Pith. sign in

REVIEW 4 major objections 5 minor 35 references

A lightweight preprocessor predicts the minimal resolution a vision-language model needs for each image–query pair, cutting prefill compute by 65–85% while keeping accuracy.

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

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

T0 review · deepseek-v4-flash

2026-08-04 08:40 UTC pith:4PZS2QLS

load-bearing objection CARES is a genuinely useful pre-tokenization resolution router for VLMs, with broad experiments, but the cross-model claims rest on a missing per-target retraining statement and the evaluation has several unaddressed leaks. the 4 major comments →

arxiv 2510.19496 v3 pith:4PZS2QLS submitted 2025-10-22 cs.CV cs.AIcs.LG

CARES: Context-Aware Resolution Selector for VLMs

Moshe Kimhi , Nimrod Shabtay , Raja Giryes , Chaim Baskin , Eli Schwartz This is my paper
classification cs.CV cs.AIcs.LG
keywords resolution selectionvision-language modelsinference efficiencyadaptive computevisual tokensANLSdynamic resolutionprefill compute
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.

CARES claims that most vision-language queries do not need high-resolution images, and that the minimal sufficient resolution for a given image and question can be predicted cheaply before the expensive model runs. It trains a small 350M proxy to classify, from a single low-resolution pass, which of three resolutions (384, 768, 1024) suffices for a frozen target VLM, using a rule that marks a resolution sufficient once answer quality (ANLS) crosses 0.85 and higher resolutions add at most 0.1. At inference the classifier's probabilities are averaged into a continuous resolution. Across five benchmarks and four target VLMs, CARES reports accuracy on par with always using native resolution while lowering prefill FLOPs (or API cost) by 65–85%. If right, it means a large share of visual tokens in current VLM deployments are unnecessary and can be avoided upstream of tokenization.

Core claim

The paper's central claim is that the input resolution of a frozen vision-language model can be treated as a decision variable controlled by a separate lightweight module, and that the optimal decision is predictable from cheap low-resolution joint image–text features. The discovery is that a simple per-example labeling rule—roll out the target VLM at a few resolutions, take the lowest one whose ANLS reaches 0.85 with no further gain above 0.1—produces training signal sufficient to learn such a selector, and that continuous interpolation over the classifier's softmax yields further compute savings without quality loss. The result holds across document and natural-image benchmarks and across

What carries the argument

The labeling rule (Eq. 2 / Algorithm 1) defines the ground-truth sufficient resolution per example: the smallest rk such that the target VLM's ANLS is at least 0.85 and no higher resolution improves it by more than 0.1. The selector is a frozen 350M proxy VLM (SmolVLM with the top layers removed) that encodes the image at the lowest resolution together with the query, and a lightweight classifier head maps the mid-layer hidden state to a distribution over the three discrete resolutions; at inference, the expected resolution under that distribution (Eq. 3) supplies a continuous control signal. The sufficiency rule is what carries the argument: it converts the continuous question 'what resolut

Load-bearing premise

The results stand or fall on the labeling rule that a resolution is 'sufficient' exactly when the target model's ANLS reaches 0.85 and no higher resolution improves it by more than 0.1—if that rule overestimates sufficiency (because ANLS tolerates partial answers or the 384/768/1024 grid misses the true threshold), then CARES learns to route to resolutions that merely satisfy the metric, and the reported compute savings come at a hidden accuracy cost.

What would settle it

Run the same benchmarks and target models with a stricter sufficiency rule (e.g., exact match on the full answer string, or ANLS with threshold 0.95 and margin 0.02) and a finer resolution grid such as {384, 512, 640, 768, 896, 1024}; if the learned selector then requires much higher resolutions or loses more than a sub-point of accuracy, the 65–85% savings are artifacts of the ANLS-based labeling.

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

If this is right

  • Current VLM pipelines can be front-ended by a resolution selector without retraining the target model, so the compute savings are immediately deployable.
  • Resolution routing is query-dependent: the same image can be served at low resolution for a coarse question and at high resolution for a fine-grained one, which static resizing cannot do.
  • The method composes with post-tokenization token-compression methods, so pixel-level and token-level savings multiply.
  • Continuous resolution selection extends the discrete training grid, giving finer cost–accuracy trade-offs than any fixed menu.
  • Cost savings apply to API-based models too, since billing scales with input tokens/resolution.

