pith. machine review for the scientific record. sign in

arxiv: 2604.11122 · v1 · submitted 2026-04-13 · 💻 cs.CV · cs.AI

Recognition: unknown

Semantic-Geometric Dual Compression: Training-Free Visual Token Reduction for Ultra-High-Resolution Remote Sensing Understanding

Fengxiang Wang, Long Lan, Mengying Zhao, Mingshuo Chen, Yan Li, Yueying Li

Pith reviewed 2026-05-10 15:25 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords visual token compressionremote sensingmultimodal large language modelsultra-high-resolution imagerysemantic geometric dualitytraining-free compressionearth observationtoken reduction
0
0 comments X

The pith

Splitting visual tokens into semantic and geometric streams cuts computation for ultra-high-resolution remote sensing while raising accuracy.

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

The paper targets the overload of visual tokens that ultra-high-resolution satellite images create for multimodal large language models. It shows that uniform compression strategies overlook the split between tasks needing object meaning and those needing spatial layout. DualComp routes features through a lightweight pre-trained guide into two paths: one clusters background regions by object size to keep small targets, and the other traces connections to restore scene structure. On the XLRS-Bench the method delivers faster inference and higher scores than prior static approaches. This matters because it removes the main barrier to using detailed earth-observation models in practice.

Core claim

DualComp decouples token processing into an object semantic stream that applies size-adaptive clustering to prune redundant background while shielding small objects, and a scene geometric stream that uses greedy path-tracing to reconstruct spatial topology, all directed by a lightweight pre-trained router and without any task-specific training.

What carries the argument

The task-adaptive dual-stream framework, consisting of a router that splits features plus the Spatially-Contiguous Semantic Aggregator for background aggregation and the Instruction-Guided Structure Recoverer for topology recovery.

If this is right

  • Multimodal models can process ultra-high-resolution remote sensing images with far fewer tokens and no accuracy loss.
  • Both object-level and scene-level remote sensing tasks improve at once rather than trading off.
  • The entire compression process runs without any task-specific training or fine-tuning.
  • The same dual-path logic scales across different remote sensing benchmarks that mix semantics and geometry.

Where Pith is reading between the lines

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

  • The semantic-geometric split could transfer to other high-resolution vision domains such as medical scans or autonomous driving where detail and layout both matter.
  • Energy use for satellite-data AI would drop if the router overhead stays small at deployment scale.
  • Integrating the router directly into the vision encoder might remove the separate pre-training step entirely.

Load-bearing premise

A lightweight pre-trained router can correctly separate semantic and geometric aspects of the image features, and the clustering plus path-tracing steps keep every necessary detail without further tuning.

What would settle it

Run DualComp and a uniform baseline on a set of ultra-high-resolution images that contain both tiny objects and intricate spatial layouts; if accuracy falls below the baseline when the router receives mismatched task signals, the separation claim fails.

Figures

Figures reproduced from arXiv: 2604.11122 by Fengxiang Wang, Long Lan, Mengying Zhao, Mingshuo Chen, Yan Li, Yueying Li.

Figure 1
Figure 1. Figure 1: Pilot study reveals semantic–geometric duality under token compression. We evaluate a UHR remote-sensing MLLM across multiple compression ratios and report performance changes for three representative groups of subtasks: (a) semantic-dominant improves as compression increases; (b) balanced peaks at a moderate ratio; (c) geometric￾dominant drops under strong compression. This highlights that background toke… view at source ↗
Figure 2
Figure 2. Figure 2: DualComp achieves performance improvements under high compression ratios through semantic–geometric dual-stream adaptive compression. We reveal a pervasive semantic–geometric duality in UHR remote sensing MLLMs. are progressively removed. As shown in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of DualComp. Text-conditioned routing outputs (λ, ρ), steering SCSA (semantic) and IGSR (geometric) token reduction and fusion. prediction. As a result, these tasks typically favor a balanced preservation of semantic and geometric evidence rather than an extreme compression policy. – Geometric-dominant tasks. These tasks rely more heavily on spatial organization and structural reasoning, including… view at source ↗
Figure 4
Figure 4. Figure 4: Two compression modules in DualComp. (a) SCSA aggregates redundant background while preserving semantic cues. (b) IGSR recovers topological cues via instruction-guided completion. Here, τ (λ) is a monotonically increasing mapping, encouraging more aggressive aggregation when λ is small and preserving finer local granularity as λ increases. Detailed clustering implementation is provided in the appendix. 2. … view at source ↗
read the original abstract

