Pith. sign in

REVIEW 4 major objections 6 minor 23 references

MAViE: A Multi-scale Adaptive Vision Encoder for Fine-grained Visual Perception and Efficient Multimodal Reasoning

T0 review · 4 major / 6 minor · reviewed 2026-07-31 · grok-4.5

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

Pith's one-line read A multi-scale vision encoder can cut about 80% of visual tokens while improving fine-grained multimodal answers and lowering time to first token.

desk verdict Clean systems design for multi-scale fusion plus question-aware token routing, but every headline number is simulated and the ablations are too tidy to trust even as placeholders. read the letter →

arxiv 2607.24424 v1 pith:I4RJDGRH submitted 2026-07-27 cs.CV

Shaofei Lei This is my paper
classification cs.CV
keywords vision-languagemodelsvisualtokencompressionmulti-scalefeaturefusionquestion-conditionedroutingadaptivebudgetfine-grainedperceptionrepresentationdistillation
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

Vision-language models usually dump every final-layer patch token from a vision transformer into the language model. That both throws away shallow detail (edges, text strokes, local structure) and makes high-resolution images expensive. This paper proposes MAViE: fuse shallow, intermediate, and deep ViT features with position-dependent gates, then keep only a question- and complexity-dependent subset of tokens scored by relevance, local variation, global semantics, and spatial coverage, recycling the rest into a handful of context tokens. Distillation from the full representation and a spatial diversity penalty are meant to limit what compression erases. The authors argue this design can raise average accuracy on standard VQA-style benchmarks while shrinking the visual sequence from hundreds of tokens to roughly one-fifth and cutting prefill latency. The manuscript supplies the full architecture and evaluation protocol; the headline numbers are explicitly design placeholders pending real training.

What carries the argument

Question-conditioned Adaptive Token Routing (QATR) on Hierarchical Multi-level Fusion (HMF) features: each token is scored by question relevance, local information, global semantics, and spatial coverage; a complexity estimator sets K between 72 and 288; unselected tokens collapse into eight context tokens, trained with full-to-compressed distillation and a spatial diversity regularizer.

What would settle it

Train MAViE end-to-end under the stated 7B protocol and measure official VQAv2, GQA, TextVQA, ScienceQA-IMG, and MMBench scores plus average token count and TTFT against full SigLIP-729; if compressed MAViE does not match or beat the full baseline on fine-grained tasks at ~146 tokens, the central claim fails.

Watch

Extended reading notes

Core claim

MAViE claims that position-dependent hierarchical fusion of multi-level ViT features, followed by question-conditioned adaptive token routing with a dynamic budget and recycled context tokens, can preserve fine-grained visual evidence better than full final-layer token streams while reducing average visual tokens by about 80% and improving mean multimodal benchmark scores and time to first token under a unified 7B language-model setup.

Load-bearing premise

The four-part routing score, eight recycled context tokens, and adaptive budget actually keep the local evidence needed for OCR, counting, and multi-object relations better than sending all final-layer tokens.

Editorial extensions

If this is right

  • VLMs can raise input resolution without a proportional jump in language-model prefill cost if routing keeps only task-relevant patches.
  • OCR- and relation-heavy questions should benefit more than pure scene classification when shallow features and coverage terms are kept.
  • Dynamic budgets can shift tokens toward text-dense or multi-object images and away from simple single-subject scenes.
  • Full-to-compressed distillation plus diversity penalties become standard tools when aggressive visual pruning is used.
  • The same router can later support multi-image, video, or iterative observe-reason-observe re-encoding.

