Pith. sign in

REVIEW 3 major objections 5 minor 54 references

Persistent Object Narratives for Token-Efficient Video Language Models

T0 review · 3 major / 5 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 SlotNarrative claims that a video can be compressed into a fixed 144-position visual interface—one identity token plus eight state tokens per persistent object—that keeps a frozen video-language model accurate on three open-ended video QA…

desk verdict A coherent slot-based interface for token-efficient VideoQA that deserves serious refereeing, though the headline numbers are weakened by an apparent test-set selection of its operating point. read the letter →

arxiv 2608.04866 v1 pith:KVV6UR7D submitted 2026-08-05 cs.CV

Junzhe Chen , Siyuan Meng , Xiaojie Guo This is my paper · ORCID
classification cs.CV
keywords videolanguagemodelstoken-efficientvisualinterfaceobject-centriclearningslotattentionpersistentobjectmemoryquestionansweringidentitytokensstate
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

SlotNarrative sets out to show that a video can be reduced to a fixed 144-position visual interface for a frozen video-language model by organizing the video into persistent object narratives instead of per-frame tokens. Each narrative is serialized as one identity token representing what the object is and eight segment-level state tokens representing what happens to it, with a parameter-free memory linking repeating slot observations across gaps and scene changes. The paper reports 75.6% accuracy on MSVD-QA, 69.8% on MSRVTT-QA, and 50.3% on ActivityNet-QA at 144 allocated visual-token positions, a favorable reported accuracy–token trade-off among compact interfaces. The reason to care is that the result suggests object-level structure, rather than simply more compression, may be what allows video-language models to stay accurate while spending very few tokens.

What carries the argument

The mechanism that carries the argument is a parameter-free persistent object memory followed by a typed tokenizer. The memory scores each reappearing slot against every inactive entry with a weighted sum of five cues—slot-feature, pooled-appearance, trajectory-state, recent-observation, and position similarity—using fixed weights $(0.3,0.2,0.2,0.1,0.2)$ and an acceptance threshold of $0.3$, then applies greedy one-to-one matching so the entry can be recovered even under a different slot index after a missing interval or scene cut. The typed tokenizer pools the matched observations into one identity token and $J=8$ state tokens per entry, binding them with a shared object-index embedding and marking token type and segment order. This converts exchangeable frame-local slots into fixed-size clip-level object narratives, which is why the interface stays at exactly $M(1+J)=144$ positions.

What would settle it

Run the parameter-free memory on a video collection with ground-truth object tracks, focusing on clips where objects leave and re-enter the frame or shots change, and compare the persistent IDs with the track IDs; chance-level re-identification accuracy would falsify the claim that identity tokens summarize persistent objects.

Watch

Extended reading notes

Core claim

The paper's central claim is that linking frame-local slot observations into persistent clip-level object entries before tokenization yields a visual interface that carries enough object-level evidence for a frozen video-language model to answer open-ended questions. Concretely, SlotNarrative keeps at most $M=16$ entries, divides the video into $J=8$ temporal segments, and serializes each entry as one identity token plus eight state tokens, for a fixed $M(1+J)=144$ allocated positions independent of the number of sampled frames. With this interface the model reports 75.6% on MSVD-QA, 69.8% on MSRVTT-QA, and 50.3% on ActivityNet-QA, improving on the closest compact slot-based system while using 25% fewer positions. The paper argues that the identity/state split is what makes this budget work: identity tokens pool stable appearance across all visible observations, state tokens retain segment-level geometry, visibility, trajectory, and re-identification events, and a shared object embedding keeps both token types bound to the same persistent entry.

Load-bearing premise

The load-bearing premise is that the fixed memory rules—the five cue weights and the 0.3 threshold—correctly re-identify the same object when it reappears after missing intervals or scene cuts, which the paper never validates against ground-truth tracks; if those associations are wrong, the identity and state tokens no longer represent persistent objects and the claimed benefit collapses.

