Pith. sign in

REVIEW 4 major objections 5 minor 48 references

At 8x token compression, CRAFT keeps about 97% of a video VLM's accuracy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review

2026-08-04 23:38 UTC pith:W556NPO7

load-bearing objection Solid, useful token compression for video VLMs with believable results; needs code, a disclosed τ/threshold, and a real orientation ablation before I'd trust the details. the 4 major comments →

arxiv 2608.01644 v1 pith:W556NPO7 submitted 2026-08-03 cs.CV cs.AI

CRAFT: Compression via Recursive Adaptive Fusion of Video Tokens for Vision-Language Models

Yu Chen , Xiaohong Li , Xiaole Wang , Jianjin Zhang , Jun Sun , Yafeng Deng This is my paper
classification cs.CV cs.AI
keywords video token compressionvision-language modelstoken mergingselect-then-fusegated fusionquery-agnostic compressionKV cache reuseefficient video understanding
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

CRAFT is a module that compresses the visual token stream a video vision-language model sends to its language backbone, and the paper's central claim is that it can cut that stream by roughly 8x while retaining about 97% of the uncompressed model's average accuracy on six video benchmarks. The method works by repeatedly pairing similar tokens using a parameter-free global similarity computed in a low-dimensional random-projection space, then fusing each pair with a small learned network: a position-aware weighting scorer assigns global weights across tokens, and a per-channel gate decides, for each dimension, how much of the source token to keep. Because a merged token is always a linear combination of original tokens, it keeps its spatio-temporal coordinates and stays in the embedding distribution the language model was pretrained on; because compression does not depend on the question, one compressed representation can be cached and reused across turns. The paper reports that CRAFT beats every training-free baseline at 8x, degrades more gracefully than they do from 2x to 32x, and transfers unchanged to two other backbones. The load-bearing premise, stated in Section 3.2, is that the larger-L2-norm token in a similar pair carries the more salient semantics, so discarding the smaller one is safe.

Core claim

The paper introduces what it calls the first end-to-end iterative trainable 'select-then-fuse' compressor for general-purpose visual token reduction. Selection is training-free and global: each round, every active token nominates its nearest neighbor by cosine similarity in a fixed random-projection space; the highest-similarity edges are truncated to budget, oriented so the higher-L2-norm endpoint survives, chain conflicts are dropped, and groups form around survivors. Fusion is learnable: Eq. (4) rewrites the survivor as x_t minus a weighted, per-channel-gated difference from its sources, so merging continuously spans hard dropping and equal-weight averaging. A two-stage curriculum trains

What carries the argument

The central object is the gated merge update x_t^new = x_t - sum_i w_i G_i ⊙ (x_t - x_i), in which x_t is the surviving representative token and {x_i} are its paired sources. w_i comes from a position-aware scorer that applies 3D rotary encodings (M-RoPE) and softmax over sources, so global spatial/temporal relations and selection similarity both influence how much each source contributes; G_i is a per-channel gate in (0,1)^D produced by an MLP with adaptive layer normalization conditioned on the merge round, zero-initialized so training starts at a neutral 0.5. Setting all G_i=1 gives weighted averaging and setting all G_i=0 gives hard pruning, so the learned gate interpolates between the t

Load-bearing premise

The load-bearing premise is that a token's vector length (L2 norm) reliably measures how important its content is, so when two tokens look similar, the longer one can stand for both and the shorter one can be thrown away.

