Pith. sign in

REVIEW 4 major objections 5 minor 40 references

Existing evaluation overstates how much visual-text compression preserves: a 95.4% OCR score can hide true text preservation of only 11.8%.

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 ↗

A new benchmark using low-predictability text shows that standard evaluations overestimate how much text visual-text compression methods actually preserve.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A genuinely new benchmark that exposes how much semantic priors inflate VTC scores, but the headline numbers are missing a no-compression ceiling control and need a revision before they can be taken at face value. the 4 major comments →

arxiv 2608.01848 v1 pith:XNMUE7HW submitted 2026-08-03 cs.CV

Decoupling semantics from vision: A framework for faithful visual-text compression evaluation

Yonghan Gao , Zehong Chen , Lijian Xu , Jingzhi Chen , Jingwei Guan , Xingyu Zeng This is my paper
classification cs.CV
keywords visual-text compressionlong-context modelingmultimodal LLM evaluationOCR fidelitysemantic prior compensationbenchmark constructionrendering evaluationtext preservation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper argues that current evaluations of visual-text compression (VTC) are overly optimistic because multimodal language models fill in missing text using linguistic context rather than visual evidence. To separate rendering quality from model reasoning, it proposes a framework: render long texts with almost no semantic predictability, ask several models to transcribe them, and use the highest word-recovery accuracy as the preservation estimate. On this benchmark, a leading VTC system that appears to preserve 95.4% of text at 10x compression actually preserves about 11.8%. The implication is that downstream accuracy measures the joint effect of rendering and model priors, not the text information that VTC designers need to optimize.

Core claim

The central claim is that VTC text preservation cannot be inferred from downstream OCR or VQA scores because those scores reflect how much linguistic context a model can exploit. The authors introduce ZeroSense, a corpus whose word sequences are sampled to have extremely low conditional probability under several language models, so no word can be guessed from its neighbors. When this corpus is rendered by VTC methods and transcribed by multiple MLLMs, recovery accuracy drops sharply relative to natural-text controls. Taking the maximum accuracy across the model set approximates the upper bound of what the rendering itself preserved, disentangling visual fidelity from evaluator capability.

What carries the argument

The machinery is the ZeroSense benchmark coupled with a model-decoupled evaluation loop. Low-predictability texts are generated by sampling words below a probability threshold and validated across multiple language models, then rendered via each VTC method. The rendered images are transcribed with a fixed OCR instruction by several MLLMs, and word-level recovery accuracy is scored against the source; the maximum score across models is reported as the preservation measure. This isolates visual fidelity by removing the statistical regularities that would otherwise let models guess unreadable text.

Load-bearing premise

The framework assumes that after probability-constrained generation and cross-model validation, no remaining statistical regularities in word order, spelling, or character patterns can be exploited by a strong OCR model; if any model still uses such cues, ZeroSense overestimates true preservation.

What would settle it

Render a set of images where the same visual glyphs appear but with characters randomly permuted within words and word order shuffled; if an OCR model recovers these at similar rates to ZeroSense, the benchmark is not isolating visual evidence. Alternatively, train a small character-level language model on the ZeroSense corpus and check whether its next-word predictions exceed chance; if they do, the corpus retains exploitable statistical structure.

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

If this is right

  • Published VTC fidelity claims likely overstate how much text survives compression; true compression ratios with acceptable preservation are lower than reported.
  • Rendering configurations should be compared by their compression-preservation trade-off on low-predictability text, not by downstream task accuracy.
  • The framework provides a practical audit protocol: any VTC method can be evaluated by rendering ZeroSense-style text and taking the best OCR recovery across multiple models.
  • If the estimates are correct, downstream task performance after VTC is weak evidence of preservation, and long-context systems should be validated on fidelity before deployment.

