited: model-based methods require intrusive backbone updates, while memory-based methods expend substantial visual-encoding computation on temporally redundant content and rely on rigid access to visual history. To address these limitations, we introduce StreamFlow, an efficient visual memory framework that enables dynamic, on-demand access to historical visual information. StreamFlow combines a lightweight, dynamics-aware mid-term memory that filters temporal redundancy before visual encoding with a latent long-term memory that consolidates historical video content into visual latents accessible to subsequent reasoning. During generation, an attention-guided retrieval mechanism injects relevant visual latents when the model's reliance on visual evidence weakens. StreamFlow achieves state-of-the-art streaming video understanding performance, reaching 67.73% overall accuracy on StreamingBench, while also delivering strong performance on offline long-video benchmarks. Relative to the vanilla setting, it improves the visual attention score (VAS) by 59.1% while reducing end-to-end latency and peak memory by 50.4% and 21.1%, respectively, enabling more visually grounded and efficient reasoning.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 6 canonical work pages

  1. [1]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631,

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631,

    arXiv

  2. [4]

    Streamkv: Streaming video question-answering with segment-based kv cache retrieval and compression

    Yilong Chen, Xiang Bai, Zhibin Wang, Chengyu Bai, Yuhan Dai, and Ming Lu. Streamkv: Streaming video question-answering with segment-based kv cache retrieval and compression. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 3120–3128, 2026b. Shangzhe Di, Zhelun Yu, Guanghao Zhang, Haoyuan Li, Hao Cheng, Bolin Li, Wanggui He...

    work page 2025

  3. [5]

    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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24108–24118, 2025a. Haoyu Fu, Dia...

    arXiv

  4. [6]

    Persistent Visual Memory: Sustaining Perception for Deep Generation in LVLMs

    Siyuan Huang, Xiaoye Qu, Yafu Li, Tong Zhu, Zefeng He, Muxin Fu, Daizong Liu, Wei-Long Zheng, and Yu Cheng. Persistent visual memory: Sustaining perception for deep generation in lvlms.arXiv preprint arXiv:2605.00814,

    work page Pith review arXiv

  5. [8]

    More thinking, less seeing? assessing amplified hallucination in multimodal reasoning models.arXiv preprint arXiv:2505.21523,

    Chengzhi Liu, Zhongxing Xu, Qingyue Wei, Juncheng Wu, James Zou, Xin Eric Wang, Yuyin Zhou, and Sheng Liu. More thinking, less seeing? assessing amplified hallucination in multimodal reasoning models.arXiv preprint arXiv:2505.21523,

    arXiv

  6. [9]

    Thinking in streaming video.arXiv preprint arXiv:2603.12938,

    Zikang Liu, Longteng Guo, Handong Li, Ru Zhen, Xingjian He, Ruyi Ji, Xiaoming Ren, Yanhao Zhang, Haonan Lu, and Jing Liu. Thinking in streaming video.arXiv preprint arXiv:2603.12938,

    arXiv

  7. [10]

    From narrow to panoramic vision: Attention-guided cold-start reshapes multimodal reasoning.arXiv preprint arXiv:2603.03825,

    Ruilin Luo, Chufan Shi, Yizhen Zhang, Cheng Yang, Songtao Jiang, Tongkun Guan, Ruizhe Chen, Ruihang Chu, Peng Wang, Mingkun Yang, et al. From narrow to panoramic vision: Attention-guided cold-start reshapes multimodal reasoning.arXiv preprint arXiv:2603.03825,

    arXiv

  8. [11]

    Watch, remember, reason: Human-view video understanding with mllms.arXiv preprint arXiv:2606.07433,

    Jiahao Meng, Yue Tan, Qi Xu, Kuan Gao, Weisong Liu, Yanwei Li, Jason Li, Lingdong Kong, Haochen Wang, Qianyu Zhou, et al. Watch, remember, reason: Human-view video understanding with mllms.arXiv preprint arXiv:2606.07433,

    arXiv

Show all 30 references
  1. [12]