Reading between the lines

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

  • If routing is the real bottleneck, failures will cluster on low-salience distant objects and ambiguous pronouns—exactly the cases the limitations section flags—suggesting a cheap full-token fallback under low router confidence.
  • The design implies that fixed-query resamplers leave money on the table whenever the question changes which spatial evidence matters.
  • A hardware-aware budget that optimizes measured latency or energy rather than a hand-set K range is a direct next experiment.
  • Success here would pressure default VLM pipelines that still project every final-layer patch by default.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The manuscript proposes MAViE, a vision-encoder front end for VLMs combining (i) hierarchical gated fusion of SigLIP layers 8/16/24 with local window enhancement (HMF, §3.2), (ii) question-conditioned token routing with a four-term score (relevance, local variation, global similarity, spatial coverage) and a complexity-adaptive budget K∈[72,288] (QATR, §3.3), and (iii) full-to-compressed distillation plus a spatial-diversity regularizer (§3.4). The headline claim (Abstract; Table 1) is ~80% token reduction (729→146), +2.2pp mean accuracy over the full-token SigLIP baseline across five benchmarks, and 228→129 ms TTFT under a unified 7B setup. However, every quantitative table (Tables 1–8 and Appendix B) is explicitly labeled simulated placeholder ("not measurements from actual runs", §4.2), and the NeurIPS checklist answers [No] on claims, reproducibility, code, and significance testing. The paper is therefore a design document with an evaluation protocol, not an empirical result.

Significance. If the claimed operating point were real — 80% fewer visual tokens with a net accuracy gain and 43% lower TTFT — it would be of clear interest to the efficient-VLM community, and the problem setting (preserving OCR strokes, small objects, and relational referents under aggressive compression) is well motivated. To the paper's credit: the disclosure of simulated results is unusually candid and repeated in the Abstract, §4.2, §5, and Appendix B; the training/evaluation protocol (§4.1) is specified in enough detail that the experiments are executable; and the method section is a coherent, if incremental, combination of known components (multi-layer fusion, prompt-aware routing, distillation). There is no code, no checkpoints, and no falsifiable prediction beyond the protocol itself. As submitted, the manuscript contributes a design sketch; its significance is contingent entirely on experiments that do not yet exist.

major comments (4)
  1. [Abstract; §4.2; Tables 1–2] The central claim has no evidential support as submitted. Abstract, Table 1, and §4.2 report +2.2pp mean accuracy, 79.97% token reduction (Eq. 26), and 228→129 ms TTFT, but §4.2 states these 'are not measurements from actual runs' and Tables 1–8 are all simulated. There are no trained checkpoints, training curves, seeds, or error bars anywhere in the paper. A results-style submission whose entire evidence base is placeholder cannot be evaluated for correctness; the honest labeling mitigates deception but does not substitute for measurement. This is decisive on its own.
  2. [Tables 3, 6, 7; Appendix B] The simulated ablations are non-neutral and therefore mislead even as illustrations. In Table 3 every component (HMF, QATR, FD, SD) improves every benchmark in exactly the order presented; in Table 6 gated fusion strictly dominates mean fusion and every layer subset on all four benchmarks; in Table 7 each scoring term adds accuracy monotonically. Real ablations essentially never behave this way, so the tables encode the conclusion rather than test it. If placeholders are retained in a revision, they should be clearly separated from the claims or removed; as written they will shape reader belief in the +2.2pp figure.
  3. [Table 8; §1; §4.4] The paper's own illustrative numbers contradict its motivating premise and its headline mechanism. Table 8 shows fixed-72 tokens (80.3 VQAv2) already exceeding the full 729-token baseline (80.0), which undercuts the Introduction's claim that aggressive pruning discards task-critical local evidence; and fixed-288 (81.5 VQAv2) beats the proposed dynamic-146 configuration (81.4), so the dynamic budget — a core contribution — is not the best-performing option in the authors' own simulated trade-off. Either the real experiments are expected to differ (in which case the placeholders are actively misleading) or the method's central design choice is unmotivated.
  4. [§3.3, Eqs. (17)–(19); §3.4, Eq. (21)] The dynamic-budget pathway appears to receive no training signal. Eq. (19) defines K via Round(c·(Kmax−Kmin)), which is non-differentiable, and no term in the objective (Eq. 21) supervises the complexity estimator c (Eq. 18). The Gumbel-Top-K relaxation (§3.3) makes token selection differentiable for fixed K but does not propagate gradient through K itself. As specified, c can only drift via weight decay; the claimed adaptive allocation (Table 5) has no stated mechanism to emerge. Relatedly, the coverage term s_cover (Eq. 17) requires a sequentially growing set S, which is consistent with hard Top-K inference but not with simultaneous Gumbel-Top-K training; the train/inference selection protocol is underspecified, and s_cover is undefined for the first selection step (S empty).
minor comments (6)
  1. [Table 2; Figure 1] Table 2, last row: the MAViE entry is typeset as 'MAViE8721 2112913.3' — column alignment is broken. Figure 1's embedded text is similarly garbled ('ImageI384×384 SigLIP ViTZ8, Z16, Z24...').
  2. [Eqs. (13), (23)] Eq. (23): τ is used as a logit offset but never defined; z_ij is defined only after the equation. Eq. (13): clarify that q is pooled from the LLM's input embedding table before any LLM forward pass, since routing precedes prefill.
  3. [§3.2–3.4] Default hyperparameters are asserted without justification or sensitivity analysis: (λr, λu, λg, λc) = (0.45, 0.20, 0.15, 0.20) in §3.3, (α, β, η) = (0.10, 0.50, 0.05) in §3.4, M=8 context tokens, γ init 0.1, and layer indices {8,16,24}. Given that all experiments are simulated, these choices are currently untested premises and should be flagged as such.
  4. [Table 1; §2] Table 1 omits the most relevant training-free and efficiency baselines discussed in §2 — SparseVLM, PruMerge, FastVLM, LLaVA-UHD — comparing only against pooling/resampling/TokenPacker-style proxies. Any real evaluation should include prompt-aware sparsification methods at matched token counts.
  5. [§2] §2 contains extended subsections (continual segmentation, broader visual applications) only loosely connected to visual-token compression; roughly a page of citations could be trimmed without loss.
  6. [Eq. (20)] Eq. (20): the aggregation weights a_mi are described as similarity-based but the normalization (softmax over unselected tokens per query, or over queries per token?) is ambiguous; this matters for whether the 8 context tokens partition or overlap in coverage.