What would settle it

Take several hundred videos whose questions hinge on a small or low-contrast object; compute the L2 norms of the tokens inside that object's bounding box and check whether they rank below the frame median. If they do, and CRAFT at 8x answers those questions worse than a control that forces the object's tokens to survive (or than a version that keeps the lower-norm token as representative), the norm-as-salience premise is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • At a fixed 8x budget, CRAFT achieves the best or tied best accuracy on all six benchmarks, so the gain is not an artifact of adaptive budget reallocation.
  • The accuracy gap over training-free baselines widens as compression tightens: at 32x CRAFT stays near 59% while baselines fall to 53.5-58%, implying learnable fusion matters most exactly where token budgets are hardest.
  • Because the compressor is query-agnostic, one forward pass over the video produces a reusable compressed KV cache; the paper reports 19.0 MB KV per sample vs 26.0-26.2 MB for baselines and the lowest prefill latency among compared methods.
  • The same merge module transfers to two other backbones with over 95% retention at 8x, indicating the mechanism does not depend on one vision encoder or token layout.
  • A LoRA-adapted variant reaches 66.9% at 10x and even exceeds the uncompressed backbone on some benchmarks, suggesting the compressor and light LLM adaptation are complementary.

Where Pith is reading between the lines

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

  • A direct extension would replace the L2-norm orientation with a learned salience scorer; if norm and task relevance diverge, a learned orientation could recover information CRAFT currently discards before any fusion happens.
  • Because every retained token is a linear combination of original tokens with preserved coordinates, CRAFT's compressed representation should compose with other sequence-level optimizations such as later-layer pruning or speculative decoding without re-alignment, which the paper does not test.
  • An adversarial evaluation targeting small, low-contrast, or briefly appearing objects would stress the norm-salience premise; on such data a variant that retains low-norm tokens by query relevance would be the natural control.
  • The position-aware gating is defined over (t,h,w), so the same select-then-fuse loop could in principle be applied to image grids or multi-view inputs with only a change in the coordinate prior.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces CRAFT, a query-agnostic visual-token compression method for video vision-language models. It operates in an iterative 'select-then-fuse' loop: a parameter-free global similarity selection (based on random-projections and ℓ2-norm orientation) decides which tokens to merge, and a lightweight learnable gated merging module (position-aware global weighting plus per-channel gating) decides how to fuse them. The authors claim that at about 8× compression CRAFT retains roughly 97% of the backbone's average accuracy, outperforms existing training-free and (on a secondary backbone) learnable compressors, transfers across three backbones, and reduces prefill latency and KV-cache memory. The method is evaluated on six video benchmarks with Qwen3.5-4B as the primary backbone, plus Qwen2.5-VL-7B and LLaVA-OV-7B for transfer.

Significance. If the claims hold, CRAFT is a practical and timely contribution: it attacks the prefill bottleneck for long-video VLMs while keeping the compressed representation in the original token space, preserving spatio-temporal coordinates, and remaining query-agnostic for KV reuse. The paper is well-structured, with explicit pseudocode, equations, and ablations; the cross-backbone transfer and the scaffold-vs-merging decomposition (Fig. 4) are particularly informative. The main empirical result—96.8% retained accuracy at a realized 8.0× compression—is consistent with the reported table. However, the validation of the core orientation heuristic is incomplete, and several hyperparameters needed for reproduction are missing. These issues are fixable but currently prevent the paper from being fully accepted as-is.