Where Pith is reading between the lines

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

  • The same logic should apply to non-English scripts: models with weaker linguistic priors for a language would show smaller gaps between natural-text and ZeroSense scores, so extending the benchmark to multiple languages would test the generality of the overestimation.
  • Comparing a model's OCR accuracy on ZeroSense versus natural text could yield a quantitative 'semantic compensation ratio' that isolates how much of its reading ability is visual versus contextual.
  • A natural production extension is to set a minimum ZeroSense preservation target for each compression ratio, giving developers a concrete acceptance test before shipping a VTC model.
Share X Bluesky LinkedIn Reddit HN

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 proposes an evaluation framework for visual-text compression (VTC) that aims to separate genuine text preservation from the semantic priors of multimodal large language models (MLLMs). The framework uses complete-text OCR, aggregates scores as the maximum over a fixed set of MLLMs, and introduces ZeroSense, a synthetic benchmark of low-predictability, single-token words. Experiments compare three rendering methods (FOX, OmniDocBench, Glyph) at multiple compression ratios and report a dramatic gap between natural-text OCR accuracy and ZeroSense accuracy; for example, DeepSeek-OCR reaches 95.4% on FOX text at 10x compression but only 11.8% on ZeroSense. The paper concludes that existing VTC evaluations substantially overestimate text preservation because MLLMs compensate for lost text via linguistic priors.

Significance. The paper addresses an important and timely problem: VTC methods are increasingly used for long-context modeling, and their evaluation is dominated by downstream tasks that can be solved from partial text and semantic inference. The proposed framework—complete-text recovery, max-over-models aggregation, and low-predictability text—is a principled step toward decoupling rendering quality from evaluator capability. The ZeroSense benchmark is a concrete, reproducible artifact, and the paper includes useful ablations on context length, dataset size, and perturbation baselines. If the missing controls are supplied, this framework could become a valuable evaluation tool for the community. The central idea is sound, but the headline quantitative claims are not yet supported because the paper lacks an uncompressed ceiling control and because the validation/evaluation model sets overlap.

major comments (4)
  1. [§5.2.1 / Table 2] The headline quantitative claim—DeepSeek-OCR preserves only 11.8% of source text at 10x FOX—requires a ceiling control that the paper does not report. ZeroSense texts are deliberately composed of rare, low-probability single-token words (Sec. 4.1), which can be hard for MLLMs to transcribe even from a perfectly rendered image. The max-over-models aggregation in Eq. (8) only gives an upper bound if at least one evaluator can read these strings from a clear image; this is never established. Without a 1x or near-lossless rendering baseline on ZeroSense, the reported 11.8% cannot be distinguished from 'ZeroSense is inherently hard to OCR.' Please add a 1x control (and ideally a very low compression, e.g., 1.5x), and report per-model ZeroSense accuracy at that ceiling.
  2. [§4.1 vs §5.1] The construction of ZeroSense uses the causal LM components of Glyph, DeepSeek-OCR, and Hunyuan in Eq. (10) to filter texts. The evaluator set in Sec. 5.1 is DeepSeek-OCR, HunyuanOCR, and Qwen2.5-VL. Thus two of the three validation models are also evaluation models. The flagship 11.8% FOX 10x number is from DeepSeek-OCR (Table A4), a validation model. Since the text was explicitly selected to be unpredictable for DeepSeek, the low score is partly by construction. At minimum, evaluate with held-out models not used in validation, or split validation/evaluation, and report the two groups separately. Qwen2.5-VL is a held-out model but is unavailable for many high-compression FOX entries.
  3. [§4.1, Eq. (9)-(10)] The claim that ZeroSense results are 'purely reflective of VTC quality' (Abstract) assumes that word-level conditional-probability constraints remove all non-visual inference. However, the corpus may retain character-level or orthographic regularities: rare words like 'draxen' still follow English-like letter patterns, which models could exploit to correct blurred or missing characters. This is a different mechanism from semantic context and is not controlled by the construction. Please measure residual character-level predictability (e.g., character-level LM perplexity) or add character-scrambling controls, and temper the 'purely reflective' language if such regularities remain.
  4. [§5.3.2, Table 6] The threshold ablation is incomplete and inconsistent with the main configuration. Sec. 5.1 fixes tau1=1e-8 and tau2=1e-5, but Table 6 reports values 1e-3, 1e-4, 1e-6 without specifying whether these are tau1 or tau2, and omits 1e-5. The text then says 'we adopt a threshold of 10^-5 as the verification condition,' which is not in the table. Since the validation threshold tau2 determines cross-model unpredictability, its effect on the preservation estimate should be reported; otherwise the claimed robustness of the threshold choice is unsupported.
minor comments (5)
  1. [§2.1] The paragraph beginning 'Later studies used visual representations to reduce the cost of long-context processing' contains a duplicate of the preceding sentence ('Before VTC, several studies had already explored processing text directly from images'). Remove the duplication.
  2. [§5.2.1] The last paragraph refers to 'Section 4.1 examines the effect of text construction,' but the ablation on text construction appears in Section 5.3.2. The citation should be corrected.
  3. [§3.3, Eq. (7)] The sentence 'the final recovery score of method, the final score of R is' is garbled and should be rewritten.
  4. [§5.2.2 / Tables 2, A4] The paper states that numbers are marked in red, but the preprint is monochrome. Use bold or an asterisk, and define the marker in the caption.
  5. [§5.1 / Table 6] The threshold nomenclature is confusing: specify whether the reported thresholds are tau1 or tau2, and include the actual values (1e-8, 1e-5) used in the main experiments for comparison.

