Pith. sign in

REVIEW 2 major objections 6 minor 42 references

Attention from only the last query token is enough to prune visual tokens before they leave the edge, keeping most answer accuracy while sending a small fraction of the image.

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

2026-07-31 22:22 UTC pith:SHYIADY4

load-bearing objection Clean systems result: last-query attention alone is enough for strong pre-cloud visual pruning, with broad tables and a real but bounded proxy-alignment caveat. the 2 major comments →

arxiv 2607.27952 v1 pith:SHYIADY4 submitted 2026-07-30 cs.CV cs.AI

LAST: The Last Query Token Guides Visual Token Pruning for Edge-Cloud Collaborative MLLM Inference

Feng Yang , Xinrui Ju , Keyang Zhang , Xiandong Meng , Rongqun Lin , Howard Leung , Shiqi Wang , Haoliang Li
show 1 more author
Chris Xing Tian
This is my paper
classification cs.CV cs.AI
keywords visual token pruningedge-cloud collaborative inferencemultimodal large language modelsquery-conditioned attentionlast query tokentraining-free pruningtoken budget
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.

Edge devices that ship full visual token sequences to cloud multimodal models pay a heavy cost in communication and cloud compute, because most patches are redundant or irrelevant to the current question. Prior pruning either ignores the question, needs the cloud model’s internals, or makes the edge do expensive aggregation and answer generation. LAST claims that a small edge vision-language model, in one decoding-free prefill, can score every visual token using only how much the last query token attends to it under causal attention—because that position already sees the whole image and the whole question. Combined with an importance-weighted diversity selection under a fixed budget, the method keeps the tokens that matter and drops the rest before transmission. On eleven benchmarks it preserves the large majority of full-token accuracy at aggressive retention ratios, with low edge overhead and less cloud work.

Core claim

Under causal attention, the last query token’s attention over visual tokens is a sufficient, lightweight, query-conditioned importance signal for pre-transmission pruning: a compact frozen edge proxy can use that single-position score plus feature-space diversity to retain a small token subset that the cloud model can consume natively, without cloud internals, multi-query aggregation, or autoregressive guidance.

What carries the argument

Last-query-token attention scoring: aggregate softmax attention from the final query position across layers and heads onto visual positions, normalize to an importance distribution, then run importance-weighted greedy k-center selection so retained tokens are both query-relevant and non-redundant.

Load-bearing premise

A small edge model that shares (or matches) the vision encoder must rank visual tokens in a way that still guides the larger cloud model’s answers correctly, without ever reading the cloud model’s own attention or training a new adapter.

What would settle it

At 12.5% retention on the same edge–cloud pairs and benchmarks, if LAST’s average relative accuracy fell to or below strong query-dependent baselines (or collapsed on dense document/chart tasks when the proxy and cloud diverge), the claim that last-query attention alone is enough would fail.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Edge–cloud multimodal systems can cut visual payload and cloud prefill cost by an order of magnitude while staying near full-token accuracy, without placing the target MLLM on the device.
  • Query-aware pruning need not wait for cloud internals or a generated draft answer; a single prefill on a small proxy is enough for selection.
  • Shared or compatible vision-token interfaces let selected edge tokens feed the cloud connector directly, avoiding re-encoding and learned compression modules.
  • Aggressive budgets especially reward combining last-token relevance with diversity rather than pure top-k or query-agnostic saliency.

Where Pith is reading between the lines

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

  • If last-token attention generalizes across more model families, edge routers could treat visual token budgets like adaptive bitrate: raise or lower retention from the same proxy scores under changing network or latency targets.
  • The same single-position signal might extend to multi-turn chat by rescoring only when the latest user turn changes what visual evidence is needed.
  • Failures would likely cluster where the small proxy systematically misreads fine text or rare layouts the cloud would have used—suggesting hybrid keep-rules for OCR-heavy tiles rather than pure attention ranking.

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

2 major / 6 minor

Summary. LAST is a training-free framework for query-dependent visual-token pruning before transmission in edge–cloud MLLM inference. A compact edge-side VLM runs a single decoding-free prefill; under causal attention, attention from the last query token to visual positions is aggregated across layers and heads as a query-conditioned importance signal (Eqs. 7–8). An importance-weighted greedy k-center rule then retains a diverse subset under a fixed budget K=⌈ρN⌉, and only those shared-encoder tokens are sent to the cloud connector. Instantiated on InternVL2-1B/8B (shared InternViT) and transferred to LLaVA-1.5-7B/13B, LAST is evaluated on 11 benchmarks at 62.5%/37.5%/12.5% retention against random, VisionZip, VisPruner, adapted FastV/SparseVLM, and SGP. The headline result is 95.4% of full-token macro accuracy at 12.5% tokens (99.3%/99.0% at milder budgets), with low score-aggregation cost and reduced cloud FLOPs.