Where Pith is reading between the lines

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

  • The 0.85/0.1 sufficiency rule is a choice about what 'sufficient' means; if ANLS rewards partially correct strings at low resolution, the selector may systematically under-allocate for tasks where the fine detail is not captured by the metric's tolerance. Whether the reported parity holds on strict-accuracy tasks outside the benchmarks (e.g., dense OCR, medical imaging) is an open question the pap
  • Because labels are generated by the same target VLM that is later evaluated, the selector learns that model's idiosyncratic resolution sensitivity; a selector trained on one family (e.g., Qwen) may over- or under-route for another, and the paper does not report cross-target transfer.
  • A testable extension: use the same labeling rule but with a strict metric (exact match on full answer strings) and a finer resolution grid (e.g., steps of 128 between 384 and 1024) to see whether the reported 65–85% savings persist or shrink; this would isolate whether ANLS's partial-credit structure is the source of the large savings.
  • If resolution needs are predictable from mid-layer features of a 350M VLM, similar cheap predictors may predict other inference cost levers (tiling strategy, number of tokens, number of reasoning steps).

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces CARES, a lightweight preprocessing module that selects the input resolution for a target VLM given an image--query pair. CARES uses a frozen 350M SmolVLM-based feature extractor at a low resolution and a small classifier head to predict the minimal sufficient resolution. Sufficiency labels are generated by Algorithm 1/Eq. (2): for each example, the target VLM (Granite-Vision, per Table 1) is evaluated at resolutions {384,768,1024}; the lowest resolution is chosen for which ANLS >= 0.85 and no higher resolution improves ANLS by more than 0.1. At inference, softmax probabilities over the discrete resolutions are combined into a continuous resolution. The method is evaluated on Ai2D, ChartQA, DocVQA, OCRBench, and SeedBench-2 across Granite-Vision 3.3-2B, InternVL3-8B, Qwen2.5-VL-72B, and GPT-4o, reporting 60--85% prefill FLOPs or cost savings with near-parity accuracy.

Significance. If the reported results are reliable, CARES is a practically useful contribution: it targets resolution selection before tokenization, is complementary to token-pruning methods, and the continuous-resolution interpolation is a clean way to use discrete supervision. The evaluation is broad, covering four target VLMs and five benchmarks. However, two load-bearing issues temper the claim as stated: the training data overlaps with the evaluation benchmarks without a stated disjoint split, and the cross-model results are presented without specifying whether separate per-target labels were generated. A further concern is that the ANLS-based sufficiency rule, which is also used as the evaluation metric on DocVQA and OCRBench, can make parity partly a consequence of the label definition. These issues are addressable with additional experiments or clarifications, but they must be resolved before the main claims can be accepted.