Circularity Check

0 steps flagged · score 0.0 of 10

No derivation circularity: MAViE is an engineering design with explicit placeholder metrics, not a self-sealing first-principles claim.

full rationale

The paper’s load-bearing content is a modular architecture (HMF gated fusion, QATR four-term routing, dynamic K, distillation, spatial diversity) plus planned external benchmarks (VQAv2, GQA, TextVQA, etc.). Equations (1)–(25) define modules and losses; they do not define a target quantity in terms of itself, nor fit a parameter on data and relabel a near-identical quantity as a prediction. Headline numbers (+2.2 pp, 729→146 tokens, 228→129 ms) are repeatedly labeled simulated placeholders for organization, not derived outputs. Related-work self-citations (remote-sensing multi-scale work) motivate design choices but are not uniqueness theorems that force the central claim. Absence of real measurements is a validity failure, not circularity of the derivation chain. Score 0; steps empty.

Assumptions & free parameters 6 free parameters · 4 assumptions · 3 invented entities

Empirical headline claims rest on an unrun training protocol plus many hand-chosen knobs (layer indices, score weights, budget bounds, loss coefficients). Domain axioms about ViT layer specialization and token redundancy are standard in the cited literature. Invented entities are the named modules and scoring terms, which lack independent external evidence until trained and tested.

