Pith. sign in

REVIEW 3 major objections 5 minor 20 references

Deleting 88.9% of a vision-language model's visual tokens can cost only 4.4 points of accuracy if saliency is diffused over a token graph, coverage is enforced per region, and dropped tokens are summarized rather than discarded.

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 · deepseek-v4-flash

2026-08-02 07:09 UTC pith:RYODIUAX

load-bearing objection Solid, well-engineered token condensation paper with a real fairness bug in the token-budget accounting that needs fixing before the SOTA claims can be trusted. the 3 major comments →

arxiv 2607.10640 v2 pith:RYODIUAX submitted 2026-07-12 cs.CV

Spectral Heat Flow for Conservative Token Condensation in Vision-Language Models

Zhaoyang Li , Yanjun Li , Wangkai Li , Yujia Chen , Tianzhu Zhang This is my paper
classification cs.CV
keywords visual token pruningtoken condensationvision-language modelsspectral heat flowkNN graph diffusionquadtree partitioningcoreset sinkstraining-free inference acceleration
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.

The paper is trying to establish that the standard speed-up recipe for vision-language models — keep the visual tokens with the highest attention scores, discard the rest — is the wrong framing, and that a 'conservative condensation' behaves much better at extreme pruning ratios. SpecFlow builds a k-nearest-neighbor graph over visual-token embeddings, diffuses a saliency seed (the attention of the encoder's global summary token) over that graph to obtain a smooth region-coherent importance field, allocates the token budget across adaptively partitioned image regions in proportion to their energy, and compresses the pruned tokens into two summary sink tokens instead of deleting them outright. On nine image-understanding benchmarks with LLaVA-1.5 the paper reports retaining 95.6% of unpruned accuracy while removing 88.9% of visual tokens, the best average among the compared training-free pruning methods at every compression level tested. A sympathetic reader cares because visual tokens, not text, drive the quadratic prefill cost of these models, so a plug-in method that removes most of them without retraining and without accuracy collapse changes what it costs to deploy them.

Core claim

SpecFlow's claim is that token importance for pruning should be a smoothed, structure-aware field, not a per-token score. It seeds saliency from the vision encoder's [CLS] attention (the global summary token) and propagates it by restart diffusion on a kNN graph built from token-feature cosine similarity — deliberately not raw self-attention, which the authors argue carries anomaly-token artifacts and region-inconsistent links. Proposition 3.1 shows the diffused field is the unique minimizer of a convex objective trading seed fidelity against graph smoothness, so the smoothing is principled rather than heuristic. Pruning then applies quadtree partitioning with energy-proportional quotas so e

What carries the argument

The load-bearing object is a sparse k-nearest-neighbor graph over visual-token embeddings with a row-stochastic softmax transition matrix (Eq. 5). The paper runs 'restart diffusion' (Eq. 2) on it, seeded by CLS attention: each token's energy is repeatedly mixed with its graph neighbors and pulled back toward the seed. The fixed point is the unique minimizer of a convex objective balancing seed fidelity against Dirichlet energy on the graph (Prop. 3.1), so the resulting spectral-heat-flow importance field is the smoothest field that stays faithful to the CLS cue. Two secondary mechanisms carry the conservation claim: quadtree splitting with energy-proportional quotas — a proportional-fairness

Load-bearing premise

The load-bearing premise is that the kNN graph built on token-feature similarity is a faithful structural substrate; if token features do not track semantic or spatial regions — anomaly tokens, feature homogenization — the diffused energy propagates misleading saliency into every region, and neither the coverage budget nor the sink tokens can recover the lost signal.

What would settle it

Reproduce the paper's own operator ablation: diffuse on raw self-attention instead of the kNN graph at 88.9% pruning and measure GQA and TextVQA; the paper reports a roughly 5-point gap (about 55.3 vs 49.6 on GQA, 54.9 vs 50.3 on TextVQA). If the two operators perform nearly identically, the structural-coherence story collapses. A second, sharper test: feed images with heavy repetitive texture, where token-feature neighborhoods are semantically flat, and check whether SpecFlow's high-compression accuracy falls back to plain CLS-attention Top-K.