major comments (4)
  1. [Section 4.1, Table 2] The training set is built from 20K samples each of TextVQA, ChartQA, DocVQA, and LLaVA-Multi, while the evaluation set includes ChartQA and DocVQA. The paper never states that the sampled training instances are disjoint from the ChartQA/DocVQA evaluation splits. If the same examples appear in both training and evaluation, the Table 2 rows for ChartQA and DocVQA do not measure generalization. Please state the exact train/test split and either confirm disjointness or report results on genuinely held-out examples.
  2. [Section 3.2/Table 1 vs. Section 3.3/Table 2] The labeling procedure is described for Granite-Vision: Table 1 says 'We process each input through a pretrained VLM (Granite-Vision) at three fixed resolutions...', and Section 3.3 states that CARES is VLM-agnostic and requires no retraining. Table 2 then reports results for InternVL3-8B, Qwen2.5-VL-72B, and GPT-4o, but the paper does not state whether separate sufficiency labels were generated for each target model. As written, the reader must assume a single Granite-labeled CARES is applied to all targets. A resolution that is sufficient for Granite-Vision need not be sufficient for Qwen2.5-VL-72B or GPT-4o, so the cross-model parity claims are unsupported. If per-target labels were generated, this must be stated explicitly; if not, the table should be reframed as a transfer experiment and the claim 'works with any VLM... no retraining required' should be qualified accordingly.
  3. [Eq. (2), Algorithm 1, Section 4.2] The sufficiency label is defined by ANLS thresholds on a three-point resolution grid, and the same ANLS metric is used to evaluate DocVQA and OCRBench. Under Eq. (2), a 384-resolution label only guarantees ANLS >= 0.85 and no improvement above 0.1 at higher resolutions; it does not guarantee that the answer is preserved or that finer-grained detail is not needed. On the training-distribution benchmarks, the observed parity is therefore partially a consequence of the labeling criterion rather than evidence of true task preservation. The paper should report answer-level agreement or exact-match comparisons between the selected resolution and native resolution, and quantify how often higher resolutions change the predicted string even when the ANLS rule labels the example 'sufficient' at a lower resolution.
  4. [Section 4.2, Tables 2 and 5] The text claims 'at most a sub-point change in macro performance' and 'no drop in accuracy', but several per-dataset differences are larger than one point. For example, Table 2 shows Qwen2.5-VL-72B on ChartQA dropping from 0.874 to 0.836 (3.8 points), Qwen2.5-VL-72B on SeedBench-2 dropping from 0.807 to 0.785 (2.2 points), and GPT-4o on OCRBench dropping from 0.770 to 0.746 (2.4 points). Table 5 also shows the continuous selector dropping Qwen2.5-VL-72B from 0.851 to 0.839. The macro-average statement may be correct, but the paper should report per-dataset worst-case differences and preferably confidence intervals, and the 'maintains accuracy' phrasing should be qualified.
minor comments (5)
  1. [Abstract and Section 1] Typos: 'ofter' should be 'often' (Abstract); 'ofthe' should be 'of the' (Section 1).
  2. [Table 1] The row label 'GTP. Carter' appears to be a typo for 'GT: Carter'. The table would benefit from a caption that explains the example columns more explicitly.
  3. [Section 4.2, Figure 2] The figure reports '2.58 vs. 7.5' TFLOPs without defining the resolution or model configuration for either number. Please add units or a table so the comparison is reproducible.
  4. [Section 4.3, Table 5] Table 5 states 'no drop in accuracy' in the surrounding text, but the continuous row for Qwen2.5-VL-72B is 0.839 versus 0.851 native. This should be reconciled with the claim in the text.
  5. [Section 4] No error bars, multiple seeds, or statistical significance tests are reported. Given that the central claim is about accuracy preservation, at least a few independent runs or bootstrap intervals would strengthen the result.

Circularity Check

0 steps flagged

No significant circularity: CARES is a standard supervised selector; its accuracy claims are measured on held-out benchmarks, not entailed by the labeling rule.

full rationale

The paper defines sufficiency labels via Algorithm 1 / Eq. 2 using the target VLM's ANLS at three resolutions, then trains a classifier to predict those labels and evaluates the downstream VLM's accuracy/ANLS on five benchmarks (Table 2). This is an ordinary supervised-learning pipeline, not a derivation: nothing in Eq. 2 guarantees the held-out benchmark scores in Table 2, and the selector could fail by choosing an insufficient resolution. The DocVQA/OCRBench rows use the same ANLS metric as the label rule, but metric consistency is not circularity unless the evaluation score is an algebraic consequence of the training labels; it is not, because the test examples and classifier outputs are not used in Eq. 2. The self-citations (WAVECLIP and Granite Vision, both with overlapping authors) are used as related work and as the data-generation VLM, respectively, and are not load-bearing evidence for the central efficiency claim. The Limitations section itself notes: 'Our supervision uses multi-resolution rollouts of a target VLM and thus inherits that model’s biases and limited language support' — a real external-validity caveat for cross-model generalization, but not a circular step. The paper's failure to state whether per-target labels were generated for each evaluated VLM, and whether the ChartQA/DocVQA training samples are disjoint from evaluation splits, are reporting/support gaps, not circularity: they affect validity of the cross-model and held-out claims, but the reported numbers are still empirically measured rather than forced by construction. No step in the derivation chain reduces to its own input by definition.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim rests on five hand-set hyperparameters (tau, delta, resolution grid, resolution range, label smoothing), five modeling assumptions — principally the ANLS-sufficiency equivalence and the transfer from SmolVLM features to target-VLM behavior — and no invented entities. The heaviest cost is the labeling rule, which defines sufficiency through the target VLM's own ANLS and thus shapes both training and evaluation.