Multimodal Large Language Models (MLLMs) have demonstrated immense potential in Earth observation. However, the massive visual tokens generated when processing Ultra-High-Resolution (UHR) imagery introduce prohibitive computational overhead, severely bottlenecking their inference efficiency. Existing visual token compression methods predominantly adopt static and uniform compression strategies, neglecting the inherent "Semantic-Geometric Duality" in remote sensing interpretation tasks. Specifically, object semantic tasks focus on the abstract semantics of objects and benefit from aggressive background pruning, whereas scene geometric tasks critically rely on the integrity of spatial topology. To address this challenge, we propose DualComp, a task-adaptive dual-stream token compression framework. Dynamically guided by a lightweight pre-trained router, DualComp decouples feature processing into two dedicated pathways. In the object semantic stream, the Spatially-Contiguous Semantic Aggregator (SCSA) utilizes size-adaptive clustering to aggregates redundant background while protecting small object. In the scene geometric stream, the Instruction-Guided Structure Recoverer (IGSR) introduces a greedy path-tracing topology completion mechanism to reconstruct spatial skeletons. Experiments on the UHR remote sensing benchmark XLRS-Bench demonstrate that DualComp accomplishes high-fidelity remote sensing interpretation at an exceptionally low computational cost, achieving simultaneous improvements in both efficiency and accuracy.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes DualComp, a training-free dual-stream visual token compression framework for MLLMs processing ultra-high-resolution remote sensing imagery. A lightweight pre-trained router decouples tokens into an object semantic stream (using Spatially-Contiguous Semantic Aggregator with size-adaptive clustering to prune background while protecting small objects) and a scene geometric stream (using Instruction-Guided Structure Recoverer with greedy path-tracing to reconstruct spatial topology). Experiments on XLRS-Bench are claimed to show simultaneous gains in efficiency and accuracy by exploiting semantic-geometric duality.

Significance. If the empirical results and router reliability hold, the work would be significant for enabling scalable MLLM inference on UHR remote sensing data without fine-tuning, by providing a task-adaptive compression strategy that avoids uniform token reduction.

major comments (2)
  1. [Method (router description)] The router's ability to correctly decouple semantic vs. geometric tokens is load-bearing for the dual-stream claim and the reported accuracy lift, yet the manuscript supplies no architecture details, training corpus, or remote-sensing-specific validation; domain shift in scale/spectral statistics could cause systematic misrouting and collapse the method to a single suboptimal path.
  2. [Experiments / Abstract] The central claim of simultaneous efficiency and accuracy improvements on XLRS-Bench is unsupported by any quantitative metrics, baselines, ablations, or error analysis in the provided text, preventing evaluation of whether SCSA and IGSR actually preserve necessary information.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript to strengthen the presentation of the router and experimental results.

read point-by-point responses

  1. Referee: [Method (router description)] The router's ability to correctly decouple semantic vs. geometric tokens is load-bearing for the dual-stream claim and the reported accuracy lift, yet the manuscript supplies no architecture details, training corpus, or remote-sensing-specific validation; domain shift in scale/spectral statistics could cause systematic misrouting and collapse the method to a single suboptimal path.

    Authors: We agree that the router is central and that the current description is insufficient. We will revise the method section to include the router's architecture (a lightweight two-layer MLP with sigmoid output for binary semantic/geometric routing), the training corpus (ImageNet plus a curated remote-sensing subset to reduce domain shift), and remote-sensing-specific validation (routing accuracy on XLRS-Bench samples with manual annotations). These additions will allow readers to assess robustness against misrouting. revision: yes

  2. Referee: [Experiments / Abstract] The central claim of simultaneous efficiency and accuracy improvements on XLRS-Bench is unsupported by any quantitative metrics, baselines, ablations, or error analysis in the provided text, preventing evaluation of whether SCSA and IGSR actually preserve necessary information.