free parameters (6)
  • score weights (λr, λu, λg, λc) = (0.45, 0.20, 0.15, 0.20)
    Default mixture of relevance/information/global/coverage terms that directly controls which tokens survive; chosen by hand in §3.3, not derived.
  • dynamic token budget bounds Kmin, Kmax = 72 and 288 (multiple of 8)
    Sets compression operating range and thus TTFT/accuracy trade-off; specified as 72 and 288 without data-driven selection procedure.
  • context token count M = 8
    Number of recycled aggregates for pruned tokens; fixed design choice affecting residual information.
  • loss coefficients α, β, η = (0.10, 0.50, 0.05)
    Weights on alignment, distillation, and diversity terms in Eq. (21); hand-set and load-bearing for compression stability claims.
  • fusion residual scale γ init and layer indices {8,16,24} = γ0=0.1; layers 8/16/24
    Which ViT depths count as shallow/mid/deep and how strongly fused features residual-add into deep features; architectural hyperparameters without ablation from real runs in the main claim path.
  • placeholder benchmark/TTFT table values = e.g. 146 tokens, +2.2 pp mean, 129 ms TTFT
    All reported accuracies, token averages, and latencies are simulated organization placeholders; the central quantitative claim currently depends on these non-measured numbers.
assumptions (4)
  • domain assumption Shallow/intermediate/deep ViT layers preferentially encode edges-text, parts-structure, and global semantics respectively, so fusing them improves OCR and local attributes over final-layer-only features.
    Stated in §1 and §3.1–3.2 as motivation for HMF; standard folklore but not proven here for SigLIP-SO400M under the planned recipe.
  • domain assumption Visual token sequences contain large task-conditional redundancy, so Top-K routing plus a few context aggregates can match or beat full-token projection into a 7B LM.
    Core efficiency premise (§1, TokenPacker citation); required for the 80% compression claim to be beneficial rather than purely lossy.
  • domain assumption A fixed 384×384 SigLIP-SO400M/14 patch grid (N=729) is an adequate base representation for the targeted VQA/OCR benchmarks when multi-scale fusion is added.
    Problem formulation §3.1; Limitations §5 admits failure on extremely small/blurred text under this axiom.
  • ad hoc to paper Gumbel-Top-K training and hard Top-K inference with the stated four-term score yield stable, question-sensitive selections transferable across the listed benchmarks.
    Training/inference rule in §3.3; no real optimization evidence yet.
invented entities (3)
  • Hierarchical Multi-level Fusion (HMF) with position-dependent gates and high-frequency window enhancement
    purpose: Produce enhanced per-token features F that keep shallow detail without wrecking deep semantics.
    Named module §3.2; mechanism is composed of known ops (DWConv, softmax gates, local attention) but packaged as the paper’s fusion block.
  • Question-conditioned Adaptive Token Routing (QATR) with complexity estimator c and coverage-updated sequential selection
    purpose: Select adaptive K tokens and recycle unselected tokens into M context tokens conditioned on Q and image complexity.
    Named module §3.3; central to the compression claim; no external falsification until real models exist.
  • Full-to-compressed representation distillation plus spatial diversity regularizer Ldiv
    purpose: Reduce information loss and regional collapse under pruning.
    Objectives in §3.4 Eqs. (24)–(25); standard loss forms applied to this router, not independently measured outside the draft.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MAViE: A Multi-scale Adaptive Vision Encoder for Fine-grained Visual Perception and Efficient Multimodal Reasoning." pith.science (2026). https://pith.science/paper/I4RJDGRH

@misc{pith2026260724424,
  author       = {Pith},
  title        = {Pith review of: MAViE: A Multi-scale Adaptive Vision Encoder for Fine-grained Visual Perception and Efficient Multimodal Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I4RJDGRH}},
  note         = {Machine review of arXiv:2607.24424}
}
read the original abstract

Vision-language models commonly project all tokens produced by a pretrained vision encoder into a large language model. However, final-layer features can discard text, local attributes, and spatial relationships, while high-resolution inputs substantially increase context length and inference latency. We introduce \method, a Multi-scale Adaptive Vision Encoder. \method uses position-dependent gates to fuse shallow, intermediate, and deep features from a vision Transformer, preserving global semantics while enhancing edges, text, and local structure. It then performs question-conditioned token routing according to question relevance, local information content, global semantics, and spatial coverage, with a token budget that adapts to image complexity. To mitigate compression loss, we further introduce full-to-compressed representation distillation and a spatial diversity regularizer. In an illustrative simulation under a unified 7B language-model framework, \method reduces the average number of SigLIP-SO400M visual tokens from 729 to 146 (approximately 80.0\%) and improves the mean score on VQAv2, GQA, TextVQA, ScienceQA-IMG, and MMBench by 2.2 percentage points, while reducing single-image time to first token from 228\,ms to 129\,ms. We provide the full model design and evaluation protocol. All reported numbers currently serve only as placeholders for paper organization and experimental design; formal claims require real training runs, independent replications, and official benchmark evaluation.

Figures

Figures reproduced from arXiv: 2607.24424 by the authors.