free parameters (5)
  • ANLS sufficiency threshold tau = 0.85
    Hand-set in Eq. 2; every training label depends on it; no sensitivity analysis is reported.
  • improvement margin delta = 0.1
    Hand-set in Eq. 2 to ignore small gains; changes which resolution is deemed sufficient.
  • resolution grid R_d = [384, 768, 1024]
    Hand-chosen in Sec 3.2; the minimal sufficient resolution is only ever measured on this grid, so labels cannot be finer than grid spacing.
  • label smoothing alpha = 0.05
    Hand-set in Sec 3.3, tuned in Table 6 for the continuous-interpolation deployment trick.
  • resolution range [r_min, r_max] = [384, 1024]
    Hand-chosen in Sec 3.2; all savings are relative to 1024, not each model's true native resolution.
axioms (5)
  • domain assumption ANLS score is a faithful measure of whether a resolution suffices for the task
    Eq. 1-2 and Algorithm 1 define supervision entirely through ANLS >= 0.85 with margin <= 0.1; if ANLS saturates while fine detail is still missing (plausible for OCR/strings), labels are systematically too low.
  • domain assumption Target VLM accuracy converges with resolution, so a minimal sufficient resolution exists
    Algorithm 1 returns the first r_k with u_k >= tau and no later gain > delta; on noisy or non-monotone curves the rule silently defaults to r_K = 1024, and the default frequency is not analyzed.
  • ad hoc to paper SmolVLM layer-16 features at 384px encode enough joint image-query information to predict the target VLM's resolution needs
    Sec 3.3; the Table 3 ablation shows this choice beats SigLIP pooling, but nothing guarantees transfer across VLMs with different resolution sensitivities; this is the core modeling bet.
  • domain assumption The 80K training sample is representative of and disjoint from the evaluation benchmarks
    Sec 4.1 samples 20K instances from TextVQA, ChartQA, DocVQA, and LLaVA-Multi; ChartQA and DocVQA are also evaluated in Table 2, and no disjoint-split statement is given.
  • standard math Standard supervised-learning assumptions (i.i.d. samples, fixed labeling procedure, cross-entropy optimization)
    Training uses cross-entropy loss over 6 epochs (Sec 3.3); the i.i.d. assumption is implicit and not tested for label noise.

pith-pipeline@v1.3.0-alltime-deepseek · 224 in / 16886 out tokens · 183475 ms · 2026-08-04T08:40:02.707430+00:00 · methodology

0 comments
read the original abstract

Large vision-language models (VLMs) commonly process images at native or high resolution to remain effective across tasks. This inflates visual tokens ofter to 97-99% of total tokens, resulting in high compute and latency, even when low-resolution images would suffice. We introduce \emph{CARES}-a \textbf{C}ontext-\textbf{A}ware \textbf{R}esolution \textbf{S}elector, a lightweight preprocessing module that, given an image-query pair, predicts the \emph{minimal} sufficient input resolution. CARES uses a compact VLM (350M) to extract features and predict when a target pretrained VLM's response converges to its peak ability to answer correctly. Though trained as a discrete classifier over a set of optional resolutions, CARES interpolates continuous resolutions at inference for fine-grained control. Across five multimodal benchmarks spanning documents and natural images, as well as diverse target VLMs, CARES preserves task performance while reducing compute by up to 80%.

Figures

Figures reproduced from arXiv: 2510.19496 by Chaim Baskin, Eli Schwartz, Moshe Kimhi, Nimrod Shabtay, Raja Giryes.

Figure 1
Figure 1. Figure 1: Overview of CARES. Given an image and its query, CARES predicts the minimal sufficient input resolution. The image is resized accordingly and, together with the query, passed to a downstream VLM. Coarse queries are routed to lower resolution; fine-grained queries that require more detail trigger higher resolution, which yields more visual tokens in the VLM. set of resolutions R and a task performance metri… view at source ↗
Figure 2
Figure 2. Figure 2: Accuracy vs. TTFT for DocVQA with Qwen2.5-VL-72B across native and fixed-resolution settings versus CARES. Bubble size indicates the number of pixels processed by the model [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Histogram of the predicted resolutions r˜ by CARES for OCRBench. Evaluation We evaluate on five public benchmarks varying from documents to natural images: Ai2D (Kembhavi et al., 2016), ChartQA (Masry et al., 2022), DocVQA (Mathew et al., 2021), OCRBench (Liu et al., 2024b), and SeedBench-2 (Li et al., 2023). For Ai2D, ChartQA, and SeedBench-2 we report exact-match accuracy. For DocVQA and OCRBench we repo… 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