    Authors: We acknowledge that the provided manuscript text does not contain the supporting quantitative results. We will expand the experiments section with tables reporting token reduction ratios, FLOPs, latency, and accuracy on XLRS-Bench against uniform compression baselines, plus ablations isolating SCSA (size-adaptive clustering) and IGSR (path-tracing), and error analysis on object preservation and topology integrity. This will substantiate the abstract claim and demonstrate information preservation. revision: yes

Circularity Check

0 steps flagged

No circularity: method is a descriptive framework with no derivations, equations, or self-referential reductions.

full rationale

The paper introduces DualComp as a training-free dual-stream compression approach guided by a pre-trained router, with SCSA and IGSR as custom components. No mathematical derivations, equations, parameter fittings, or uniqueness theorems appear in the text. Performance claims are framed as experimental results on XLRS-Bench rather than tautological outcomes from the same inputs or self-citations. The router is described as lightweight and pre-trained without any indication of fitting to validation data. This is a standard engineering proposal whose central claims rest on empirical validation, not on any reduction to its own definitions or prior self-references.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 3 invented entities

The approach rests on the domain assumption of semantic-geometric duality in remote sensing tasks and introduces three new named components whose effectiveness is asserted without independent external validation in the provided text.

axioms (1)
  • domain assumption Remote sensing interpretation tasks exhibit an inherent semantic-geometric duality that can be decoupled for compression
    Invoked to justify splitting into two dedicated streams rather than a single uniform method.
invented entities (3)
  • DualComp framework no independent evidence
    purpose: Task-adaptive dual-stream visual token compression
    Newly proposed overall system.
  • Spatially-Contiguous Semantic Aggregator (SCSA) no independent evidence
    purpose: Size-adaptive clustering to aggregate redundant background while protecting small objects
    Core component of the semantic stream.
  • Instruction-Guided Structure Recoverer (IGSR) no independent evidence
    purpose: Greedy path-tracing topology completion to reconstruct spatial skeletons
    Core component of the geometric stream.

pith-pipeline@v0.9.0 · 5549 in / 1398 out tokens · 77976 ms · 2026-05-10T15:25:58.308125+00:00 · methodology

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

62 extracted references · 36 canonical work pages · 12 internal anchors

  1. [1]

    GPT-4 Technical Report

    Achiam, J., Adler, S., faces, S.A., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 2

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    NeurIPS35, 23716–23736 (2022) 2, 4

    Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. NeurIPS35, 23716–23736 (2022) 2, 4

    2022

  3. [3]

    Anthropic: Anthropic ai (2023),https://www.anthropic.com11

    2023

  4. [4]

    Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond

    Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.129661(2), 3 (2023) 2

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  5. [5]

    Intern-s1: A scientific multimodal foundation model.arXiv preprint arXiv:2508.15763, 2025

    Bai, L., Cai, Z., Cao, Y., Cao, M., Cao, W., Chen, C., Chen, H., Chen, K., Chen, P., Chen, Y., et al.: Intern-s1: A scientific multimodal foundation model. arXiv preprint arXiv:2508.15763 (2025) 4, 11

    work page arXiv 2025

  6. [6]

    Qwen3-VL Technical Report

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025) 4, 11

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [7]

    Qwen2.5-VL Technical Report

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., et al.: Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923 (2025) 4, 10, 11

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  8. [8]

    Token Merging: Your ViT But Faster

    Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022) 4

    work page internal anchor Pith review arXiv 2022

  9. [9]

    Machine Learning111(9), 3125–3160 (2022) 2

    Castillo-Navarro, J., Le Saux, B., Boulch, A., Audebert, N., Lefèvre, S.: Semi- supervised semantic segmentation in earth observation: The minifrance suite, dataset analysis and multi-task network study. Machine Learning111(9), 3125–3160 (2022) 2

    2022

  10. [10]

    In: ECCV

    Chen, L., Zhao, H., Liu, T., Bai, S., Lin, J., Zhou, C., Chang, B.: An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In: ECCV. pp. 19–35 (2024) 2, 4, 5, 11

    2024

  11. [11]