Significance. The paper cleanly reframes visual token pruning as a pre-cloud selection problem rather than post-ingestion sparsification inside the target MLLM, which matches a realistic edge–cloud token interface. The last-query-token observation is simple, training-free, and well motivated by causal masking; combined with importance-weighted k-center it yields consistent gains over both query-agnostic and prior compact-guidance baselines, especially under aggressive budgets and on dense document/chart tasks. Strengths include broad multi-benchmark evaluation, attention-source and selection-strategy ablations (Tables 4–5), efficiency breakdowns (Table 2, Fig. 3), and a cross-family LLaVA check (Table 3). If the proxy–cloud ranking alignment continues to hold in deployment families with compatible token layouts, the method is a practical contribution to collaborative multimodal inference.

major comments (2)
  1. [Method: Edge–Cloud Token Interface] Edge–Cloud Token Interface and Eqs. (1)–(4): the pipeline’s correctness depends on a shared (or layout-compatible) vision encoder so that edge-selected encoder tokens XS are consumed unchanged by ϕc. InternVL2-1B/8B satisfies this by construction; LLaVA-1.5-7B/13B stays inside one family. The central claim is empirically supported in these settings, but the manuscript should state the compatible-encoder/token-layout requirement as an explicit deployment assumption and briefly discuss failure modes or needed adapters when edge and cloud encoders diverge. Without that scope statement, readers may over-generalize the “no cloud internals / direct transmit” claim.
  2. [Table 2; Efficiency Analysis] Table 2 and the “low edge-side selection overhead” claim: Agg. is indeed only 2.51 ms, but full Edge latency is ~237 ms versus 118.55 ms for Full Tokens (no proxy). The dominant cost is the proxy prefill, not score aggregation. The paper should qualify when LAST reduces end-to-end cost (cloud FLOPs/memory and transmission payload versus added edge prefill), and ideally report or bound wall-clock cloud decode and a simple transmission model, so the efficiency narrative is not read as “edge is almost free.”
minor comments (6)
  1. [Table 1] Table 1: at 12.5% retention LAST leads on Avg./Rel. but trails SparseVLM on MME (2097 vs 2160) and SQAIMG (89.7 vs 91.1). A short note that the method optimizes macro trade-off rather than every single benchmark would prevent over-reading “consistently strongest.”
  2. [Comparison Methods] FastV† and SparseVLM† are adapted to the proxy for pre-transmission pruning. One sentence on what was kept vs. dropped from their original in-decoder schedules would make the comparison fully transparent.
  3. [Figure 4] Figure 4 is persuasive for ChartQA/DocVQA; adding one failure or near-miss example would balance the qualitative section.
  4. [Last-Query-Token Attention Scoring] Notation: ewi / ew vs. si is clear in Eqs. (7)–(8), but the text sometimes says “last query token” without restating p⋆; a single reminder that p⋆ is the final text position after multimodal assembly would help.
  5. [Abstract / Introduction] Minor typos/spacing in the compiled text (e.g., missing spaces in “cloudMLLM”, “visual-tokense-quences”, “query-guidedmethodsoften”) should be cleaned in production.
  6. [Related Work] Related Work cites several 2026 works; ensure camera-ready citations match final venues/status for reproducibility of the comparison set.

Circularity Check

0 steps flagged

No significant circularity: empirical pruning method evaluated on external benchmarks with frozen models; no fitted-as-prediction or self-definitional loop.

full rationale

LAST is a training-free systems/method paper. Its load-bearing chain is observational and empirical, not algebraic: under causal attention the last query token attends to the full visual prefix and query (Eqs. 5–8), those frozen proxy attention weights plus importance-weighted k-center selection (Eqs. 9–11) choose a token subset, and performance is measured against public benchmarks and a full-token reference (Tables 1–5). Importance scores are read from a frozen edge VLM, not fitted to maximize Rel.%. Claimed accuracy retention (e.g., 95.4% at 12.5% tokens) is an external measurement, not a restatement of a fitted constant or a definition. Baselines (VisionZip, VisPruner, FastV, SparseVLM, SGP) and ablations (attention source, Top-k vs k-center) are independent controls. Self-citations are absent as load-bearing uniqueness or ansatz imports; related-work citations are prior pruning methods used as competitors. No equation equates the reported result to its inputs by construction. Score 0 is appropriate.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 2 invented entities