Circularity Check

0 steps flagged

No significant circularity; the framework's construction and evaluation are distinct measurements, though the missing uncompressed ceiling is a validity concern.

full rationale

The paper's derivation chain is not circular. ZeroSense construction (Sec. 4.1) uses probability-constrained generation and validation (Eqs. 9-10) to obtain low-predictability text, while the preservation metric (Eqs. 7-8) is a separate OCR accuracy measurement on rendered images. The OCR accuracy is not computed from the probability thresholds; it is an empirical result from actual visual recognition. The evaluation models overlap with the validation models (DeepSeek-OCR and Hunyuan appear in both sets), but this is a deliberate attempt to suppress semantic priors, not a case where the result reduces to the input probabilities by construction. The central quantitative claim (11.8% at 10x on FOX layout) is empirical and could in principle be higher if the rendering preserved text well and the models recognized it. The only self-citation is reference [21] in Related Work, which merely lists subsequent work and is not load-bearing for the framework's validity. The paper does, however, omit an uncompressed (1x) or near-lossless baseline on ZeroSense; without such a ceiling, low OCR scores could reflect inherent difficulty of reading synthetic rare words rather than compression-induced information loss. This is a missing control that threatens the interpretation of the 'preservation' estimate, but it is not a circularity in the derivation. No equation is defined in terms of another, no fitted parameter is renamed as a prediction, and no uniqueness theorem is imported from the authors' prior work. The framework is self-contained against external benchmarks (FOX, OmniDocBench) and uses a fixed set of evaluation models; the reported divergence between natural-text and ZeroSense scores is an empirical observation, not an analytical tautology.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The central claim rests on two hand-chosen probability thresholds (tau1, tau2) and on three domain assumptions about the relationship between probability, semantics, model capability, and accuracy. No new physical entities are postulated.

free parameters (2)
  • tau1 = 1e-8
    Threshold for the candidate word set in probability-constrained generation (Eq. 9). Hand-chosen; no principled criterion is given, and the paper does not ablate tau1.
  • tau2 = 1e-5
    Validation probability threshold for retaining a sequence (Eq. 10). Hand-chosen as a compromise between strictness and computational feasibility; Table 6 shows the benchmark's accuracy scores vary substantially with tau2 (0.850 at 1e-3 down to 0.588 at 1e-6 at 2.5x), so the absolute scores are sensitive to this choice.
axioms (3)
  • domain assumption Low conditional word probability under a set of LLMs implies low semantic correlation for any MLLM.
    Section 4.1: sequences surviving Eq. (10) are taken to be unpredictable, so any correct OCR must be visually grounded. This is the core assumption enabling ZeroSense.
  • domain assumption Maximum OCR accuracy over a finite set of MLLMs approximates the upper bound of recoverable textual information.
    Section 3.2, Model-Decoupled Evaluation: the paper asserts that 'at least one sufficiently capable model should be able to recover it' but does not prove that three models span the capability space.
  • domain assumption Word-level OCR accuracy between recovered and source text measures textual information preservation.
    Section 3.3, Eq. (7): character/word accuracy is used as the preservation metric, ignoring word order, layout, and structure, which could be relevant for downstream tasks.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Decoupling semantics from vision: A framework for faithful visual-text compression evaluation." pith.science (2026). https://pith.science/paper/XNMUE7HW