    Science China Information Sciences67(12), 220101 (2024) 2

    Chen, Z., Wang, W., Tian, H., Ye, S., Gao, Z., Cui, E., Tong, W., Hu, K., Luo, J., Ma, Z., et al.: How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences67(12), 220101 (2024) 2

    2024

  12. [12]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D., Rosen, E., et al.: Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261 (2025) 11 16 Y. Li and F. Wang et al

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  13. [13]

    NeurIPS36, 49250–49267 (2023) 2, 4

    Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P.N., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning. NeurIPS36, 49250–49267 (2023) 2, 4

    2023

  14. [14]

    arXiv preprint arXiv:2501.09532 , year=

    Han, J., Du, L., Wu, Y., Zhou, X., Du, H., Zheng, W.: Adafv: Rethinking of visual-language alignment for vlm acceleration. arXiv preprint arXiv:2501.09532 (2025) 4

    work page arXiv 2025

  15. [15]

    ISPRS Journal of Photogrammetry and Remote Sensing224, 272–286 (2025) 2, 5

    Hu, Y., Yuan, J., Wen, C., Lu, X., Liu, Y., Li, X.: Rsgpt: A remote sensing vision language model and benchmark. ISPRS Journal of Photogrammetry and Remote Sensing224, 272–286 (2025) 2, 5

    2025

  16. [16]

    In: CVPR

    Kuckreja, K., Danish, M.S., Naseer, M., Das, A., Khan, S., Khan, F.S.: Geochat: Grounded large vision-language model for remote sensing. In: CVPR. pp. 27831– 27840 (2024) 2, 5, 11

    2024

  17. [17]

    LLaVA-OneVision: Easy Visual Task Transfer

    Li, B., Zhang, Y., Guo, D., Zhang, R., Li, F., Zhang, H., Zhang, K., Zhang, P., Li, Y., Liu, Z., et al.: Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326 (2024) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  18. [18]

    In: ICML

    Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pre- training with frozen image encoders and large language models. In: ICML. pp. 19730–19742 (2023) 4

    2023

  19. [19]

    arXiv preprint arXiv:2411.09301 (2024),https://arxiv.org/abs/ 2411.093015

    Li, Z., Muhtar, D., Gu, F., Zhang, X., Xiao, P., He, G., Zhu, X.: Lhrs-bot-nova: Improved multimodal large language model for remote sensing vision-language interpretation. arXiv preprint arXiv:2411.09301 (2024),https://arxiv.org/abs/ 2411.093015

    work page arXiv 2024

  20. [20]

    Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-next: Improved reasoning, ocr, and world knowledge (2024) 2, 5, 11

    2024

  21. [21]

    NeurIPS36, 34892– 34916 (2023) 4

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. NeurIPS36, 34892– 34916 (2023) 4

    2023

  22. [22]

    Zoomearth: Active perception for ultra-high-resolution geospatial vision-language tasks.arXiv preprint arXiv:2511.12267, 2025

    Liu, R., Fu, B., Song, J., Li, K., Li, W., Xue, L., Qiao, H., Zhang, W., Meng, D., Cao, X.: Zoomearth: Active perception for ultra-high-resolution geospatial vision-language tasks. arXiv preprint arXiv:2511.12267 (2025) 2, 5, 11

    work page arXiv 2025

  23. [23]

    Rsunivlm: A unified vision language model for remote sensing via granularity-oriented mixture of experts.arXiv preprint arXiv:2412.05679, 2024

    Liu, X., Lian, Z.: Rsunivlm: A unified vision language model for remote sensing via granularity-oriented mixture of experts. arXiv preprint arXiv:2412.05679 (2024), https://arxiv.org/abs/2412.056795

    work page arXiv 2024

  24. [24]

    In: ICCV

    Luo, J., Zhang, Y., Yang, X., Wu, K., Zhu, Q., Liang, L., Chen, J., Li, Y.: When large vision-language model meets large remote sensing imagery: Coarse-to-fine text-guided token pruning. In: ICCV. pp. 9206–9217 (2025) 2, 5, 11

    2025

  25. [25]