Editorial extensions

If this is right

  • With 144 allocated positions independent of sampled-frame count, the language-model context no longer grows with video length; the same layout applies whatever the number of input frames.
  • Against the closest compact object-centric baseline at 192 positions, the interface uses 25% fewer tokens while improving the reported accuracy by 0.7 points on MSVD-QA, 0.2 on MSRVTT-QA, and 2.0 on ActivityNet-QA.
  • Ablations show association before serialization is the main lever: removing the persistent memory costs 3.2, 2.1, and 2.5 points on the three benchmarks, and the trajectory encoder adds a further 1.4–2.7 points on top of memory alone.
  • At nearby token budgets, cutting the object budget from 16 to 8 entries costs 4.0 accuracy points, while halving the temporal segments from 8 to 4 costs only 1.3 points, so object coverage matters more than temporal resolution.
  • Identity-only and state-only variants fall 6.2 and 7.9 points short of the full interface, whereas the typed combination holds 50.3% on ActivityNet-QA, showing the two token types are complementary.

Reading between the lines

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

  • The paper does not test the interface on arbitrarily long videos or with a growing number of objects; a natural extension is a streaming or incremental inventory where the 144 positions are reused across time windows.
  • Since the paper's own linear probes find almost no decodable velocity or frame-order signal in the trajectory state, the next accuracy gains for temporal questions likely lie in a stronger temporal encoder, not in adding more tokens.
  • The authors note the persistent IDs are algorithmic associations rather than ground-truth tracks, so a track-level audit of the memory would determine whether the QA gains come from correct re-identification or from grouping that is merely useful in aggregate.
  • The identity/state serialization is a general pattern: any domain where an entity's stable properties should be reported separately from its changes could reuse the two-token-type layout, although the paper demonstrates it only for video QA.
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 / 5 minor

Summary. The paper proposes SlotNarrative, a slot-based visual interface for frozen Video-LLMs. A frozen SigLIP backbone and a recurrent slot encoder produce per-frame object-like observations; a parameter-free memory with hand-set matching weights links these observations into persistent clip-level object entries, and a typed tokenizer serializes each retained entry into one identity token and J state tokens. With M=16 objects and J=8 segments, the interface allocates a fixed 144 visual-token positions per video, independent of sampled frame count. The visual modules and Qwen2-7B-Instruct are frozen; only projectors, layer norms, and typed embeddings are trained on VideoInstruct-100K. The paper reports 75.6%, 69.8%, and 50.3% accuracy on MSVD-QA, MSRVTT-QA, and ActivityNet-QA, respectively, and claims a favorable accuracy-to-token trade-off among compact Video-LLM interfaces. Within-model ablations show that trajectory encoding, persistent memory, and typed identity/state tokens each contribute to the reported accuracy. The paper is explicit that cross-paper comparisons are not controlled and that association IDs are algorithmic rather than ground-truth tracks.

Significance. If the results hold, the contribution is significant: a compact, object-structured visual interface that reduces video input to a fixed 144-token budget while retaining competitive VideoQA accuracy, with a frozen language model and a parameter-free association memory. The paper ships unusually detailed reproducibility settings (seed, decoding, judge prompt, hardware, training schedule) and includes within-model ablations and representation diagnostics that support the contribution of the memory and typed tokens. The authors also clearly flag the main limitations: cross-paper comparisons are operating-point comparisons, results are single runs, and persistent association is not evaluated against ground-truth tracks. Those strengths make the paper worth serious consideration, but the headline accuracy/token trade-off currently rests on two load-bearing issues that need to be resolved before the central claim can be trusted.