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

If this is right

  • Prefill cost, which scales quadratically with visual-token count, drops superlinearly: at pruning ratio R the saving approaches 2R − R² when attention dominates — about 99% of the quadratic attention term at 88.9% pruning.
  • Aggressive budgets stop being catastrophic: the paper reports about 95.6% of unpruned accuracy at 88.9% compression on LLaVA-1.5, the best average among compared baselines, with graceful degradation as the budget shrinks.
  • The method is drop-in: no fine-tuning, plug-and-play with FlashAttention, and a pruning overhead measured at roughly 1% of end-to-end latency.
  • Structure-aware condensation transfers across architectures and modalities in the paper's experiments: LLaVA-NeXT at 88.9% pruning, video QA with 455 of 2,048 tokens kept, and Qwen2.5-VL all show the same pattern.

Where Pith is reading between the lines

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

  • Editorial inference: the mechanism predicts that SpecFlow's advantage over Top-K pruning grows as attention gets spikier, since diffusion and coverage budgeting specifically counteract spike concentration; a test that sharpens the CLS seed (fewer heads, higher temperature) should widen the gap.
  • Editorial inference: the method's premise cuts the other way on feature-homogenized inputs — heavy repetitive texture, dense small objects, or uniform backgrounds — where kNN neighborhoods carry little region structure; the paper's own limitation appendix names fine-grained evidence (small objects, dense text, counting) as the risk zone where mean-plus-residual sinks may be too thin.
  • Editorial inference: the measured decoding constant suggests that after aggressive prefill pruning the next bottleneck is decoding, which shrinks only linearly with token count; further end-to-end gains would need KV-cache or decode-side compression, which this paper does not address.
  • Editorial inference: the recipe — diffuse importance, enforce coverage, sink the rest — is generic for any transformer consuming spatially arranged tokens, so it could port to high-resolution OCR or long-video inputs, with the caveat that coverage partitioning would need to become spatiotemporal.

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

3 major / 5 minor

Summary. The paper proposes SpecFlow, a training-free visual token condensation method for vision-language models. It constructs a kNN graph over visual token embeddings, diffuses CLS-attention-derived importance via restart diffusion (spectral heat flow), selects tokens with a quadtree-based energy-proportional allocation to enforce spatial coverage, and appends two 'sink' tokens (mean and residual of pruned tokens) to preserve summary statistics. The method is evaluated on nine image understanding benchmarks with LLaVA-1.5, on LLaVA-NeXT, on two video QA benchmarks, and on Qwen2.5-VL, claiming state-of-the-art average relative performance at 66.7%, 77.8%, and 88.9% token reductions. The appendix provides proofs of diffusion contraction/mass preservation, the Dirichlet interpretation, proportional allocation optimality, and sink-token properties, along with wall-clock, memory, and FLOPs measurements.

Significance. If the empirical claims hold, SpecFlow provides a useful training-free alternative to existing token pruning methods, with a clear conceptual shift from point-wise ranking to graph-diffused importance plus spatial coverage. The mathematical lemmas in Appendices A.1–A.3 (contraction, mass preservation, Dirichlet view, proportional fairness, sink optimality) are standard but correctly proven and clearly presented. The paper also includes welcome engineering details: wall-clock breakdown, GPU memory, and FLOPs comparisons. However, the empirical comparison is not yet fully convincing because of a token-budget accounting inconsistency, a mismatched table caption, and the absence of code, error bars, or significance tests. The central 'outperforms SOTA' claim is proportionate only if the budget comparison is fair and the reported margins are statistically supported.