35 extracted references · 19 linked inside Pith

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    Pith/arXiv arXiv 2023

  2. [2]

    Kazi Hasan Ibn Arif, JinYi Yoon, Dimitrios S Nikolopoulos, Hans Vandierendonck, Deepu John, and Bo Ji. 2025. Hired: Attention-guided token dropping for efficient inference of high-resolution vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 1773--1781

    2025

  3. [3]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, and 8 others. 2025. Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923

    Pith/arXiv arXiv 2025

  4. [4]

    Lucas Beyer, Pavel Izmailov, Alexander Kolesnikov, Mathilde Caron, Simon Kornblith, Xiaohua Zhai, Matthias Minderer, Michael Tschannen, Ibrahim Alabdulmohsin, and Filip Pavetic. 2023. https://arxiv.org/abs/2212.08013 Flexivit: One model for all patch sizes . Preprint, arXiv:2212.08013

    Pith/arXiv arXiv 2023

  5. [5]

    Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. 2023. Token merging: Your ViT but faster. In International Conference on Learning Representations

    2023

  6. [6]

    Mu Cai, Jianwei Yang, Jianfeng Gao, and Yong Jae Lee. 2025. Matryoshka multimodal models. Proceedings of the International Conference on Learning Representation

    2025

  7. [7]

    Gritsenko, Mario Lucic, and Neil Houlsby

    Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Peter Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim Alabdulmohsin, Avital Oliver, Piotr Padlewski, Alexey A. Gritsenko, Mario Lucic, and Neil Houlsby. 2023. Patch n pack: Navit, a vision transformer for any aspect ratio and resolution. In Thirty-...

    2023

  8. [8]

    Junshan Hu, Jialiang Mao, Zhikang Liu, Zhongpu Xia, Peng Jia, and Xianpeng Lang. 2025. https://arxiv.org/abs/2504.03154 Tokenflex: Unified vlm training for flexible visual tokens inference . Preprint, arXiv:2504.03154

    Pith/arXiv arXiv 2025

  9. [9]

    Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. 2024. Mantis: Interleaved multi-image instruction tuning. arXiv preprint arXiv:2405.01483

    Pith/arXiv arXiv 2024

  10. [10]

    Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. 2016. https://arxiv.org/abs/1603.07396 A diagram is worth a dozen images . Preprint, arXiv:1603.07396

    Pith/arXiv arXiv 2016

  11. [11]

    Moshe Kimhi, Erez Koifman, Ehud Rivlin, Eli Schwartz, and Chaim Baskin. 2025. https://arxiv.org/abs/2509.21153 Waveclip: Wavelet tokenization for adaptive-resolution clip . Preprint, arXiv:2509.21153

    arXiv 2025

  12. [12]

    Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. 2023. Seed-bench-2: Benchmarking multimodal large language models. arXiv preprint arXiv:2311.17092

    Pith/arXiv arXiv 2023

  13. [13]

    Li, Sachin Goyal, Joao D

    Kevin Y. Li, Sachin Goyal, Joao D. Semedo, and J. Zico Kolter. 2024. https://arxiv.org/abs/2411.03312 Inference optimal vlms need only one visual token but larger models . Preprint, arXiv:2411.03312

    Pith/arXiv arXiv 2024

  14. [14]

    Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. 2022. https://arxiv.org/abs/2202.07800 Not all patches are what you need: Expediting vision transformers via token reorganizations . Preprint, arXiv:2202.07800

    Pith/arXiv arXiv 2022

  15. [15]

    Zhihang Lin, Mingbao Lin, Luxi Lin, and Rongrong Ji. 2025. Boosting multimodal large language models with visual tokens withdrawal for rapid inference. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 5334--5342

    2025

  16. [16]

    Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. 2024 a . https://llava-vl.github.io/blog/2024-01-30-llava-next/ Llava-next: Improved reasoning, ocr, and world knowledge

    2024

  17. [17]