    Livevlm: Efficient online video understanding via streaming-oriented kv cache and retrieval.arXiv preprint arXiv:2505.15269,

    Zhenyu Ning, Guangda Liu, Qihao Jin, Chengwei Li, Wenchao Ding, Minyi Guo, and Jieru Zhao. Livevlm: Efficient online video understanding via streaming-oriented kv cache and retrieval.arXiv preprint arXiv:2505.15269,

    arXiv

  2. [13]

    Cognitive memory in large language models

    Lianlei Shan, Shixian Luo, Zezhou Zhu, Yu Yuan, and Yong Wu. Cognitive memory in large language models. arXiv preprint arXiv:2504.02441,

    arXiv

  3. [14]

    A simple baseline for streaming video understanding

    Yujiao Shen, Shulin Tian, Jingkang Yang, and Ziwei Liu. A simple baseline for streaming video understanding. arXiv preprint arXiv:2604.02317,

  4. [15]

    Towards a dynamic and fixed-budget memory bank for efficient streaming video understanding.arXiv preprint arXiv:2606.25658,

    Baiyang Song, Yuli Lin, Qiong Wu, Tao Chen, Jun Peng, Xiao Chen, Yiyi Zhou, and Rongrong Ji. Towards a dynamic and fixed-budget memory bank for efficient streaming video understanding.arXiv preprint arXiv:2606.25658,

    arXiv

  5. [16]

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

    EnxinSong,WenhaoChai,GuanhongWang,YuchengZhang,HaoyangZhou,FeiyangWu,HaozheChi,XunGuo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449,

    arXiv

  6. [18]

    Onevision-encoder: Codec-aligned sparsity as a foundational principle for multimodal intelligence.arXiv preprint arXiv:2602.08683,

    Feilong Tang, Xiang An, Yunyao Yan, Yin Xie, Bin Qin, Kaicheng Yang, Yifei Shen, Yuanhan Zhang, Chunyuan Li, Shikun Feng, et al. Onevision-encoder: Codec-aligned sparsity as a foundational principle for multimodal intelligence.arXiv preprint arXiv:2602.08683,

  7. [19]

    Qwen3.5-omni technical report.arXiv preprint arXiv:2604.15804,

    Qwen Team. Qwen3.5-omni technical report.arXiv preprint arXiv:2604.15804,

    arXiv

  8. [20]

    Star: A benchmark for situated reasoning in real-world videos.arXiv preprint arXiv:2405.09711, 2024a

    Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos.arXiv preprint arXiv:2405.09711, 2024a. Shiwei Wu, Joya Chen, Kevin Qinghong Lin, Qimeng Wang, Yan Gao, Qianli Xu, Tong Xu, Yao Hu, Enhong Chen, ...

    2026 arXiv

  9. [21]

    Fluxmem: Adaptive hierarchical memory for streaming video understanding.arXiv preprint arXiv:2603.02096,

    Yiweng Xie, Bo He, Junke Wang, Xiangyu Zheng, Ziyi Ye, and Zuxuan Wu. Fluxmem: Adaptive hierarchical memory for streaming video understanding.arXiv preprint arXiv:2603.02096,

  10. [22]

    Streammem: Query-agnostic kv cache memory for streaming video understanding.arXiv preprint arXiv:2508.15717,

    Yanlai Yang, Zhuokai Zhao, Satya Narayan Shukla, Aashu Singh, Shlok Kumar Mishra, Lizhu Zhang, and Mengye Ren. Streammem: Query-agnostic kv cache memory for streaming video understanding.arXiv preprint arXiv:2508.15717,

    arXiv

  11. [23]

    Clevrer: Collision events for video representation and reasoning.arXiv preprint arXiv:1910.01442,

    Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning.arXiv preprint arXiv:1910.01442,

    1910 arXiv

  12. [25]

    Memory in large language models: Mechanisms, evaluation and evolution.arXiv preprint arXiv:2509.18868,

    Dianxing Zhang, Wendong Li, Kani Song, Jiaye Lu, Gang Li, Liuchun Yang, and Sheng Li. Memory in large language models: Mechanisms, evaluation and evolution.arXiv preprint arXiv:2509.18868,