major comments (3)
  1. [Section 4.1 vs. Section 4.3 / Appendix D.2] The statement in Section 4.1 that QA annotations are never used for training or model selection appears to be contradicted by the budget ablations in Section 4.3 and Appendix D.2. Appendix D.2 reports ActivityNet-QA results for 80-, 144-, and 216-position variants and says that the modest gain beyond 144 positions 'supports the selected operating point.' If the default M=16, J=8 was fixed after inspecting these test-set numbers, then the reported headline accuracies (75.6%, 69.8%, 50.3%) are a test-selected operating point and the favorable token-efficiency comparison is not an unbiased estimate. The manuscript should state explicitly whether the default was fixed before any eval-set ablations were run, and if not, the selection should be repeated on a validation split or the reported numbers should be labeled as exploratory rather than as the primary result.
  2. [Section 4.3, Table S4] The claim of diminishing returns beyond the 144-position setting is supported by very small differences from a single training run: Table S4 shows 216 positions giving 50.9% versus 50.3% at 144 positions (+0.6) and 272 positions giving 50.6% (+0.3). With no error bars or multiple seeds, these differences are within the range of normal training noise, and the conclusion that 144 positions is the right operating point is not statistically supported. The paper should provide multiple seeds or error bars for at least the budget and component ablations, or soften the claim about the selected operating point.
  3. [Section 3.3, Appendix A.3, Section 4.4] The persistent association stage is load-bearing: the identity and state tokens are only meaningful if recurring slot observations are correctly linked to the same clip-level object. Yet the paper explicitly states that the IDs are 'algorithmic associations rather than ground-truth tracks' (Section 3.3) and that the qualitative results 'do not measure association accuracy' (Section 4.4). The same-entry appearance probe in Appendix D.4 measures consistency among model-assigned associations, not correctness. I am not claiming the associations are wrong; I am asking for a direct measurement. The authors should evaluate association quality against ground-truth tracks on a suitable annotated benchmark (e.g., DAVIS, TAO, or a manually labeled subset of ActivityNet) and report metrics such as ID switches, merges, and re-identification precision, and ideally relate this to the downstream QA gains.
minor comments (5)
  1. [Abstract] The abstract contains formatting artifacts with missing spaces (e.g., 'Videolargelanguagemodels'), which should be corrected in the camera-ready version.
  2. [Table 1 and Section 4.2] The token-count comparison would be clearer if the caption specified for each baseline whether the reported number is allocated capacity, active tokens, or a rate (e.g., '729/GOP' and '2/frame'), and whether the 144 value is the pre-mask allocation or the average active count.
  3. [Figure 5] The sentence describing the dashed segment as joining 'the two non-dominated plotted points' is ambiguous; please specify which points are connected and why the segment illustrates the trade-off.
  4. [Appendix D.4] The temporal-signal diagnostic is said to be undefined when c1 - cx is below 1e-6; the paper should state how such cases are handled in the reported summary value.
  5. [Appendix C.5] The evaluation protocol says malformed judge responses are retried up to three times, but it does not state what happens if all retries are malformed; this should be specified (e.g., counted as incorrect or excluded).

Circularity Check

0 steps flagged · score 0.0 of 10

No load-bearing circularity: the fixed 144-token interface is evaluated against external VideoQA benchmarks and no reported result reduces by construction to a fitted parameter or self-citation.

full rationale

The paper's central claim is that a 144-position slot-based visual interface, using one identity token and J state tokens per persistent object entry, achieves competitive VideoQA accuracy on MSVD-QA, MSRVTT-QA, and ActivityNet-QA. This claim is tested against held-out benchmark annotations, and Section 4.1 states that the QA annotations are never used for training or model selection. The matching weights, threshold, and inventory ranking are fixed hand-set rules or learned on unlabeled ActivityNet videos and VideoInstruct conversations, not fitted to the reported accuracy numbers; no equation in the paper defines the reported accuracy in terms of these constants. The identity/state token decomposition is a design choice rather than a derivation that assumes the result it explains, and the paper explicitly disclaims that its persistent IDs are ground-truth tracks: "These IDs are algorithmic associations rather than ground-truth tracks," and the Appendix D.4 probe "measures model consistency rather than ground-truth re-identification accuracy." The paper contains no load-bearing self-citations by the current authors; cited training objectives and baselines are external prior work. One validity concern exists outside circularity: Section D.2 ablates M/J budgets on the ActivityNet-QA evaluation split and concludes the modest gain "supports the selected operating point," which could indicate test-set operating-point selection and would weaken the unbiasedness of the reported accuracy. That is a statistical contamination concern, not a circular derivation, so it does not raise the circularity score. No step was found where a prediction equals its input by construction, a fitted parameter is renamed as a prediction, or a central premise reduces to a self-citation chain.