major comments (3)
  1. [§4.3 (Eq. 16) and Tables 1–4, 7] The token-budget accounting is internally inconsistent. Eq. (16) appends u_mean and u_res to the selected set, so the decoder processes K+2 tokens, yet the tables and abstract report budgets of 192/128/64 and pruning ratios 66.7%/77.8%/88.9% as if exactly K tokens were fed. At K=64, the true pruning ratio is 1−66/576 ≈ 88.5%, not 88.9%, and attention FLOPs scale by (66/64)^2 ≈ 6%. The reported SOTA margins at 88.9% are small (0.7 pp vs. HoloV in Table 1; 0.2 pp in Table 2), so the comparison may not be fair. Appendix E concedes that budget accounting is not yet standardized. The authors must count the sinks in the budget or give the same sinks to all baselines and recompute ratios, relative performance, and FLOPs.
  2. [Table 2 and §5.2] Table 2's caption says 'Video QA benchmarks', but the table reports LLaVA-NeXT image-understanding results on GQA, MMB, etc., with an upper bound of 2880 tokens. This mismatch obscures which experiments support the video claim; the actual video results are in Table 3. Correct the caption and align §5.2 with the table contents.
  3. [§5.1–5.2 and Tables 1–4] The empirical claims are reported without error bars, significance tests, or multiple seeds. The headline claim of 'outperforming SOTA' rests on average margins as small as 0.2–0.7 percentage points (e.g., 98.7% vs 98.5% in Table 1; 95.8% vs 95.6% in Table 2). Without variance estimates or a reproducible code release, the reader cannot distinguish real improvement from noise. Provide per-seed results, significance tests, and a working code link (the URL is not verifiable from the preprint).
minor comments (5)
  1. [§5.3] The text says 'As illustrated in Figure 3' for the diffusion-operator ablation, but Figure 3 is the pruning-strategy ablation; the diffusion ablation is Figure 2. Please fix the cross-reference.
  2. [Table 1] The average column for ToMe at 192 tokens reports 88.5% despite missing MMB-CN and VizWiz entries (shown as '–'). The averaging procedure over missing entries should be stated, or the missing values should be obtained and reported for consistency.
  3. [Table 4] The Average row mixes MME (a raw score, 2304) with percentage metrics (86.1, 84.7, 84.8) and then reports 100.0%. This normalization is unclear; MME scores are typically hundreds, not percentages. Please clarify the normalization and state it explicitly in the caption.
  4. [Appendix B.2] Only the number of diffusion steps (2) and minimum crop size (4) are given. The diffusion strength α, kNN temperature τ, neighbor count k, CLS head set H, quadtree variance threshold δ, and sink-token count are listed as free parameters but their values and sensitivity are not reported. Please provide a full hyperparameter table and, ideally, a sensitivity analysis.
  5. [Throughout] Several typos and formatting issues appear: 'LLaV A' with a space instead of 'LLaVA' (e.g., in Tables and text), 'V on Luxburg' should be 'Von Luxburg', and the abstract/table text uses inconsistent spacing. These should be corrected.

Circularity Check

0 steps flagged

No significant circularity; SpecFlow's benchmark outcomes are empirical and not encoded in its construction.

full rationale

SpecFlow's claimed outputs are benchmark accuracies. The selection procedure (kNN graph Eq. 5, diffusion Eq. 2, quadtree allocation Eqs. 8-14, sink construction Eq. 15) is parameterized by hyperparameters (k, tau, alpha, T_diff, delta, m) and by the seed e^(0)=CLS attention; no parameter is fitted to the nine benchmark scores in Tables 1-4. The method is training-free, and the reported numbers come from external evaluation, so the 'predictions' are not equivalent to the construction by definition. The mathematical claims (Prop. 3.1: diffusion equals Dirichlet-regularized smoothing; Prop. 4.1: energy-proportional allocation equals Kelly fairness; Lemmas A.5-A.6: mean sink optimality and residual radius) are standard results that the paper proves rather than assumes, and they do not carry the empirical claim. The self-citations (Li et al. 2025b,c) appear only as examples of recent VLM/vision work in the introduction and are not load-bearing for the pruning mechanism. The skeptic's point about the two appended sink tokens making the effective decoder sequence K+2 at K=64 is a legitimate fairness/accounting issue for the SOTA comparison, and Appendix E explicitly notes the authors 'plan to standardize token-budget accounting (including any appended sinks)'; however, this affects whether the reported pruning ratio and FLOPs are comparable to baselines, not whether any derivation reduces to its inputs. Thus no circular step meets the quoted-reduction standard required for a positive circularity finding.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 1 invented entities

The method mostly relies on standard math and well-established components. The main burden is on hyperparameters chosen by hand without sensitivity analysis and on the domain assumption that CLIP token geometry supports kNN-based diffusion.