    Yuliang Liu, Zhang Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xu-Cheng Yin, Cheng-Lin Liu, Lianwen Jin, and Xiang Bai. 2024 b . https://doi.org/10.1007/s11432-024-4235-6 Ocrbench: on the hidden mystery of ocr in large multimodal models . Science China Information Sciences, 67(12)

    doi:10.1007/s11432-024-4235-6 2024

  18. [18]

    Andrés Marafioti, Orr Zohar, Miquel Farré, Merve Noyan, Elie Bakouch, Pedro Cuenca, Cyril Zakka, Loubna Ben Allal, Anton Lozhkov, Nouamane Tazi, Vaibhav Srivastav, Joshua Lochner, Hugo Larcher, Mathieu Morlon, Lewis Tunstall, Leandro von Werra, and Thomas Wolf. 2025. Smolvlm: Redefining small and efficient multimodal models. arXiv preprint arXiv:2504.05299

    Pith/arXiv arXiv 2025

  19. [19]

    Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. 2022. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244

    Pith/arXiv arXiv 2022

  20. [20]

    Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. 2021. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200--2209

    2021

  21. [21]

    Hadas Orgad, Michael Toker, Zorik Gekhman, Roi Reichart, Idan Szpektor, Hadas Kotek, and Yonatan Belinkov. 2024. Llms know more than they show: On the intrinsic representation of llm hallucinations. arXiv preprint arXiv:2410.02707

    Pith/arXiv arXiv 2024

  22. [22]

    Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. 2021. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In Advances in Neural Information Processing Systems (NeurIPS)

    2021

  23. [23]

    Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. 2019. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317--8326

    2019

  24. [24]

    Granite Vision Team, Leonid Karlinsky, Assaf Arbelle, Abraham Daniels, Ahmed Nassar, Amit Alfassi, Bo Wu, Eli Schwartz, Dhiraj Joshi, Jovana Kondic, and 1 others. 2025. Granite vision: a lightweight, open-source multimodal model for enterprise intelligence. arXiv preprint arXiv:2502.09927

    Pith/arXiv arXiv 2025

  25. [25]

    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. 2024. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191

    Pith/arXiv arXiv 2024

  26. [26]

    Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, and Dahua Lin. 2025. https://arxiv.org/abs/2410.17247 Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction . Preprint, arXiv:2410.17247

    Pith/arXiv arXiv 2025

  27. [27]

    Jiarui Zhang, Mahyar Khayatkhoei, Prateek Chhikara, and Filip Ilievski. 2025 a . https://arxiv.org/abs/2502.17422 MLLM s know where to look: Training-free perception of small visual details with multimodal LLM s . In The Thirteenth International Conference on Learning Representations

    Pith/arXiv arXiv 2025

  28. [28]

    Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu, Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuanhan Zhang, Jingkang Yang, Chunyuan Li, and Ziwei Liu. 2024. https://arxiv.org/abs/2407.12772 Lmms-eval: Reality check on the evaluation of large multimodal models . Preprint, arXiv:2407.12772

    Pith/arXiv arXiv 2024

  29. [29]

    Shaolei Zhang, Qingkai Fang, Zhe Yang, and Yang Feng. 2025 b . Llava-mini: Efficient image and video large multimodal models with one vision token. In International Conference on Learning Representations (ICLR)

    2025

  30. [30]

    Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, and 1 others. 2025 c . Sparsevlm: Visual token sparsification for efficient vision-language model inference. In International Conference on Machine Learning

    2025

  31. [31]

    Wangbo Zhao, Yizeng Han, Jiasheng Tang, Zhikai Li, Yibing Song, Kai Wang, Zhangyang Wang, and Yang You. 2024. A stitch in time saves nine: Small vlm is a precise guidance for accelerating large vlms. arXiv preprint arXiv:2412.03324

    Pith/arXiv arXiv 2024

  32. [32]

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, and 1 others. 2025. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479

    Pith/arXiv arXiv 2025

  33. [33]

    Mingjian Zhu, Kai Han, Enhua Wu, Qiulin Zhang, Ying Nie, Zhenzhong Lan, and Yunhe Wang. 2021. Dynamic resolution network. In Advances in Neural Information Processing Systems (NeurIPS), Red Hook, NY, USA

    2021

  34. [34]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  35. [35]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...