Assumptions & free parameters 8 free parameters · 7 assumptions · 0 invented entities

The central claim rests on slot attention reliably finding objects, on hand-set memory rules preserving identity across gaps, on attention-derived presence labels, and on the adequacy of the ActivityNet/VideoInstruct training data and GPT-3.5 evaluation. None of these are demonstrated with ground-truth tracking or fully controlled baselines.

free parameters (8)
  • Memory cue weight vector = (0.3, 0.2, 0.2, 0.1, 0.2)
    Hand-set weights for slot-feature, pooled-appearance, trajectory-state, recent-observation, and position cues in the matching score, Eq. (1), Appendix A.3. Not learned and not varied in ablations.
  • Re-identification acceptance threshold = 0.3
    Hand-chosen threshold on the normalized matching score in Appendix A.3; controls whether a reappearing slot is linked to an existing entry.
  • Gaussian position bandwidth = 0.1
    Bandwidth for the velocity-predicted centroid proximity cue in Appendix A.3.
  • Scene-cut detection threshold = 0.35
    HSV-histogram Bhattacharyya distance threshold that triggers recurrence resets and matching-cue changes, Appendix A.4.
  • Existence and visibility confidence thresholds = 1[c>3] and 1[c>5]
    Attention-concentration derived labels used as presence supervision in Stage 2, Appendix A.2. The cutoffs 3 and 5 are arbitrary.
  • EMA momentum for memory summaries = 0.9
    Exponential moving average update rate for entry summaries in Appendix A.3.
  • Inventory ranking coefficient = 0.5
    Fixed coefficient on the state-change term in the inventory score Imp_v, Appendix B.1. Not varied across datasets.
  • Object and segment counts = M=16, J=8
    Interface capacity chosen for the default 144-token layout. Ablations in Table S4 show nearby budgets perform similarly, so the choice is not sharply constrained.
assumptions (7)
  • domain assumption Slot attention groups visual features into object-like units that correspond to real objects in the video.
    Used throughout Stage 1 and Figure 2; no ground-truth masks are available. If slots do not align with objects, the narratives are not object narratives.
  • domain assumption Recurrent slot initialization provides short-range identity within a shot.
    Invoked in Section 3.2 and used for trajectory prediction targets in Stage 2, Appendix C.3.
  • domain assumption Attention-pooled raw SigLIP features preserve semantic appearance of grouped regions.
    Used for identity prototypes and state appearance in Appendix B.2; the paper's diagnostics support this partially but not with ground-truth reference.
  • ad hoc to paper The hand-set memory matching rules can re-identify objects across gaps and scene cuts.
    The fixed weights, threshold, and invisible-to-visible trigger in Section 3.3 and Appendix A.3 are not validated against ground-truth tracks.
  • ad hoc to paper Attention concentration thresholds provide reliable existence and visibility labels.
    Confidence labels c>3 and c>5 in Appendix A.2 replace missing object masks and determine memory update behavior.
  • domain assumption VideoInstruct-100K ActivityNet videos are sufficient to train the visual interface and language alignment.
    All three stages use this single video collection; no other visual pretraining data is used.
  • domain assumption GPT-3.5 judge accuracy and quality scores reflect answer correctness.
    All reported metrics depend on the GPT-3.5 judge prompt and model version, which the paper acknowledges may differ across published reports.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Persistent Object Narratives for Token-Efficient Video Language Models." pith.science (2026). https://pith.science/paper/KVV6UR7D

@misc{pith2026260804866,
  author       = {Pith},
  title        = {Pith review of: Persistent Object Narratives for Token-Efficient Video Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KVV6UR7D}},
  note         = {Machine review of arXiv:2608.04866}
}
read the original abstract

Video large language models (Video-LLMs) have made strong progress in open-ended video understanding. However, their visual interfaces remain token-intensive and provide limited explicit structure for linking recurring object evidence across time. We introduce SlotNarrative, a slot-based interface that organizes a video into persistent object narratives represented by compact object-state tokens. Rather than compressing frame-wise features before establishing temporal correspondence, SlotNarrative first groups visual features into object-like slots and then associates recurring observations with clip-level object entries through a lightweight, parameter-free memory that integrates multiple complementary matching cues. Each retained entry is serialized into two token types: an identity token that summarizes persistent object appearance and a set of state tokens that encode segment-level appearance, geometry, visibility, and trajectory information. This design yields an interface of only 144 allocated visual-token positions for a frozen Video-LLM, independent of the number of sampled frames. Across multiple datasets, SlotNarrative achieves a favorable trade-off between accuracy and visual-token count compared with prior compact Video-LLM interfaces. Experimental results establish persistent object narratives as a compact, structured, and temporally organized visual interface for Video-LLMs. Our code will be made publicly available.