major comments (4)
  1. [§4.5 / Table 3; §3.2, Eq. (4)] The ablation in Table 3 does not actually test the orientation assumption stated in §3.2 (larger ℓ2-norm endpoint becomes the representative). The rows 'mean mean' vs 'mean target' differ in the reduction rule—equal-weight averaging of the whole group vs. keeping only the target—not in which endpoint is chosen as target. The text's claim that the +1.00 improvement 'validates our assumption' is therefore unsupported: the comparison shows that, given the norm-based target, hard-pruning sources beats averaging them; it does not test whether keeping the larger-norm token is better than keeping the smaller-norm token or choosing randomly. An ablation comparing 'larger-norm target', 'smaller-norm target', and random orientation under the same fusion rule is needed. This is load-bearing because orientation is the only content-dependent heuristic in the selection stage, and all subsequent rounds
  2. [§3.3, Eq. (6)] The softmax temperature τ in Eq. (6) is never given a value anywhere in the manuscript, including the hyperparameter table (Table 4) and the appendices. The position-aware weighting scorer is a core learnable module, so the method is not fully specified without knowing whether τ is fixed (and at what value) or learned. Please report τ explicitly, or remove it from the equation if it is always 1.
  3. [§3.1, §3.2, Algorithm 1; Table 1] The adaptive threshold mode uses a similarity threshold sθ (called θ in §3.1) to stop merging, and Table 1 reports the realized 8.0× compression under this mode. However, the actual threshold value(s) used in the experiments are never reported. The training schedule in Appendix B.2 samples thresholds, but the inference-time threshold for Table 1 is absent. Without this value, the headline result (8.0×, 65.3 average accuracy, 96.8% retention) cannot be reproduced. Please specify sθ (per benchmark, if it varies) or describe the procedure used to obtain the 8.0× operating point.
  4. [§4.2 / Table 1, Block A] The headline comparison in Block A pits CRAFT—trained on ~242K video-caption/QA pairs—against five training-free baselines. The abstract's broader claim of 'consistently outperforms prior state-of-the-art token-compression methods' is not supported by this comparison alone. The paper does include learnable baselines in Block B and a useful decomposition in Fig. 4, but the main table should include the same-scaffold selection-only row (as in Fig. 4) or a similarly positioned trained-fusion baseline so the margin can be attributed to the learned components rather than to the training data. At minimum, qualify the state-of-the-art claim to the appropriate comparison class.
minor comments (5)
  1. [§3.2, Eq. (2)] The text says 'Here · denotes ℓ2 normalization' but the equation only shows X and W; please clarify which quantities are normalized (e.g., with an overbar on X or Esim).
  2. [Algorithm 1, line 8] 'drop chain-conflicting edges' is stated but not detailed in the pseudocode; the text explanation in §3.2 is clear but a brief comment or a reference would help.
  3. [Table 2] The column header 'FLOPs↑' is ambiguous; consider something like 'FLOPs reduction (×)' to make the direction explicit.
  4. [Fig. 3] The legend entries 'CRAFT (ours)' and 'CRAFT (ours) · ratio' are hard to distinguish; rename the second to 'CRAFT (fixed-ratio mode)' or similar.
  5. [Appendix B.2] The table lists many hyperparameters but omits τ and sθ; adding them here would resolve the reproducibility concerns raised above.

Circularity Check

0 steps flagged

No significant circularity: the retention figures are measured outcomes of an externally evaluated pipeline, not derived constants or fitted values.

full rationale

The paper's central claim—roughly 97% backbone accuracy retention at ~8× compression—is an empirical measurement on six external benchmarks, not a quantity derived from the method's definitions. The token selection is parameter-free (random-projection similarity, Eq. 2–3) and the fusion network is trained with a standard cross-entropy objective on frozen backbone outputs (§3.4), so the compressor is not fitted to the benchmarks on which it is evaluated; the training data are explicitly de-duplicated against the evaluation benchmarks (Appendix B.3). No load-bearing step reduces by construction to its inputs: Eq. (4) is a general gated fusion form whose degenerate cases correspond to averaging and pruning, and the claimed retention is not an algebraic consequence of that equation. The L2-norm orientation heuristic is an assumption, but it is not defined in terms of the paper's predictions; the ablation in §4.5 provides partial empirical support, even if the comparison does not fully isolate the orientation choice—that is a correctness/experimental-design concern, not circularity. There are no self-citations used as a uniqueness theorem or as justification for the core design, and no known result is merely renamed. The derivation chain is self-contained against external benchmarks, so the absence of circularity is the appropriate finding.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The central claim rests on standard ML training, approximate similarity search, and two untested domain assumptions: the L2-norm salience proxy and the distribution-preservation of linear token fusion. The method introduces no new physical or formal entities. The free parameters are the usual deep-learning hyperparameters, plus one unstated value (tau) and an implicitly tuned adaptive threshold.