Figure 1
Figure 1. Overview of MAViE. Multi-level features are fused by [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

23 extracted references · 18 linked inside Pith

  1. [5]

    Global-local attention network for semantic segmentation in aerial images

    Ming Li, Lianlei Shan, Xiang Li, Yang Bai, Dong Zhou, Weiqiang Wang, Kai Lv, Bin Luo, and Shibao Chen. Global-local attention network for semantic segmentation in aerial images. In International Conference on Pattern Recognition, pages 5704–5711, 2021a. Ming Li, Lianlei Shan, Weiqiang Wang, Kai Lv, Bin Luo, and Shibao Chen. Building lightweight semantic s...

    arXiv

  2. [8]

    Yingen Liu, Fan Wu, Ruihui Li, Zhuo Tang, and Kenli Li

    doi: 10.1109/ICBAIE66852.2025.11326614. Yingen Liu, Fan Wu, Ruihui Li, Zhuo Tang, and Kenli Li. PAR: Prompt-aware token reduction method for efficient large multimodal models.arXiv preprint arXiv:2410.07278,

    arXiv 2025

  3. [9]

    MMBench: Is your multi-modal model an all-around player?arXiv preprint arXiv:2307.06281, 2023b

    Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. MMBench: Is your multi-modal model an all-around player?arXiv preprint arXiv:2307.06281, 2023b. Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hiera...

    arXiv

  4. [10]

    LLM-CoT enhanced graph neural recommendation with harmonized group policy optimization.arXiv preprint arXiv:2505.12396,

    Haoyu Luo, Bowen Wu, Haoran Jia, Qiang Zhu, and Lianlei Shan. LLM-CoT enhanced graph neural recommendation with harmonized group policy optimization.arXiv preprint arXiv:2505.12396,

    arXiv

  5. [11]

    Synthetic lung x-ray generation through cross-attention and affinity transformation.arXiv preprint arXiv:2503.07209,

    Rui Pi and Lianlei Shan. Synthetic lung x-ray generation through cross-attention and affinity transformation.arXiv preprint arXiv:2503.07209,

    arXiv

  6. [12]

    Boosting general trimap-free matting in the real-world image

    Lianlei Shan and Wenzhang Zhou. Boosting general trimap-free matting in the real-world image. arXiv preprint arXiv:2405.17916,

    arXiv

  7. [13]

    Boosting semantic segmentation of aerial images via decoupled and multilevel compaction and dispersion.IEEE Transactions on Geoscience and Remote Sensing, 61:1–16, 2023a

    Lianlei Shan, Weiqiang Wang, Kai Lv, and Bin Luo. Boosting semantic segmentation of aerial images via decoupled and multilevel compaction and dispersion.IEEE Transactions on Geoscience and Remote Sensing, 61:1–16, 2023a. Lianlei Shan, Guangyu Zhao, Jing Xie, Peng Cheng, Xiang Li, and Zhen Wang. A data-related patch proposal for semantic segmentation of ae...

    arXiv

  8. [14]

    LLaV A-PruMerge: Adaptive token reduction for efficient large multimodal models.arXiv preprint arXiv:2403.15388,

    Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. LLaV A-PruMerge: Adaptive token reduction for efficient large multimodal models.arXiv preprint arXiv:2403.15388,

    arXiv

Show all 23 references
  1. [16]

    Qwen2-VL: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191,

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-VL: Enhancing vision-language model’s pe...

    arXiv

  2. [17]

    GDGS: 3d gaussian splatting via geometry-guided initialization and dynamic density control.arXiv preprint arXiv:2507.00363,

    Xin Wang and Lianlei Shan. GDGS: 3d gaussian splatting via geometry-guided initialization and dynamic density control.arXiv preprint arXiv:2507.00363,

    arXiv

  3. [18]

    RecLLM-R1: A two-stage training paradigm with reinforcement learning and chain-of-thought.arXiv preprint arXiv:2506.19235,

    Yuxuan Xie, Xinyi Ren, Yiming Qi, Yifan Hu, and Lianlei Shan. RecLLM-R1: A two-stage training paradigm with reinforcement learning and chain-of-thought.arXiv preprint arXiv:2506.19235,

    arXiv

  4. [19]

    LLaV A-UHD: An LMM perceiving any aspect ratio and high-resolution images.arXiv preprint arXiv:2403.11703,

    Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. LLaV A-UHD: An LMM perceiving any aspect ratio and high-resolution images.arXiv preprint arXiv:2403.11703,

    arXiv

  5. [20]

    KV-Efficient VLA: A method to speed up vision language models with RNN-gated chunked KV cache.arXiv preprint arXiv:2509.21354,

    Wanshun Xu and Long Zhuang. KV-Efficient VLA: A method to speed up vision language models with RNN-gated chunked KV cache.arXiv preprint arXiv:2509.21354,

  6. [21]

    A global-local cross-attention network for ultra-high resolution remote sensing image semantic segmentation.arXiv preprint arXiv:2506.19406, 2025a

    Chen Yi and Lianlei Shan. A global-local cross-attention network for ultra-high resolution remote sensing image semantic segmentation.arXiv preprint arXiv:2506.19406, 2025a. Qian Yi and Lianlei Shan. GeolocSFT: Efficient visual geolocation via supervised fine-tuning of multimo...

    arXiv

  7. [22]

    Asymmetric Mamba–CNN collaborative architecture for large-size remote sensing image semantic segmentation.IEEE Transactions on Geoscience and Remote Sensing, 63:2002419, 2025a

    Jian Zhang, Ming Chen, Yifan Zhao, Lianlei Shan, Cheng Li, Hao Hu, Xin Ge, Qiang Zhu, and Bin Xu. Asymmetric Mamba–CNN collaborative architecture for large-size remote sensing image semantic segmentation.IEEE Transactions on Geoscience and Remote Sensing, 63:2002419, 2025a. do...

    2025

  8. [23]

    DynRsl-VLM: Enhancing autonomous driving perception with dynamic resolution vision-language models.arXiv preprint arXiv:2503.11265,

    Xirui Zhou, Lianlei Shan, and Xiaolin Gui. DynRsl-VLM: Enhancing autonomous driving perception with dynamic resolution vision-language models.arXiv preprint arXiv:2503.11265,

    arXiv

  9. [2017]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. InAdvances in Neural Information Processing Systems, volume 36, pages 34892–34916, 2023a. 9 Shuo Liu and Lianlei Shan. NeuroV oxel-LM: Language-aligned 3d perception via dynamic voxeliza- tion ...

    arXiv

  10. [2019]

    GMM-based comprehensive feature extraction and relative distance preservation for few-shot cross-modal retrieval.arXiv preprint arXiv:2505.13306,

    Chen Sun, Wei Li, Xiang Li, Yang Liu, and Lianlei Shan. GMM-based comprehensive feature extraction and relative distance preservation for few-shot cross-modal retrieval.arXiv preprint arXiv:2505.13306,

    arXiv

  11. [2021]

    Transform dual-branch attention net: Efficient semantic segmentation of ultra-high-resolution remote sensing images

    8 Bin Du, Lianlei Shan, Xin Shao, Dong Zhang, Xin Wang, and Jian Wu. Transform dual-branch attention net: Efficient semantic segmentation of ultra-high-resolution remote sensing images. Remote Sensing, 17(3):540, 2025a. Yi Du, Zhen Zeng, Yu Tian, Zhe Zhang, Xin Zhang, and Lian...

    arXiv

  12. [2022]

    Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond.arXiv preprint arXiv:2308.12966,

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond.arXiv preprint arXiv:2308.12966,

    arXiv

  13. [2023]

    The binary quantized neural network for dense prediction via specially designed upsampling and attention

    Xiaohan Ding, Lianlei Shan, Guangyu Zhao, Ming Wu, Wenzhang Zhou, and Wei Li. The binary quantized neural network for dense prediction via specially designed upsampling and attention. arXiv preprint arXiv:2405.17776,

    arXiv

  14. [2024]

    LLaV A-OneVision: Easy visual task transfer

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaV A-OneVision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024a. Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BL...

    arXiv

  15. [2025]

    To- kenPacker: Efficient visual projector for multimodal LLM.arXiv preprint arXiv:2407.02392, 2024b

    Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. To- kenPacker: Efficient visual projector for multimodal LLM.arXiv preprint arXiv:2407.02392, 2024b. Xiang Li, Lianlei Shan, Ming Li, and Weiqiang Wang. Energy minimum regularization in conti...

    arXiv

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.