Figures

Figures reproduced from arXiv: 2608.04866 by the authors.

Figure 1
Figure 1. Visual interfaces for Video-LLMs. (a) Dense frame tokens are costly and leave object identity implicit. (b) Per-frame [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of SlotNarrative. A slot encoder groups SigLIP features into temporally propagated observations. Parameter-free memory links observations across missing intervals, a fixed inventory retains clip-level entries, and trainable typed projectors represent each entry with one identity token and multiple segment-level state tokens. or slot reassignment. Dual-State Slot Attention (Tran et al. 2026) separates appear… view at source ↗
Figure 3
Figure 3. Persistent association across missing observations. The colored regions show [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Typed object-state tokenization. Slot masks pool backbone appearance along observations linked by persistent memory. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Reported ActivityNet-QA accuracy versus per [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Successful ActivityNet-QA examples. Each row pairs six event-aligned frames with the question, reference, and [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 44 canonical work pages

  1. [1]

    Maaz, Muhammad and Rasheed, Hanoona and Khan, Salman and Khan, Fahad , booktitle=

    work page

  2. [2]

    Lin, Bin and Ye, Yang and Zhu, Bin and Cui, Jiaxi and Ning, Munan and Jin, Peng and Yuan, Li , booktitle=

    work page

  3. [3]

    Cheng, Zesen and Leng, Sicong and Zhang, Hang and Xin, Yifei and Li, Xin and Chen, Guanzheng and Zhu, Yongxin and Zhang, Wenqi and Luo, Ziyang and Zhao, Deli and others , journal=

    work page

  4. [4]

    Li, Bo and Zhang, Yuanhan and Guo, Dong and Zhang, Renrui and Li, Feng and Zhang, Hao and Zhang, Kaichen and Zhang, Peiyuan and Li, Yanwei and Liu, Ziwei and others , journal=

    work page

  5. [5]

    Li, Hongyu and Chen, Jinyu and Wei, Ziyu and Huang, Shaofei and Hui, Tianrui and Gao, Jialin and Wei, Xiaoming and Liu, Si , booktitle=

    work page

  6. [6]

    Zhang, Yuanhan and Wu, Jinming and Li, Wei and Li, Bo and Ma, Zejun and Liu, Ziwei and Li, Chunyuan , journal=

  7. [7]

    2024 , organization=

    Li, Yanwei and Wang, Chengyao and Jia, Jiaya , booktitle=. 2024 , organization=

    work page 2024

  8. [8]

    Song, Enxin and Chai, Wenhao and Wang, Guanhong and Zhang, Yucheng and Zhou, Haoyang and Wu, Feiyang and Chi, Haozhe and Guo, Xun and Ye, Tian and Zhang, Yanting and others , booktitle=

    work page

Show all 54 references
  1. [9]

    and Soran, Bilge and Krishnamoorthi, Raghuraman and Elhoseiny, Mohamed and Chandra, Vikas , booktitle=

    Shen, Xiaoqian and Xiong, Yunyang and Zhao, Changsheng and Wu, Lemeng and Chen, Jun and Zhu, Chenchen and Liu, Zechun and Xiao, Fanyi and Varadarajan, Balakrishnan and Bordes, Florian and Liu, Zhuang and Xu, Hu and Kim, Hyunwoo J. and Soran, Bilge and Krishnamoorthi, Raghurama...