@misc{pith2026260801848,
  author       = {Pith},
  title        = {Pith review of: Decoupling semantics from vision: A framework for faithful visual-text compression evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XNMUE7HW}},
  note         = {Machine review of arXiv:2608.01848}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent visual-text compression (VTC) methods, typified by DeepSeek-OCR, report impressive high token compression ratios for long-context modeling tasks by leveraging text-to-image rendering. However, existing evaluation protocols heavily rely on downstream task performance. Such evaluation metrics fail to accurately measure text preservation due to the strong inherent linguistic priors of Multimodal Large Language Models (MLLMs). In this work, we introduce a new evaluation framework that decouples MLLMs' capabilities to faithfully assess VTC quality. Within this framework, we further introduce the ZeroSense Benchmark to ensure low semantic correlation of testing samples. By eliminating textual dependencies, our benchmark guarantees that the evaluation results are purely reflective of VTC quality, unaffected by the semantic inference capabilities of downstream models. Extensive experiments across multiple datasets demonstrate that VTC quality and downstream task accuracy diverge significantly, highlighting the necessity of our decoupled evaluation framework.

Figures

Figures reproduced from arXiv: 2608.01848 by Jingwei Guan, Jingzhi Chen, Lijian Xu, Xingyu Zeng, Yonghan Gao, Zehong Chen.

Figure 1
Figure 1. Figure 1: Semantic confounds in evaluating visual-text compression. (a) Semantic Priors Compensation: The model [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: This overview illustrates our proposed evaluation framework. ZeroSense constructs low-predictability texts [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: ZeroSense text construction pipeline. Low-probability words are sampled to form long sequences, which are [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] 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

40 extracted references · 5 linked inside Pith

  1. [1]

    Long context is not long at all: A prospector of long-dependency data for large language models

    Longze Chen, Ziqiang Liu, Wanwei He, Yinhe Zheng, Hao Sun, Yunshui Li, Run Luo, and Min Yang. Long context is not long at all: A prospector of long-dependency data for large language models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8222–8234, Bangkok, Thailand, August 2024. As...

    2024

  2. [2]

    Longlora: Efficient fine-tuning of long-context large language models, 2024

    Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhijian Liu, Song Han, and Jiaya Jia. Longlora: Efficient fine-tuning of long-context large language models, 2024

    2024

  3. [3]

    Minimax-m1: Scaling test-time compute efficiently with lightning attention.arXiv preprint arXiv:2506.13585, 2025

    Aili Chen, Aonian Li, Bangwei Gong, et al. Minimax-m1: Scaling test-time compute efficiently with lightning attention.arXiv preprint arXiv:2506.13585, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Deepseek-ocr: Contexts optical compression.arXiv preprint arXiv:2510.18234, 2025

    Haoran Wei, Yaofeng Sun, and Yukun Li. Deepseek-ocr: Contexts optical compression.arXiv preprint arXiv:2510.18234, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Glyph: Scaling context windows via visual-text compression

    Jiale Cheng, Yusen Liu, Xinyu Zhang, et al. Glyph: Scaling context windows via visual-text compression. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 37145–37158, 2026

    2026

  6. [6]

    Vtc-r1: Vision-text compression for efficient long-context reasoning, 2026

    Yibo Wang, Yongcheng Jing, Shunyu Liu, Hao Guan, Rong cheng Tu, Chengyu Wang, Jun Huang, and Dacheng Tao. Vtc-r1: Vision-text compression for efficient long-context reasoning, 2026

    2026

  7. [7]