  13. [26]

    Hermes: Kv cache as hierarchical memory for efficient streaming video understanding.arXiv preprint arXiv:2601.14724,

    Haowei Zhang, Shudong Yang, Jinlan Fu, See-Kiong Ng, and Xipeng Qiu. Hermes: Kv cache as hierarchical memory for efficient streaming video understanding.arXiv preprint arXiv:2601.14724,

    arXiv

  14. [27]

    Llava-video: Video instruction tuning with synthetic data.arXiv preprint arXiv:2410.02713,

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

    arXiv

  15. [28]

    An empirical study on parameter-efficient fine-tuning for multimodal large language models

    12 StreamFlow: Dynamic Memory Flows for Streaming Video Understanding Xiongtao Zhou, Jie He, Yuhua Ke, Guangyao Zhu, Víctor Gutiérrez-Basulto, and Jeff Pan. An empirical study on parameter-efficient fine-tuning for multimodal large language models. InFindings of the Associatio...

    2024

  16. [29]

    From seconds to hours: Reviewing multimodal large language models on comprehensive long video understanding.arXiv preprint arXiv:2409.18938,

    Heqing Zou, Tianze Luo, Guiyang Xie, Fengmao Lv, Guangcong Wang, Junyang Chen, Zhuochen Wang, Hansheng Zhang, Huaijian Zhang, et al. From seconds to hours: Reviewing multimodal large language models on comprehensive long video understanding.arXiv preprint arXiv:2409.18938,

    arXiv

  17. [30]

    Temporal distance, content similarity, and merge frequency determine event consolidation, while a spatiotemporal window stores recent observations

    organizes long-term visual history into event-level trees through a Persistent Event Memory Forest. Temporal distance, content similarity, and merge frequency determine event consolidation, while a spatiotemporal window stores recent observations. • ReKV(Dietal.,2025)storeshis...

    2025

  18. [2019]

    Vismem: Latent vision memory unlocks potential of vision-language models.arXiv preprint arXiv:2511.11007,

    Xinlei Yu, Chengming Xu, Guibin Zhang, Zhangquan Chen, Yudong Zhang, Yongbo He, Peng-Tao Jiang, Jiangning Zhang, Xiaobin Hu, and Shuicheng Yan. Vismem: Latent vision memory unlocks potential of vision-language models.arXiv preprint arXiv:2511.11007,

  19. [2023]

    Cradle: Empowering foundation agents towards general computer control.arXiv preprint arXiv:2403.03186,

    Weihao Tan, Wentao Zhang, Xinrun Xu, Haochong Xia, Ziluo Ding, Boyu Li, Bohan Zhou, Junpeng Yue, Jiechuan Jiang, Yewen Li, et al. Cradle: Empowering foundation agents towards general computer control.arXiv preprint arXiv:2403.03186,

    arXiv

  20. [2024]

    Scaling the long video understanding of multimodal large language models via visual memory mechanism.arXiv preprint arXiv:2603.29252, 2026a

    Tao Chen, Kun Zhang, Qiong Wu, Xiao Chen, Chao Chang, Xiaoshuai Sun, Yiyi Zhou, and Rongrong Ji. Scaling the long video understanding of multimodal large language models via visual memory mechanism.arXiv preprint arXiv:2603.29252, 2026a. Xueyi Chen, Keda Tao, Kele Shao, and Hu...

  21. [2025]

    arXiv preprint arXiv:2410.24164,

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.𝜋0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164,

    arXiv

  22. [2026]

    Streamingbench: Assessing the gap for mllms to achieve streaming video understanding

    Junming Lin, Zheng Fang, Chi Chen, Haoxuan Cheng, Zihao Wan, Fuwen Luo, Ziyue Wang, Peng Li, Yang Liu, and Maosong Sun. Streamingbench: Assessing the gap for mllms to achieve streaming video understanding. InICASSP 2026-2026 IEEE International Conference on Acoustics, Speech a...

    2026

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.