  2. [10]

    Object-Centric Learning with

    Locatello, Francesco and Weissenborn, Dirk and Unterthiner, Thomas and Mahendran, Aravindh and Heigold, Georg and Uszkoreit, Jakob and Dosovitskiy, Alexey and Kipf, Thomas , booktitle=. Object-Centric Learning with

  3. [11]

    International Conference on Learning Representations , year=

    Bridging the Gap to Real-World Object-Centric Learning , author=. International Conference on Learning Representations , year=

  4. [12]

    International Conference on Learning Representations , year=

    Conditional Object-Centric Learning from Video , author=. International Conference on Learning Representations , year=

  5. [13]

    Advances in Neural Information Processing Systems , volume=

    Object-Centric Learning for Real-World Videos by Predicting Temporal Feature Similarities , author=. Advances in Neural Information Processing Systems , volume=

  6. [14]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Temporally Consistent Object-Centric Learning by Contrasting Slots , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  7. [15]

    Advances in Neural Information Processing Systems , volume=

    Slot-VLM: Object-Event Slots for Video-Language Modeling , author=. Advances in Neural Information Processing Systems , volume=

  8. [16]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    VideoOrion: Tokenizing Object Dynamics in Videos , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  9. [17]

    Efficient Motion-Aware Video

    Zhao, Zijia and Huo, Yuqi and Yue, Tongtian and Guo, Longteng and Lu, Haoyu and Wang, Bingning and Chen, Weipeng and Liu, Jing , booktitle=. Efficient Motion-Aware Video

  10. [18]

    Yashima, Daichi and Kurita, Shuhei and Oda, Yusuke and Sugiura, Komei , booktitle=

  11. [19]

    Proceedings of the 25th ACM International Conference on Multimedia , pages=

    Video Question Answering via Gradually Refined Attention over Appearance and Motion , author=. Proceedings of the 25th ACM International Conference on Multimedia , pages=

  12. [20]

    Yu, Zhou and Xu, Dejing and Yu, Jun and Yu, Ting and Zhao, Zhou and Zhuang, Yueting and Tao, Dacheng , booktitle=

  13. [21]

    Zhang, Hang and Li, Xin and Bing, Lidong , booktitle=

  14. [22]

    Jin, Peng and Takanobu, Ryuichi and Zhang, Wancai and Cao, Xiaochun and Yuan, Li , booktitle=

  15. [23]

    Wang, Peng and Bai, Shuai and Tan, Sinan and Wang, Shijie and Fan, Zhihao and Bai, Jinze and Chen, Keqin and Liu, Xuejing and Wang, Jialin and Ge, Wenbin and others , journal=

  16. [24]

    2023 , organization=

    Li, Junnan and Li, Dongxu and Savarese, Silvio and Hoi, Steven , booktitle=. 2023 , organization=

    2023

  17. [25]

    Dai, Wenliang and Li, Junnan and Li, Dongxu and Tiong, Anthony and Zhao, Junqi and Wang, Weisheng and Li, Boyang and Fung, Pascale N and Hoi, Steven , booktitle=

  18. [26]

    Ryoo, Michael S and Piergiovanni, AJ and Arnab, Anurag and Dehghani, Mostafa and Angelova, Anelia , booktitle=

  19. [27]

    Token Merging: Your

    Bolya, Daniel and Fu, Cheng-Yang and Dai, Xiaoliang and Zhang, Peizhao and Feichtenhofer, Christoph and Hoffman, Judy , booktitle=. Token Merging: Your

  20. [28]

    Yang, Senqiao and Chen, Yukang and Tian, Zhuotao and Wang, Chengyao and Li, Jingyao and Yu, Bei and Jia, Jiaya , booktitle=

  21. [29]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Unified Spatiotemporal Token Compression for Video-LLMs at Ultra-Low Retention , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  22. [30]

    Shen, Leqi and Gong, Guoqiang and He, Tao and Zhang, Yifeng and Liu, Pengzhang and Zhao, Sicheng and Ding, Guiguang , booktitle=

  23. [31]

    Shao, Kele and Tao, Keda and Qin, Can and You, Haoxuan and Sui, Yang and Wang, Huan , booktitle=

  24. [32]

    Ma, Fan and Jin, Xiaojie and Wang, Heng and Xian, Yuchen and Feng, Jiashi and Yang, Yi , booktitle=