    Focus anywhere for fine-grained multi-page document understanding, 2024

    Chenglong Liu, Haoran Wei, Jinyue Chen, Lingyu Kong, Zheng Ge, Zining Zhu, Liang Zhao, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Focus anywhere for fine-grained multi-page document understanding, 2024

    2024

  8. [8]

    Omnidocbench: Benchmarking diverse pdf document parsing with comprehensive annotations

    Linke Ouyang, Yuan Qu, Hongbin Zhou, et al. Omnidocbench: Benchmarking diverse pdf document parsing with comprehensive annotations. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24838–24848, June 2025

    2025

  9. [9]

    Ocr-free document understanding transformer

    Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. InEuropean Conference on Computer Vision, pages 498–517. Springer, 2022

    2022

  10. [10]

    Language modelling with pixels

    Phillip Rust, Jonas F Lotz, Emanuele Bugliarello, Elizabeth Salesky, Miryam de Lhoneux, and Desmond Elliott. Language modelling with pixels. InThe Eleventh International Conference on Learning Representations, pages 1–32. OpenReview, 2023

    2023

  11. [11]

    Pix2struct: Screenshot parsing as pretraining for visual language understanding

    Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. InInternational Conference on Machine Learning, pages 18893–18912. PMLR, 2023

    2023

  12. [12]

    Multilingual pretraining for pixel language models

    Ilker Kesen, Jonas F Lotz, Ingo Ziegler, Phillip Rust, and Desmond Elliott. Multilingual pretraining for pixel language models. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 29582–29599, 2025

    2025

  13. [13]

    See the text: From tokenization to visual reading.ArXiv, abs/2510.18840, 2025

    Ling Xing, Alex Wang, Rui Yan, Hongyu Qu, Zechao Li, and Jinhui Tang. See the text: From tokenization to visual reading.ArXiv, abs/2510.18840, 2025

    arXiv 2025

  14. [14]

    Leveraging visual tokens for extended text contexts in multi-modal learning.Advances in Neural Information Processing Systems, 37:14325–14348, 2024

    Alex Jinpeng Wang, Linjie Li, Yiqi Lin, Min Li, Lijuan Wang, and Mike Zheng Shou. Leveraging visual tokens for extended text contexts in multi-modal learning.Advances in Neural Information Processing Systems, 37:14325–14348, 2024

    2024

  15. [15]

    From text to pixel: Advancing long-context understanding in mllms, 2024

    Yujie Lu, Xiujun Li, Tsu-Jui Fu, Miguel Eckstein, and William Yang Wang. From text to pixel: Advancing long-context understanding in mllms, 2024

    2024

  16. [16]

    Vision-centric token compression in large language model.Advances in Neural Information Processing Systems, 38:33080–33110, 2026

    Ling Xing, Alex Jinpeng Wang, Rui Yan, Xiangbo Shu, and Jinhui Tang. Vision-centric token compression in large language model.Advances in Neural Information Processing Systems, 38:33080–33110, 2026

    2026

  17. [17]

    Text or pixels? evaluating efficiency and understanding of LLMs with visual text inputs

    Yanhong Li, Zixuan Lan, and Jiawei Zhou. Text or pixels? evaluating efficiency and understanding of LLMs with visual text inputs. InFindings of the Association for Computational Linguistics: EMNLP 2025, pages 10564–10578, Suzhou, China, November 2025. Association for Computational Linguistics

    2025

  18. [18]

    Global context compression with interleaved vision-text transformation, 2026

    Dian Jiao, Jiaxin Duan, Shuai Zhao, Jiabing Leng, Yiran Zhang, and Feng Huang. Global context compression with interleaved vision-text transformation, 2026

    2026

  19. [19]

    Render-of-thought: Rendering textual chain-of-thought as images for visual latent reasoning

    Yifan Wang, Shiyu Li, Peiming Li, Xiaochen Yang, Zheng Wei, and Yang Tang. Render-of-thought: Rendering textual chain-of-thought as images for visual latent reasoning. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 45236–45253, San Diego, California, United States, July 2026. Associ...