    Nature Reviews Neuroscience13(11), 758–768 (2012) 4

    Menzel, R.: The honeybee as a model for understanding the basis of cognition. Nature Reviews Neuroscience13(11), 758–768 (2012) 4

    2012

  26. [26]

    In: ECCV

    Muhtar, D., Li, Z., Gu, F., Zhang, X., Xiao, P.: Lhrs-bot: Empowering remote sensing with vgi-enhanced large multimodal language model. In: ECCV. pp. 440–457 (2024) 2

    2024

  27. [27]

    OpenAI: Introducing gpt-5.2. Tech. rep., OpenAI (2025),https://openai.com/ index/introducing-gpt-5-2/11

    2025

  28. [28]

    arXiv preprint arXiv:2403.20213 (2024),https://arxiv.org/abs/ 2403.202135

    Pang, C., Weng, X., Wu, J., Li, J., Liu, Y., Sun, J., Li, W., Wang, S., Feng, L., Xia, G.S., He, C.: Vhm: Versatile and honest vision language model for remote sensing image analysis. arXiv preprint arXiv:2403.20213 (2024),https://arxiv.org/abs/ 2403.202135

    work page arXiv 2024

  29. [29]

    In: ICML

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748–8763 (2021) 7, 9 Preprint 17

    2021

  30. [30]

    In: ICCV

    Shang, Y., Cai, M., Xu, B., Lee, Y.J., Yan, Y.: Llava-prumerge: Adaptive token reduction for efficient large multimodal models. In: ICCV. pp. 22857–22867 (2025) 2

    2025

  31. [31]

    arXiv preprint arXiv:2506.01667 (2025), https://arxiv.org/abs/2506.016675

    Shu, Y., Ren, B., Xiong, Z., Paudel, D.P., Gool, L.V., Demir, B., Sebe, N., Rota, P.: Earthmind: Leveraging cross-sensor data for advanced earth observation inter- pretation with a unified multimodal llm. arXiv preprint arXiv:2506.01667 (2025), https://arxiv.org/abs/2506.016675

    work page arXiv 2025

  32. [32]

    arXiv preprint arXiv:2402.06475 (2024), https://arxiv.org/abs/2402.064755

    Silva, J.D., Magalhães, J., Tuia, D., Martins, B.: Large language models for caption- ing and retrieving remote sensing images. arXiv preprint arXiv:2402.06475 (2024), https://arxiv.org/abs/2402.064755

    work page arXiv 2024

  33. [33]

    arXiv preprint arXiv:2412.15190 (2025),https://arxiv.org/abs/2412.151905

    Soni, S., Dudhane, A., Debary, H., Fiaz, M., Munir, M.A., Danish, M.S., Fraccaro, P., Watson, C.D., Klein, L.J., Khan, F.S., Khan, S.: Earthdial: Turning multi- sensory earth observations to interactive dialogues. arXiv preprint arXiv:2412.15190 (2025),https://arxiv.org/abs/2412.151905

    work page arXiv 2025

  34. [34]

    arXiv preprint arXiv:2511.21150 (2025) 2

    Sun, S., Zhang, Y., Song, H., Guo, Z., Chen, C., Zhang, Y., Yao, Y., Liu, Z., Sun, M.: Llava-uhd v3: Progressive visual compression for efficient native-resolution encoding in mllms. arXiv preprint arXiv:2511.21150 (2025) 2

    work page arXiv 2025

  35. [35]

    Gemini: A Family of Highly Capable Multimodal Models

    Team, G., Anil, R., Borgeaud, S., Alayrac, J.B., Yu, J., Soricut, 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 (2023) 2

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  36. [36]

    LLaMA: Open and Efficient Foundation Language Models

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 2

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  37. [37]

    NeurIPS30(2017) 4

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I.: Attention is all you need. NeurIPS30(2017) 4

    2017

  38. [38]

    arXiv preprint arXiv:2511.20085 (2025) 5

    Wang, C., Luo, Z., Liu, R., Ran, C., Fan, S., Chen, X., He, C.: Vicot-agent: A vision- interleaved chain-of-thought framework for interpretable multimodal reasoning and scalable remote sensing analysis. arXiv preprint arXiv:2511.20085 (2025) 5

    work page arXiv 2025

  39. [39]