The work is an empirical systems method on frozen VLMs. It inherits standard transformer causal attention and public benchmark protocols; it adds design choices (last-token score, multiplicative k-center, shared encoder interface) rather than new physical entities or fitted laws. Free parameters are operational budgets and implementation hyperparameters, not constants fit to prove a theory.

free parameters (3)
  • retention ratio ρ (token budget K=⌈ρN⌉) = evaluated at 0.625, 0.375, 0.125
    User-chosen operating point evaluated at 62.5%, 37.5%, 12.5% (and LLaVA fixed counts 128/64/32); central trade-off curves depend on these settings.
  • importance-weighted k-center objective (product e_w * δ) = jt = argmax e_w_i * δ_i^(t-1)
    Multiplicative combination of normalized last-token attention and feature coverage distance is a hand-chosen selection rule; ablation shows it beats Top-k but is not derived from an optimality theorem.
  • edge proxy / cloud model pairing = InternVL2-1B→8B; LLaVA-1.5-7B→13B
    Primary results use InternVL2-1B proxy with InternVL2-8B cloud (shared ViT); secondary LLaVA-1.5-7B/13B. Claim strength depends on these capacity and interface choices.
axioms (4)
  • ad hoc to paper Under causal self-attention, the last query token’s attention over visual positions is a sufficient query-conditioned relevance signal for pruning.
    Core methodological premise in Last-Query-Token Attention Scoring; supported by ablation vs first/middle/all-mean/+answer but not proved universally.
  • domain assumption Edge and cloud models can share the same vision-encoder token layout so selected encoder tokens are consumable by the cloud connector without re-encoding or alignment training.
    Stated in Edge–Cloud Token Interface; enables the pre-transmission pipeline for InternVL and, with fixed counts, LLaVA experiments.
  • domain assumption Standard softmax causal attention and frozen pretrained VLM behavior are valid substrates for importance estimation without fine-tuning.
    Training-free claim rests on off-the-shelf InternVL/LLaVA components and official templates.
  • domain assumption Public VQA/document/hallucination benchmarks and macro-averaged Rel.% are adequate proxies for edge-cloud pruning utility.
    Entire experimental section evaluates on ChartQA, DocVQA, TextVQA, VQAv2, GQA, VizWiz, SQA-IMG, POPE, MME, MMBench-CN, MM-Vet.
invented entities (2)
  • LAST last-query-token importance signal independent evidence
    purpose: Replace multi-token aggregation or answer-generation guidance with a single-position attention sum across layers/heads.
    Defined in Eqs. 7–8 as s_i summed from A^(ℓ,h)[p★, p_i]; central mechanism of the method.
  • Importance-weighted greedy k-center token selector independent evidence
    purpose: Balance query relevance and feature-space diversity under fixed K.
    Selection rule in Eqs. 9–11; ablation Table 5 contrasts it with Top-k on same scores.

reviewed 2026-07-31 · how reviews work

0 comments
Cite this review

Pith. "Pith review of LAST: The Last Query Token Guides Visual Token Pruning for Edge-Cloud Collaborative MLLM Inference." pith.science (2026). https://pith.science/paper/SHYIADY4

@misc{pith2026260727952,
  author       = {Pith},
  title        = {Pith review of: LAST: The Last Query Token Guides Visual Token Pruning for Edge-Cloud Collaborative MLLM Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SHYIADY4}},
  note         = {Machine review of arXiv:2607.27952}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multimodal foundation models are reshaping edge-cloud visual intelligence from task-specific feature pipelines into token-based interfaces, where edge devices encode visual inputs into tokens for a general-purpose cloud MLLM. However, dense visual-token sequences increase cloud-side inference costs. Existing pruning methods mainly target centralized inference: vision-driven methods can operate before cloud execution but are typically query-agnostic, whereas query-guided methods often rely on internal states of the target MLLM and cannot determine token relevance before transmission. Compact guidance models offer an alternative, but existing designs may require costly attention aggregation or auxiliary generation. We propose LAST, a training-free framework for query-dependent visual token pruning in edge-cloud collaborative MLLM inference. LAST uses a compact edge-side VLM as a guidance proxy and derives a lightweight importance signal from the last query token's attention to visual tokens. Under causal attention, the last query token can attend to the full visual sequence and the entire query context, enabling query-aware pruning without cloud-model access, autoregressive generation, or costly aggregation over multiple query positions. LAST then retains a diverse set of query-relevant visual tokens under a fixed token budget. We evaluate LAST on 11 multimodal benchmarks under multiple token budgets against pruning methods with different guidance strategies. Experiments show that LAST consistently achieves the strongest performance, preserving 95.4% of the full-token accuracy while retaining only 12.5% of the visual tokens, with low edge-side selection overhead and reduced cloud-side computation.