free parameters (7)
  • softmax temperature tau (Eq. 6) = not stated
    Appears in the source-weight softmax but no value is given anywhere in the main text or appendices; it is a free tunable hyperparameter of the learned weighting.
  • similarity threshold for adaptive mode = not stated explicitly; training samples from {0.75,0.8,0.85,0.9,0.95}
    Controls when merging stops in adaptive mode; the evaluation threshold appears chosen so the realized ratio averages 8.0x across datasets, matching baselines.
  • random projection dimension dsim = 128
    Hand-chosen dimension of the low-rank random-projection space used for approximate cosine similarity.
  • low-rank projection rank r for Wq, Wk = 128
    Rank of the position-aware weighting projections in Eq. 5.
  • gating network hidden dimension dh = 512
    Intermediate dimension of the per-channel gate MLP in Eq. 7.
  • merge module parameters (weighting scorer + gate) = about 5M
    All neural network weights of the compressor, trained end-to-end on caption and QA data. The central claim depends on the success of this training.
  • LoRA rank and alpha (CRAFT+ only) = rank 32, alpha 64
    Used only for the gray-outlined CRAFT+ variant that adapts the LLM; not part of the main claim.
axioms (5)
  • domain assumption Globally similar tokens are redundant and can be merged without losing critical semantics
    The entire selection scaffold in Section 3.2 rests on this premise. It is validated only indirectly by final benchmark accuracy, and the paper acknowledges that aggressive merging can lose critical details.
  • domain assumption A larger L2 norm indicates more salient semantic information
    Stated in Section 3.2: 'based on the assumption that a larger norm indicates more salient semantic information'. This decides which token is retained in every merge group.
  • domain assumption Linear combinations of original tokens stay within the pre-trained LM input distribution
    Claimed in the abstract and Section 3.3 to justify avoiding re-alignment training. The paper does not verify this empirically; it is a premise for the main efficiency and accuracy story.
  • standard math Random projections preserve cosine similarity sufficiently for pairing (Johnson-Lindenstrauss)
    Used in Eqs. 2-3 to justify computing similarity in a fixed 128-dimensional random projection space. The JL lemma provides a standard probabilistic guarantee.
  • domain assumption The same compressor design transfers across backbones with only re-instantiation at different width
    Appendix A.2 states the merge module is re-instantiated and re-trained on each backbone with the same recipe. The paper frames this as 'transfer', assuming no architecture-specific tuning is needed. This is plausible but unproven beyond the three tested backbones.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CRAFT: Compression via Recursive Adaptive Fusion of Video Tokens for Vision-Language Models." pith.science (2026). https://pith.science/paper/W556NPO7

@misc{pith2026260801644,
  author       = {Pith},
  title        = {Pith review of: CRAFT: Compression via Recursive Adaptive Fusion of Video Tokens for Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W556NPO7}},
  note         = {Machine review of arXiv:2608.01644}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In video understanding, vision-language models (VLMs) must ingest massive numbers of visual tokens, causing the computational and memory cost of the prefill stage to rise sharply. Such visual sequences are highly redundant along the spatio-temporal dimension, yet a high compression ratio is often accompanied by the loss of critical details. Existing token-compression methods either employ heuristic, training-free compression with limited content adaptivity or introduce additional modules that require expensive alignment training, leaving the trade-off between efficiency and adaptivity unresolved. To alleviate this limitation, we propose CRAFT: Compression via Recursive Adaptive Fusion of Video Tokens. CRAFT recursively merges tokens by decoupling parameter-free token selection from learnable token fusion: global similarity determines which tokens to merge, while a position-aware weighting module and a content-adaptive channel-wise gate learn how to fuse them. The whole compression pipeline is query-agnostic. Because every retained token is a linear combination of the original tokens, CRAFT preserves their true spatio-temporal coordinates and stays aligned with the pre-trained language model's input distribution. Experiments on multiple representative video benchmarks show that CRAFT consistently outperforms prior state-of-the-art token-compression methods. At about $8\times$ compression, it retains roughly $97\%$ of the backbone's average accuracy and shows significant efficiency improvement.