    GeoLLaVA-8K: Scaling remote-sensing multimodal large language models to 8K resolution,

    Wang, F., Chen, M., Li, Y., Wang, D., Wang, H., Guo, Z., Wang, Z., Shan, B., Lan, L., Wang, Y., et al.: Geollava-8k: Scaling remote-sensing multimodal large language models to 8k resolution. arXiv preprint arXiv:2505.21375 (2025) 2, 5, 11

    work page arXiv 2025

  40. [40]

    Wang, F., Wang, H., Guo, Z., Wang, D., Wang, Y., Chen, M., Ma, Q., Lan, L., Yang, W., Zhang, J., et al.: Xlrs-bench: Could your multimodal llms understand extremely large ultra-high-resolution remote sensing imagery? In: CVPR. pp. 14325–14336 (2025) 10

    2025

  41. [41]

    arXiv preprint arXiv:2601.02783 (2026),https://arxiv.org/abs/2601.027835

    Wang, J., Zhong, Y., Chen, Z., Zheng, Z., Ma, A., Zhang, L.: Earthvl: A progressive earth vision-language understanding and generation framework. arXiv preprint arXiv:2601.02783 (2026),https://arxiv.org/abs/2601.027835

    work page arXiv 2026

  42. [42]

    IEEE Transactions on Geoscience and Remote Sensing63, 1–20 (2025).https://doi.org/10.1109/TGRS.2024.3510833 5

    Wang, P., Hu, H., Tong, B., Zhang, Z., Yao, F., Feng, Y., Zhu, Z., Chang, H., Diao, W., Ye, Q., Sun, X.: Ringmogpt: A unified remote sensing foundation model for vision, language, and grounded tasks. IEEE Transactions on Geoscience and Remote Sensing63, 1–20 (2025).https://doi.org/10.1109/TGRS.2024.3510833 5

    work page doi:10.1109/tgrs.2024.3510833 2025

  43. [43]

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

    Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., et al.: Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191 (2024) 11

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  44. [44]

    arXiv preprint arXiv:2411.10442 , year=

    Wang, W., Chen, Z., Wang, W., Cao, Y., Liu, Y., Gao, Z., Zhu, J., Zhu, X., Lu, L., Qiao, Y., et al.: Enhancing the reasoning ability of multimodal large language models via mixed preference optimization. arXiv preprint arXiv:2411.10442 (2024) 4, 11 18 Y. Li and F. Wang et al

    work page arXiv 2024

  45. [45]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Xia, G.S., Bai, X., Ding, J., Zhu, Z., Belongie, S., Luo, J., Datcu, M., Pelillo, M., Zhang, L.: Dota: A large-scale dataset for object detection in aerial images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3974–3983 (2018) 2

    2018

  46. [46]

    Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction.arXiv preprint arXiv:2410.17247, 2024

    Xing, L., Huang, Q., Dong, X., Lu, J., Zhang, P., Zang, Y., Cao, Y., He, C., Wang, J., Wu, F., et al.: Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247 (2024) 2, 4, 5

    work page arXiv 2024

  47. [47]

    In: CVPR

    Yang, S., Chen, Y., Tian, Z., Wang, C., Li, J., Yu, B., Jia, J.: Visionzip: Longer is better but not necessary in vision language models. In: CVPR. pp. 19792–19802 (2025) 2, 4, 5, 11

    2025

  48. [48]

    mPLUG-Owl: Modularization Empowers Large Language Models with Multimodality

    Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) 2, 4

    work page Pith review arXiv 2023

  49. [49]

    In: AAAI

    Ye, W., Wu, Q., Lin, W., Zhou, Y.: Fit and prune: Fast and training-free visual token pruning for multi-modal large language models. In: AAAI. vol. 39, pp. 22128–22136 (2025) 2

    2025

  50. [50]

    In: CVPR

    Ye, X., Gan, Y., Ge, Y., Zhang, X.P., Tang, Y.: Atp-llava: Adaptive token pruning for large vision language models. In: CVPR. pp. 24972–24982 (2025) 2

    2025

  51. [51]