Figures

Figures reproduced from arXiv: 2607.27952 by Chris Xing Tian, Feng Yang, Haoliang Li, Howard Leung, Keyang Zhang, Rongqun Lin, Shiqi Wang, Xiandong Meng, Xinrui Ju.

Figure 1
Figure 1. Figure 1: Two paradigms of edge-cloud visual intelligence: [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of LAST. A compact edge-side LLM performs a single decoding-free prefill pass and uses the last query token’s attention to estimate query-conditioned visual importance. Tokens are then selected by importance-weighted k-center selection. The selected features are then transmitted to the cloud model for final response generation. candidate visual tokens to the current query, and only a fixed￾budget … view at source ↗
Figure 3
Figure 3. Figure 3: Accuracy–efficiency trade-off on Edge-Cloud Model. The average efficiency score jointly captures the prefill-stage speedup and GPU memory reduction relative to the full-token baseline across all datasets. differences mainly arise from edge-side guidance. As shown in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of LAST compared to query￾agnostic methods. 12.5% retained tokens are highlighted with . Thumbnails employed in InternVL are presented in the left corner. spectively. These examples suggest that query conditioning can help preserve ground-truth evidence when it occupies only a small portion of a visually dense input. Choosing the Attention Source. We compare different at￾tention sources for d… 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

42 extracted references · 8 linked inside Pith

  1. [1]

    Science China Information Sciences , volume=

    How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites , author=. Science China Information Sciences , volume=. 2024 , publisher=

    2024

  2. [2]

    Findings of the association for computational linguistics: ACL 2022 , pages=

    Chartqa: A benchmark for question answering about charts with visual and logical reasoning , author=. Findings of the association for computational linguistics: ACL 2022 , pages=

    2022

  3. [3]

    Proceedings of the IEEE/CVF winter conference on applications of computer vision , pages=

    Docvqa: A dataset for vqa on document images , author=. Proceedings of the IEEE/CVF winter conference on applications of computer vision , pages=

  4. [4]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Towards vqa models that can read , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  5. [5]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Making the v in vqa matter: Elevating the role of image understanding in visual question answering , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  6. [6]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Gqa: A new dataset for real-world visual reasoning and compositional question answering , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  7. [7]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Vizwiz grand challenge: Answering visual questions from blind people , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  8. [8]

    Advances in neural information processing systems , volume=

    Learn to explain: Multimodal reasoning via thought chains for science question answering , author=. Advances in neural information processing systems , volume=

  9. [9]

    Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    Evaluating object hallucination in large vision-language models , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    2023

  10. [10]

    Advances in Neural Information Processing Systems , volume=

    Mme: A comprehensive evaluation benchmark for multimodal large language models , author=. Advances in Neural Information Processing Systems , volume=

  11. [11]

    European conference on computer vision , pages=

    Mmbench: Is your multi-modal model an all-around player? , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  12. [12]

    arXiv preprint arXiv:2308.02490 , year=

    Mm-vet: Evaluating large multimodal models for integrated capabilities , author=. arXiv preprint arXiv:2308.02490 , year=

    Pith/arXiv arXiv

  13. [13]

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

    Visionzip: Longer is better but not necessary in vision language models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  14. [14]

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

    Beyond text-visual attention: Exploiting visual cues for effective token pruning in vlms , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  15. [15]

    European Conference on Computer Vision , pages=

    An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  16. [16]

    arXiv preprint arXiv:2410.04417 , year=

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

    Pith/arXiv arXiv

  17. [17]

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

    A stitch in time saves nine: Small vlm is a precise guidance for accelerating large vlms , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  18. [18]

    2023 , eprint =

    Achiam, Josh and Adler, Steven and Agarwal, Sandhini and Ahmad, Lama and Akkaya, Ilge and Aleman, Florencia Leoni and Almeida, Diogo and Altenschmidt, Janko and Altman, Sam and Anadkat, Shyamal and others , journal =. 2023 , eprint =

    2023

  19. [19]

    arXiv preprint arXiv:2309.16609 , year =

    Qwen Technical Report , author =. arXiv preprint arXiv:2309.16609 , year =. 2309.16609 , archivePrefix =

    Pith/arXiv arXiv

  20. [20]