free parameters (8)
  • diffusion strength α
    Controls propagation in Eq. (2); value not reported in main text; no sensitivity analysis.
  • kNN softmax temperature τ
    Used in Eq. (5) to convert cosine similarities to transition probabilities; value not specified.
  • kNN neighbor count k
    Defines graph sparsity in Eq. (5); value not specified.
  • CLS attention head set H
    Seed energy averages over a chosen subset of heads (Alg. 1); head count/indices not specified.
  • quadtree variance threshold δ
    Determines split criterion in Eq. (8); value not specified.
  • minimum crop size m = 4
    Set to 4 in Appendix B.2; controls finest quadtree granularity.
  • diffusion steps T_diff = 2
    Set to 2 in Appendix B.2; approximation to the fixed-point solution.
  • sink token count = 2
    Fixed to one mean plus one residual sink; chosen by hand without ablating the count.
axioms (6)
  • domain assumption A kNN graph built from CLIP token feature similarity reflects semantic region structure.
    Central to HeatFlow; if features do not align with regions, diffusion propagates misleading energy. Invoked in §3.2 and §4.1.
  • domain assumption CLS attention is a reliable seed of visual token importance.
    Used as initial energy in Eq. (6); grounded in CLIP design but not proven for all inputs.
  • standard math Energy-proportional budget allocation is an optimal fairness objective.
    Proven in Appendix A.2 via Kelly's concave program; mathematically sound.
  • standard math Restart diffusion converges to a unique fixed point and preserves mass.
    Proven in Appendix A.1 using contraction and row-stochasticity; correct.
  • domain assumption Mean + residual sink tokens sufficiently preserve pruned information.
    The paper acknowledges this may fail for fine-grained evidence (Appendix E); no strong guarantee.
  • domain assumption Benchmark suites measure the intended VLM capabilities.
    Standard practice; no independent verification of benchmark validity.
invented entities (1)
  • Coreset sink tokens (mean and residual) no independent evidence
    purpose: Compress all pruned visual tokens into two additional tokens to preserve context and diversity.
    The sink tokens have no falsifiable handle outside the paper's own benchmarks; their adequacy is only evaluated indirectly through downstream accuracy.

pith-pipeline@v1.3.0-alltime-deepseek · 229 in / 7275 out tokens · 124755 ms · 2026-08-02T07:09:22.102482+00:00 · methodology

0 comments
read the original abstract

Vision-Language Models (VLMs) are costly at inference time because they must process long sequences of visual tokens. Existing token pruning methods often degrade under high compression by blindly discarding information, breaking spatial structure or collapsing diversity. We propose SpecFlow, a training-free framework that shifts the paradigm from destructive pruning to conservative condensation, strictly enforcing spatial coverage and statistical conservation to ensure stability. Treating visual tokens as nodes in a $k$NN graph, SpecFlow (i) computes a stable importance field via spectral heat flow to preserve structural coherence, (ii) allocates budgets via adaptive spatial partitioning to guarantee coverage, and (iii) aggregates discarded information into coreset sinks to maintain statistical conservation. The method is plug-and-play, requires no fine-tuning, and is compatible with FlashAttention. Experiments confirm that our SpecFlow outperforms SOTA methods across tasks, VLM architectures, and pruning ratios. Notably, LLaVA-1.5 with SpecFlow retains 95.6% of original performance despite pruning 88.9% of visual tokens, offering an exceptional efficiency-accuracy balance. Code is available at https://github.com/Lzy-dot/SpecFlow

Figures

Figures reproduced from arXiv: 2607.10640 by Tianzhu Zhang, Wangkai Li, Yanjun Li, Yujia Chen, Zhaoyang Li.

Figure 1
Figure 1. Figure 1: Global Top-K pruning vs. our proposed SpecFlow. (a) Global Top-K pruning based on [CLS] attention can yield fragmented selections due to spiky attention distributions, often cre￾ating spatial holes within objects. (b) SpecFlow diffuses attention￾derived energy on a kNN token graph and applies coverage-aware regional budgeting, yielding region-coherent retained tokens under high compression. et al., 2023; Z… view at source ↗
Figure 2
Figure 2. Figure 2: Ablation study on diffusion operators. Our KNN graph (built from token feature similarity) outperforms raw self-attention (Token attn) and [CLS]-initialized attention propagation ([CLS] attn) [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation on pruning strategies. Our Quadtree outper￾forms Top-K (direct selection of top-K tokens) and Uni Crop (uniformly partitioned crops), validating adaptive spatial pruning. 5.3. Ablation Studies Ablation on Diffusion Operator. We validate the design choice of constructing a KNN graph based on token feature similarity by comparing it against two variants: raw visual token self-attention (Token attn) … view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative visualization of token selection and adaptive splitting. (a) Visualization of our energy-based quadtree splitting, showing the induced partitions and the final retained-token layouts across representative cases (b) Comparison of retained visual tokens at different pruning ratios for SpecFlow versus FastV and HoloV [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Cumulative mass (sorted) over visual token proportion for different im￾portance measures. propagate importance to each other, which mitigates the frag￾mentation of coherent objects under aggressive pruning and prevents a small set of peak tokens from dominating the budget. Meanwhile, the dif￾fused energy remains suffi￾ciently discriminative, avoid￾ing the overly smooth behavior that text-vision attention c… view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of raw [CLS] attention and HeatFlow-diffused energy. Left: Raw attention exhibits a spiky distribution. Right: Diffused energy is smoothed and covers coherent semantic regions. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗

Review history (2 revisions) →

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

20 extracted references · 12 linked inside Pith

  1. [1]

    Therefore, ∥T(e)− T(e′)∥1 =α∥W ⊤(e−e ′)∥1 ≤α∥W ⊤∥1∥e−e ′∥1 =α∥e−e ′∥1, proving Eq. (23). By Banach’s fixed-point theorem (Kreyszig, 1991), T has a unique fixed point e⋆. The fixed-point equationT(e ⋆) =e ⋆ is equivalent to Eq. (24). The linear convergence bound follows directly from contraction. A.2 Mass preservation (justifies optional renormalization). ...

    1991

  2. [8]

    X., and Wen, J.-R

    Li, Y ., Du, Y ., Zhou, K., Wang, J., Zhao, W. X., and Wen, J.-R. Evaluating object hallucination in large vision- language models.arXiv:2305.10355,

    Pith/arXiv arXiv

  3. [10]

    M., Hauth, A., Millican, K., et al

    Team, G., Anil, R., Borgeaud, S., Alayrac, J.-B., Yu, J., Sori- cut, R., Schalkwyk, J., Dai, A. M., Hauth, A., Millican, K., et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,

    Pith/arXiv arXiv

  4. [12]

    Pyramid- drop: Accelerating your large vision-language models via pyramid visual redundancy reduction.arXiv preprint arXiv:2410.17247,

    Xing, L., Huang, Q., Dong, X., Lu, J., Zhang, P., Zang, Y ., Cao, Y ., He, C., Wang, J., Wu, F., et al. Pyramid- drop: Accelerating your large vision-language models via pyramid visual redundancy reduction.arXiv preprint arXiv:2410.17247,

    Pith/arXiv arXiv

  5. [13]

    Sparsevlm: Visual token sparsification for effi- cient vision-language model inference.arXiv preprint arXiv:2410.04417, 2024b

    Zhang, Y ., Fan, C.-K., Ma, J., Zheng, W., Huang, T., Cheng, K., Gudovskiy, D., Okuno, T., Nakata, Y ., Keutzer, K., et al. Sparsevlm: Visual token sparsification for effi- cient vision-language model inference.arXiv preprint arXiv:2410.04417, 2024b. Zhao, H. H., Zhou, P., Gao, D., Bai, Z., and Shou, M. Z. Lova3: Learning to visual question answering, ask...

    Pith/arXiv arXiv

  6. [15]

    Additional Theoretical Details A.1

    12 Spectral Heat Flow for Conservative Token Condensation in Vision-Language Models A. Additional Theoretical Details A.1. HeatFlow Diffusion: Fixed Point, Mass Preservation, and Dirichlet View This appendix provides proofs and auxiliary lemmas for Proposition 3.1. Throughout, W∈R N×N is the nonnegative row-stochastic transition matrix in Eq. (5) (i.e.,W1...

    2012

  7. [17]

    A.3 Dirichlet view under symmetrized affinity (proof of Prop

    Since it holds att= 0, it holds for allt. A.3 Dirichlet view under symmetrized affinity (proof of Prop. 3.1).Proposition 3.1 additionally assumes that W is obtained by row-normalizing a symmetric affinity matrix. Concretely, defineS∈R N×N andD= diag(S1)such that S=S ⊤ ≥0, D ii >0, W=D −1S.(26) This holds, for instance, when we use a symmetrized KNN edge s...

    2004

  8. [18]

    Finally, sinceW ⊤ = (D−1S)⊤ =SD −1, we obtain (I−αW ⊤)e= (1−α)e (0), which matches the fixed-point equation

    Substitutingf=D −1eandf 0 =D −1e(0) gives (D−αS)D −1e= (1−α)e (0) ⇐ ⇒(I−αSD −1)e= (1−α)e (0). Finally, sinceW ⊤ = (D−1S)⊤ =SD −1, we obtain (I−αW ⊤)e= (1−α)e (0), which matches the fixed-point equation. By uniqueness of the minimizer, the optimizer ofJ(e)is exactlye ⋆. A.2. Quota Allocation and Integer Rounding for Quadtree Pruning This appendix provides ...