Figures

Figures reproduced from arXiv: 2608.01644 by Jianjin Zhang, Jun Sun, Xiaohong Li, Xiaole Wang, Yafeng Deng, Yu Chen.

Figure 1
Figure 1. Figure 1: Left: six-benchmark accuracy comparison at a fixed 8× compression ratio, normalized per benchmark to the best score among the six methods shown (outer vertex = best). Right: benchmark-averaged accuracy; the dashed line marks the uncompressed backbone (67.4). take into account the structural characteristics of video, so merging only within local temporal windows or within a single frame cannot eliminate the… view at source ↗
Figure 2
Figure 2. Figure 2: CRAFT performs Training-Free Global Selection and Learnable Gated Merging iteratively. Note that in Training-Free [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Accuracy–compression trade-off on Qwen3.5-4B: [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Selection scaffold vs. learnable merging, at ∼8×. Scissor-Merge couples LLaVA-Scissor’s scaffold with our learnable merging module. Scissor is the original LLaVA￾Scissor, and CRAFT (selection only) is our scaffold with the learnable merging replaced by equal-weight averaging. All methods run in their threshold-driven adaptive mode, and all trained variants are stage-1-only checkpoints. hyperparameters as C… view at source ↗
Figure 5
Figure 5. Figure 5: Training-data composition of the two-stage curriculum (§3.4). [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    International Conference on Learning Representations (ICLR) , year =

    Token Merging: Your ViT But Faster , author =. International Conference on Learning Representations (ICLR) , year =. 2210.09461 , archivePrefix =

    Pith/arXiv arXiv

  2. [2]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =. 2106.02034 , archivePrefix =

    Pith/arXiv arXiv

  3. [3]

    International Conference on Learning Representations (ICLR) , year =

    Not All Patches are What You Need: Expediting Vision Transformers via Token Reorganizations , author =. International Conference on Learning Representations (ICLR) , year =. 2202.07800 , archivePrefix =

    Pith/arXiv arXiv

  4. [4]

    Johnson and Joram Lindenstrauss , journal =

    William B. Johnson and Joram Lindenstrauss , journal =. Extensions of

    work page

  5. [5]

    Database-friendly Random Projections:

    Dimitris Achlioptas , journal =. Database-friendly Random Projections:

    work page

  6. [6]

    European Conference on Computer Vision (ECCV) , year =

    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 (ECCV) , year =. 2403.06764 , archivePrefix =

    Pith/arXiv arXiv

  7. [7]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =. 2411.15024 , archivePrefix =

    Pith/arXiv arXiv

  8. [8]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    HoliTom: Holistic Token Merging for Fast Video Large Language Models , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =. 2505.21334 , archivePrefix =

    arXiv

  9. [9]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    VisionZip: Longer is Better but Not Necessary in Vision Language Models , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =. 2412.04467 , archivePrefix =

    arXiv

  10. [10]

    Findings of the Association for Computational Linguistics (ACL Findings) , year =

    PruneVid: Visual Token Pruning for Efficient Video Large Language Models , author =. Findings of the Association for Computational Linguistics (ACL Findings) , year =. 2412.16117 , archivePrefix =

    Pith/arXiv arXiv

  11. [11]

    International Conference on Machine Learning (ICML) , year =

    LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding , author =. International Conference on Machine Learning (ICML) , year =. 2410.17434 , archivePrefix =

    Pith/arXiv arXiv

  12. [12]

    International Conference on Machine Learning (ICML) , year =

    BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models , author =. International Conference on Machine Learning (ICML) , year =. 2301.12597 , archivePrefix =

    Pith/arXiv arXiv

  13. [13]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Flamingo: a Visual Language Model for Few-Shot Learning , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =. 2204.14198 , archivePrefix =

    Pith/arXiv arXiv

  14. [14]