    2023 , eprint =

    Bai, Jinze and Bai, Shuai and Yang, Shusheng and Wang, Shijie and Tan, Sinan and Wang, Peng and Lin, Junyang and Zhou, Chang and Zhou, Jingren , journal =. 2023 , eprint =

    2023

  21. [21]

    arXiv preprint arXiv:2512.16349 , year =

    Collaborative Edge-to-Server Inference for Vision-Language Models , author =. arXiv preprint arXiv:2512.16349 , year =. 2512.16349 , archivePrefix =

    Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2604.26508 , year =

    Progressive Semantic Communication for Efficient Edge-Cloud Vision-Language Models , author =. arXiv preprint arXiv:2604.26508 , year =. 2604.26508 , archivePrefix =

    Pith/arXiv arXiv

  23. [23]

    Xu, Tong and Shi, Hailong and Gao, Xingyu , booktitle =

  24. [24]

    Du, Chenxi and Deng, Yongheng and Liu, Jiani and Zhang, Yujia and Chen, Xi and Ren, Ju , booktitle =

  25. [25]

    Prune Redundancy, Preserve Essence: Vision Token Compression in

    Fang, Zhengyao and Lyu, Pengyuan and Zhang, Chengquan and Lu, Guangming and Yu, Jun and Pei, Wenjie , journal =. Prune Redundancy, Preserve Essence: Vision Token Compression in. 2026 , eprint =

    2026

  26. [26]

    arXiv preprint arXiv:2310.03744 , year =

    Improved Baselines with Visual Instruction Tuning , author =. arXiv preprint arXiv:2310.03744 , year =. 2310.03744 , archivePrefix =

    Pith/arXiv arXiv

  27. [27]

    2026 , eprint =

    Jia, Yuanyuan and Tang, Shunpu and Yang, Qianqian , journal =. 2026 , eprint =

    2026

  28. [28]

    2024 , eprint =

    Zhang, Mingjin and Cao, Jiannong and Shen, Xiaoming and Cui, Zeyang , journal =. 2024 , eprint =

    2024

  29. [29]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) , pages =

    Token Level Routing Inference System for Edge Devices , author =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) , pages =. 2025 , address =

    2025

  30. [30]

    Hao, Zixu and Jiang, Huiqiang and Jiang, Shiqi and Ren, Ju and Cao, Ting , booktitle =. Hybrid. 2024 , publisher =

    2024

  31. [31]

    2024 , eprint =

    Jin, Hongpeng and Wu, Yanzhao , journal =. 2024 , eprint =

    2024

  32. [32]

    , journal =

    Yuan, Liangqi and Han, Dong-Jun and Wang, Shiqiang and Brinton, Christopher G. , journal =. Device--Cloud Collaborative. 2025 , eprint =

    2025

  33. [33]

    2605.18853 , archivePrefix =

    arXiv preprint arXiv:2605.18853 , year =. 2605.18853 , archivePrefix =

    Pith/arXiv arXiv

  34. [34]

    Wang, Ziyao and Chen, Chen and Li, Jingtao and Zhuang, Weiming and Huang, Jiabo and Li, Ang and Lyu, Lingjuan , booktitle =

  35. [35]

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

    Rethinking Token Reduction for Large Vision-Language Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  36. [36]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Improved baselines with visual instruction tuning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  37. [37]

    2019 IEEE International Conference on Multimedia and Expo (ICME) , pages =

    Towards Digital Retina in Smart Cities: A Model Generation, Utilization and Communication Paradigm , author =. 2019 IEEE International Conference on Multimedia and Expo (ICME) , pages =. 2019 , doi =

    2019

  38. [38]

    IEEE Transactions on Circuits and Systems for Video Technology , volume =

    Digital Retina: A Way to Make the City Brain More Efficient by Visual Coding , author =. IEEE Transactions on Circuits and Systems for Video Technology , volume =. 2021 , doi =

    2021

  39. [39]

    2023 , eprint=

    Multimodal Foundation Models: From Specialists to General-Purpose Assistants , author=. 2023 , eprint=

    2023

  40. [40]

    arXiv preprint arXiv:2407.10671 , year =

    Qwen2 Technical Report , author =. arXiv preprint arXiv:2407.10671 , year =. doi:10.48550/arXiv.2407.10671 , url =. 2407.10671 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.10671

  41. [41]

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

    QuietPrune: Query-Guided Early Token Pruning for Vision-Language Models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  42. [42]

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

    IF-Prune: Information-Flow Guided Token Pruning for Efficient Vision-Language Models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

This paper was first reviewed by grok-4.5 on July 31, 2026.