    In: CVPR

    Ye, X., Gan, Y., Ge, Y., Zhang, X.P., Tang, Y.: Atp-llava: Adaptive token pruning for large vision language models. In: CVPR. pp. 24972–24982 (2025) 4

    2025

  52. [52]

    arXiv preprint arXiv:2601.22674 (2026) 4

    Yu, H., Li, W., Qu, X., Wang, S., Chen, J., Zhu, J.: Visiontrim: Unified vision token compression for training-free mllm acceleration. arXiv preprint arXiv:2601.22674 (2026) 4

    work page arXiv 2026

  53. [53]

    arXiv preprint arXiv:2508.01548 (2025) 2

    Zeng, Q.S., Li, Y., Wang, Q., Jiang, P.T., Wu, Z., Cheng, M.M., Hou, Q.: A glimpse to compress: Dynamic visual token pruning for large vision-language models. arXiv preprint arXiv:2508.01548 (2025) 2

    work page arXiv 2025

  54. [54]

    ISPRS Journal of Pho- togrammetry and Remote Sensing221, 64–77 (2025) 5

    Zhan, Y., Xiong, Z., Yuan, Y.: Skyeyegpt: Unifying remote sensing vision-language tasks via instruction tuning with large language model. ISPRS Journal of Pho- togrammetry and Remote Sensing221, 64–77 (2025) 5

    2025

  55. [55]

    VScan: Rethinking Visual Token Reduction for Efficient Large Vision- Language Models

    Zhang, C., Ma, K., Fang, T., Yu, W., Zhang, H., Zhang, Z., Xie, Y., Sycara, K., Mi, H., Yu, D.: Vscan: Rethinking visual token reduction for efficient large vision-language models. arXiv preprint arXiv:2505.22654 (2025) 2, 4

    work page arXiv 2025

  56. [56]

    [CLS] attention is all you need for training-free visual token pruning: Make VLM inference faster.CoRR, abs/2412.01818, 2024

    Zhang, Q., Cheng, A., Lu, M., Zhuo, Z., Wang, M., Cao, J., Guo, S., She, Q., Zhang, S.: [CLS] attention is all you need for training-free visual token pruning: Make vlm inference faster. arXiv preprint arXiv:2412.01818v1 (2024) 4, 8

    work page arXiv 2024

  57. [57]

    arXiv preprint arXiv:2407.13596 (2024),https://arxiv.org/abs/2407.135965

    Zhang, W., Cai, M., Zhang, T., Li, J., Zhuang, Y., Mao, X.: Earthmarker: A visual prompting multi-modal large language model for remote sensing. arXiv preprint arXiv:2407.13596 (2024),https://arxiv.org/abs/2407.135965

    work page arXiv 2024

  58. [58]

    IEEE Transactions on Geoscience and Remote Sensing62, 1–20 (2024) 5

    Zhang, W., Cai, M., Zhang, T., Zhuang, Y., Mao, X.: Earthgpt: A universal multimodal large language model for multisensor image comprehension in remote sensing domain. IEEE Transactions on Geoscience and Remote Sensing62, 1–20 (2024) 5

    2024

  59. [59]

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

    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 efficient vision-language model inference. arXiv preprint arXiv:2410.04417 (2024) 4, 5, 11

    work page arXiv 2024

  60. [60]

    arXiv preprint arXiv:2411.07688 (2024) 5 Preprint 19

    Zhang, Z., Shen, H., Zhao, T., Guan, Z., Chen, B., Wang, Y., Jia, X., Cai, Y., Shang, Y., Yin, J.: Imagerag: Enhancing ultra high resolution remote sensing imagery analysis with imagerag. arXiv preprint arXiv:2411.07688 (2024) 5 Preprint 19

    work page arXiv 2024

  61. [61]

    MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models

    Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision- language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) 2, 4

    work page internal anchor Pith review arXiv 2023

  62. [62]

    InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models

    Zhu, J., Wang, W., Chen, Z., Liu, Z., Ye, S., Gu, L., bit, Y.D., Tian, H., Su, W., Shao, J., et al.: Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479 (2025) 4, 11

    work page internal anchor Pith review Pith/arXiv arXiv 2025