  25. [33]

    International Conference on Learning Representations , year =

    Mu Cai and Jianwei Yang and Jianfeng Gao and Yong Jae Lee , title =. International Conference on Learning Representations , year =

  26. [34]

    Ryoo, Michael S. and Zhou, Honglu and Kendre, Shrikant and Qin, Can and Xue, Le and Shu, Manli and Park, Jongwoo and Ranasinghe, Kanchana and Savarese, Silvio and Xu, Ran and Xiong, Caiming and Niebles, Juan Carlos , journal=

  27. [35]

    He, Bo and Li, Hengduo and Jang, Young Kyun and Jia, Menglin and Cao, Xuefei and Shah, Ashish and Shrivastava, Abhinav and Lim, Ser-Nam , booktitle=

  28. [36]

    Lan, Xiaohan and Yuan, Yitian and Jie, Zequn and Ma, Lin , journal=

  29. [37]

    European Conference on Computer Vision , pages=

    An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Models , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  30. [38]

    Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops , pages=

    STORM: Token-Efficient Long Video Understanding for Multimodal LLMs , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops , pages=

  31. [39]

    Zhang, Haichao and Fu, Yun , booktitle=

  32. [40]

    arXiv preprint arXiv:1901.11390 , year=

    MONet: Unsupervised Scene Decomposition and Representation , author=. arXiv preprint arXiv:1901.11390 , year=

    1901 arXiv

  33. [41]

    Proceedings of the 36th International Conference on Machine Learning , volume=

    Multi-Object Representation Learning with Iterative Variational Inference , author=. Proceedings of the 36th International Conference on Machine Learning , volume=. 2019 , publisher=

    2019

  34. [42]

    International Conference on Learning Representations , year=

    Illiterate DALL-E Learns to Compose , author=. International Conference on Learning Representations , year=

  35. [43]

    Advances in Neural Information Processing Systems , volume=

    SAVi++: Towards End-to-End Object-Centric Learning from Real-World Videos , author=. Advances in Neural Information Processing Systems , volume=

  36. [44]

    Advances in Neural Information Processing Systems , volume=

    Simple Unsupervised Object-Centric Learning for Complex and Naturalistic Videos , author=. Advances in Neural Information Processing Systems , volume=

  37. [45]

    International Conference on Learning Representations , year=

    SlotFormer: Unsupervised Visual Dynamics Simulation with Object-Centric Models , author=. International Conference on Learning Representations , year=

  38. [46]

    and Ionescu, Radu Tudor , journal=

    Grigore, Diana-Nicoleta and Madan, Neelu and Mogelmose, Andreas and Moeslund, Thomas B. and Ionescu, Radu Tudor , journal=

  39. [47]

    arXiv preprint arXiv:2606.12601 , year=

    Dual-State Slot Attention: Decoupling Appearance and Identity for Video Object-Centric Learning , author=. arXiv preprint arXiv:2606.12601 , year=

    arXiv

  40. [48]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    One Trajectory, One Token: Grounded Video Tokenization via Panoptic Sub-object Trajectory , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  41. [49]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    TrajTok: Learning Trajectory Tokens Enhances Video Understanding , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  42. [50]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Sigmoid Loss for Language Image Pre-Training , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  43. [51]

    International Conference on Learning Representations , year=

    From Vicious to Virtuous Cycles: Synergistic Representation Learning for Unsupervised Video Object-Centric Learning , author=. International Conference on Learning Representations , year=

  44. [52]

    arXiv preprint arXiv:2407.10671 , year=

    Qwen2 Technical Report , author=. arXiv preprint arXiv:2407.10671 , year=

    arXiv

  45. [53]

    arXiv preprint arXiv:2405.07798 , year=

    FreeVA: Offline MLLM as Training-Free Video Assistant , author=. arXiv preprint arXiv:2405.07798 , year=

    arXiv

  46. [54]

    International Conference on Learning Representations , year=

    SlowFast-LLaVA: A Strong Training-Free Baseline for Video Large Language Models , author=. International Conference on Learning Representations , year=

Pith tools

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