    International Journal of Computer Vision (IJCV) , year =

    TokenPacker: Efficient Visual Projector for Multimodal LLM , author =. International Journal of Computer Vision (IJCV) , year =. 2407.02392 , archivePrefix =

    Pith/arXiv arXiv

  15. [15]

    PVC: Progressive Visual Token Compression for Unified Image and Video Processing in Large Vision-Language Models

    PVC: Progressive Visual Token Compression for Unified Image and Video Processing in Large Vision-Language Models , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =. 2412.09613 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    arXiv preprint arXiv:2507.02279 , year =

    LaCo: Efficient Layer-wise Compression of Visual Tokens for Multimodal Large Language Models , author =. arXiv preprint arXiv:2507.02279 , year =. 2507.02279 , archivePrefix =

    Pith/arXiv arXiv

  17. [17]

    arXiv preprint arXiv:2503.21307 , year =

    InternVL-X: Advancing and Accelerating InternVL Series with Efficient Visual Token Compression , author =. arXiv preprint arXiv:2503.21307 , year =. 2503.21307 , archivePrefix =

    Pith/arXiv arXiv

  18. [18]

    International Conference on Learning Representations (ICLR) , year =

    MARC: Memory-Augmented RL Token Compression for Efficient Video Understanding , author =. International Conference on Learning Representations (ICLR) , year =. 2510.07915 , archivePrefix =

    arXiv

  19. [19]

    Transactions on Machine Learning Research (TMLR) , year =

    When Tokens Talk Too Much: A Survey of Multimodal Long-Context Token Compression across Images, Videos, and Audios , author =. Transactions on Machine Learning Research (TMLR) , year =. 2507.20198 , archivePrefix =

    arXiv

  20. [20]

    arXiv preprint arXiv:2505.18227 , year =

    Token Reduction Should Go Beyond Efficiency in Generative Models -- From Vision, Language to Multimodality , author =. arXiv preprint arXiv:2505.18227 , year =. 2505.18227 , archivePrefix =

    arXiv

  21. [21]

    Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track , year =

    EgoSchema: A Diagnostic Benchmark for Very Long-form Video Language Understanding , author =. Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track , year =. 2308.09126 , archivePrefix =

    Pith/arXiv arXiv

  22. [22]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    MVBench: A Comprehensive Multi-modal Video Understanding Benchmark , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =. 2311.17005 , archivePrefix =

    Pith/arXiv arXiv

  23. [23]

    Findings of the Association for Computational Linguistics (ACL Findings) , year =

    TempCompass: Do Video LLMs Really Understand Videos? , author =. Findings of the Association for Computational Linguistics (ACL Findings) , year =. 2403.00476 , archivePrefix =

    Pith/arXiv arXiv

  24. [24]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =. 2405.21075 , archivePrefix =

    Pith/arXiv arXiv

  25. [25]

    Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track , year =

    LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding , author =. Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track , year =. 2407.15754 , archivePrefix =

    Pith/arXiv arXiv

  26. [26]

    arXiv preprint arXiv:2409.12191 , year =

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution , author =. arXiv preprint arXiv:2409.12191 , year =. 2409.12191 , archivePrefix =

    Pith/arXiv arXiv

  27. [27]

    arXiv preprint arXiv:2502.13923 , year =

    Qwen2.5-VL Technical Report , author =. arXiv preprint arXiv:2502.13923 , year =. 2502.13923 , archivePrefix =

    Pith/arXiv arXiv

  28. [28]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Video-XL: Extra-Long Vision Language Model for Hour-Scale Video Understanding , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =. 2409.14485 , archivePrefix =

    Pith/arXiv arXiv

  29. [29]

    International Conference on Learning Representations (ICLR) , year =

    LongVILA: Scaling Long-Context Visual Language Models for Long Videos , author =. International Conference on Learning Representations (ICLR) , year =. 2408.10188 , archivePrefix =