    1998

  9. [19]

    descendingM(c)

    Then the resulting integer quotas satisfy X c∈C qc =K,0≤q c ≤ |c| ∀c∈ C.(30) In particular, the “descendingM(c)” filling strategy in Sec. 4.2 is feasible. Proof. By construction, 0≤q (0) c ≤ |c|and P c q(0) c ≤ P c ˆqc =K , hence R≥0 . The total remaining capacity after the initial step is X c (|c| −q(0) c )≥ X c |c| − X c q(0) c =N− X c q(0) c ≥K− X c q(...

    2019

  10. [20]

    Beyond the raw images, the dataset provides object-level information such as locations and attributes

    GQA is a compositional visual question answering benchmark built around three core elements: images, structured scene annotations (scene graphs), and questions. Beyond the raw images, the dataset provides object-level information such as locations and attributes. Its questions are designed to probe not only visual recognition but also relational and compo...

    2023

  11. [416]

    Wen, Z., Gao, Y ., Wang, S., Zhang, J., Zhang, Q., Li, W., He, C., and Zhang, L

    Springer, 2024b. Wen, Z., Gao, Y ., Wang, S., Zhang, J., Zhang, Q., Li, W., He, C., and Zhang, L. Stop looking for important tokens in multimodal language models: Duplication matters more. arXiv preprint arXiv:2502.11494,

    Pith/arXiv arXiv

  12. [1998]

    Videopoet: A large language model for zero- shot video generation.arXiv preprint arXiv:2312.14125,

    Kondratyuk, D., Yu, L., Gu, X., Lezama, J., Huang, J., Schindler, G., Hornung, R., Birodkar, V ., Yan, J., Chiu, M.-C., et al. Videopoet: A large language model for zero- shot video generation.arXiv preprint arXiv:2312.14125,

    Pith/arXiv arXiv

  13. [2003]

    Don’t just chase” highlighted tokens” in mllms: Revisiting visual holistic context reten- tion.arXiv preprint arXiv:2510.02912,

    Zou, X., Lu, D., Wang, Y ., Yan, Y ., Lyu, Y ., Zheng, X., Zhang, L., and Hu, X. Don’t just chase” highlighted tokens” in mllms: Revisiting visual holistic context reten- tion.arXiv preprint arXiv:2510.02912,

    arXiv

  14. [2006]

    Flashattention-2: Faster attention with bet- ter parallelism and work partitioning.arXiv preprint arXiv:2307.08691,

    Dao, T. Flashattention-2: Faster attention with bet- ter parallelism and work partitioning.arXiv preprint arXiv:2307.08691,

    Pith/arXiv arXiv

  15. [2010]

    Token merging: Your vit but faster.arXiv preprint arXiv:2210.09461,

    Bolya, D., Fu, C.-Y ., Dai, X., Zhang, P., Feichtenhofer, C., and Hoffman, J. Token merging: Your vit but faster.arXiv preprint arXiv:2210.09461,

    Pith/arXiv arXiv

  16. [2011]

    MME: A comprehen- sive evaluation benchmark for multimodal large language models.arXiv:2306.13394,

    Fu, C., Chen, P., Shen, Y ., Qin, Y ., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al. MME: A comprehen- sive evaluation benchmark for multimodal large language models.arXiv:2306.13394,

    Pith/arXiv arXiv

  17. [2019]

    D., Bouadjenek, M

    Huynh, N. D., Bouadjenek, M. R., Aryal, S., Razzak, I., and Hacid, H. Visual question answering: from early developments to recent advances–a survey.arXiv preprint arXiv:2501.03939,

    Pith/arXiv arXiv

  18. [2022]

    An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929,

    Dosovitskiy, A. An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929,

    Pith/arXiv arXiv 2010

  19. [2024]

    Mini-gemini: Mining the potential of multi-modality vision language models.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 2025a

    Li, Y ., Zhang, Y ., Wang, C., Zhong, Z., Chen, Y ., Chu, R., Liu, S., and Jia, J. Mini-gemini: Mining the potential of multi-modality vision language models.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 2025a. Li, Z., Qian, D., Su, K., Diao, Q., Xia, X., Liu, C., Yang, W., Zhang, T., and Yuan, Z. Bindweave: Subject-consistent video ge...

    arXiv 2024

  20. [2025]

    Qwen-vl: A frontier large vision- language model with versatile abilities.arXiv preprint arXiv:2308.12966,

    Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., and Zhou, J. Qwen-vl: A frontier large vision- language model with versatile abilities.arXiv preprint arXiv:2308.12966,

    Pith/arXiv arXiv