    Pith/arXiv arXiv

  30. [30]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    MLVU: Benchmarking Multi-task Long Video Understanding , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =. 2406.04264 , archivePrefix =

    Pith/arXiv arXiv

  31. [31]

    arXiv preprint arXiv:2510.16598 , year =

    VisionSelector: End-to-End Learnable Visual Token Compression for Efficient Multimodal LLMs , author =. arXiv preprint arXiv:2510.16598 , year =. 2510.16598 , archivePrefix =

    arXiv

  32. [32]

    Proceedings of the 33rd ACM International Conference on Multimedia (ACM MM) , year =

    TimeChat-Online: 80\ author =. Proceedings of the 33rd ACM International Conference on Multimedia (ACM MM) , year =. 2504.17343 , archivePrefix =

    Pith/arXiv arXiv

  33. [33]

    International Conference on Learning Representations (ICLR) , year =

    LoRA: Low-Rank Adaptation of Large Language Models , author =. International Conference on Learning Representations (ICLR) , year =. 2106.09685 , archivePrefix =

    Pith/arXiv arXiv

  34. [34]

    IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Scalable Diffusion Models with Transformers , author =. IEEE/CVF International Conference on Computer Vision (ICCV) , year =. 2212.09748 , archivePrefix =

    Pith/arXiv arXiv

  35. [35]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Masked Autoencoders Are Scalable Vision Learners , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =. 2111.06377 , archivePrefix =

    Pith/arXiv arXiv

  36. [36]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =. 2203.12602 , archivePrefix =

    Pith/arXiv arXiv

  37. [37]

    International Conference on Machine Learning (ICML) , year =

    SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference , author =. International Conference on Machine Learning (ICML) , year =. 2410.04417 , archivePrefix =

    Pith/arXiv arXiv

  38. [38]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    PyramidDrop: Accelerating Your Large Vision-Language Models via Pyramid Visual Redundancy Reduction , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =. 2410.17247 , archivePrefix =

    Pith/arXiv arXiv

  39. [39]

    arXiv preprint arXiv:2506.03990 , year =

    DynTok: Dynamic Compression of Visual Tokens for Efficient and Effective Video Understanding , author =. arXiv preprint arXiv:2506.03990 , year =. 2506.03990 , archivePrefix =

    Pith/arXiv arXiv

  40. [40]

    Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    Video Compression Commander: Plug-and-Play Inference Acceleration for Video Large Language Models , author =. Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =. 2505.14454 , archivePrefix =

    arXiv

  41. [41]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =. 2503.02175 , archivePrefix =

    Pith/arXiv arXiv

  42. [42]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    FastVID: Dynamic Density Pruning for Fast Video Large Language Models , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =. 2503.11187 , archivePrefix =

    arXiv

  43. [43]

    IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Multi-Granular Spatio-Temporal Token Merging for Training-Free Acceleration of Video LLMs , author =. IEEE/CVF International Conference on Computer Vision (ICCV) , year =. 2507.07990 , archivePrefix =

    Pith/arXiv arXiv

  44. [44]

    arXiv preprint arXiv:2506.21862 , year =

    LLaVA-Scissor: Token Compression with Semantic Connected Components for Video LLMs , author =. arXiv preprint arXiv:2506.21862 , year =. 2506.21862 , archivePrefix =

    Pith/arXiv arXiv

  45. [45]

    IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Dynamic-VLM: Simple Dynamic Visual Token Compression for VideoLLM , author =. IEEE/CVF International Conference on Computer Vision (ICCV) , year =. 2412.09530 , archivePrefix =

    Pith/arXiv arXiv

  46. [46]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =. 2512.06866 , archivePrefix =

    arXiv

  47. [47]

    2024 , eprint=

    LLaVA-OneVision: Easy Visual Task Transfer , author=. 2024 , eprint=

    2024

  48. [48]

    2026 , eprint=

    Qwen3.5-Omni Technical Report , author=. 2026 , eprint=

    2026

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.