    2026

  20. [20]

    Lensvlm: Selective context expansion for compressed visual representation of text, 2026

    Roy Xie, Dan Friedman, Donghan Yu, Bowen Pan, Christopher Fifty, Jang-Hyun Kim, Xianzhi Du, Zhe Gan, Vivek Rathod, and Bhuwan Dhingra. Lensvlm: Selective context expansion for compressed visual representation of text, 2026

    2026

  21. [21]

    Zerosense: How vision matters in long context compression.arXiv preprint arXiv:2603.11846, 2026

    Yonghan Gao, Zehong Chen, Lijian Xu, Jingzhi Chen, Jingwei Guan, and Xingyu Zeng. Zerosense: How vision matters in long context compression.arXiv preprint arXiv:2603.11846, 2026

    arXiv 2026

  22. [22]

    One leaf reveals the season: Occlusion-based contrastive learning with semantic-aware views for efficient visual representation

    Xiaoyu Yang, Lijian Xu, Hongsheng Li, and Shaoting Zhang. One leaf reveals the season: Occlusion-based contrastive learning with semantic-aware views for efficient visual representation. InInternational Conference on Machine Learning, pages 71425–71440, 2025

    2025

  23. [23]

    Scalar: Spatial- concept alignment for robust vision in harsh open world.Pattern Recognition, page 113203, 2026

    Xiaoyu Yang, Lijian Xu, Xingyu Zeng, Xiaosong Wang, Hongsheng Li, and Shaoting Zhang. Scalar: Spatial- concept alignment for robust vision in harsh open world.Pattern Recognition, page 113203, 2026

    2026

  24. [24]

    Beyond surrogate gradients: Fully differentiable token pruning for vision-language models.arXiv preprint arXiv:2605.28051, 2026

    Landi He, Mingde Yao, Shawn Young, and Lijian Xu. Beyond surrogate gradients: Fully differentiable token pruning for vision-language models.arXiv preprint arXiv:2605.28051, 2026

    Pith/arXiv arXiv 2026

  25. [25]

    The model knows which tokens matter:automatic token selection via noise gating.arXiv preprint arXiv:2603.07135, 2026

    Landi He, Xiaoyu Yang, and Lijian Xu. The model knows which tokens matter:automatic token selection via noise gating.arXiv preprint arXiv:2603.07135, 2026

    arXiv 2026

  26. [26]

    Stepwise token selection for efficient multimodal large language models

    Landi He, Shawn Young, and Lijian Xu. Stepwise token selection for efficient multimodal large language models. WAICA, 2026

    2026

  27. [27]

    Fewer tokens, greater scaling: Self-adaptive visual bases for efficient and expansive representation learning.arXiv preprint arXiv:2511.19515, 2026

    Shawn Young, Xingyu Zeng, and Lijian Xu. Fewer tokens, greater scaling: Self-adaptive visual bases for efficient and expansive representation learning.arXiv preprint arXiv:2511.19515, 2026

    arXiv 2026

  28. [28]

    Ruler: What’s the real context size of your long-context language models?, 2024

    Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, and Boris Ginsburg. Ruler: What’s the real context size of your long-context language models?, 2024

    2024

  29. [29]

    Longbench: A bilingual, multitask benchmark for long context understanding.Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 2024

    Yushi Bai, Xin Lv, Jiajie Zhang, Hongchang Lyu, Jiankai Tang, Zhidian Huang, Zhengxiao Du, Xiao Liu, Aohan Zeng, Lei Hou, et al. Longbench: A bilingual, multitask benchmark for long context understanding.Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 2024

    2024

  30. [30]

    Vtcbench: Can vision-language models understand long context with vision-text compression?, 2025

    Hongbo Zhao, Meng Wang, Fei Zhu, Wenzhuo Liu, Bolin Ni, Fanhu Zeng, Gaofeng Meng, and Zhaoxiang Zhang. Vtcbench: Can vision-language models understand long context with vision-text compression?, 2025

    2025

  31. [31]

    Fico: Evaluating vision- language models under visual fidelity and compression at scale

    Jianhong Tu, Nicholas Crispino, Kyle Montgomery, Chenguang Wang, and Dawn Song. Fico: Evaluating vision- language models under visual fidelity and compression at scale. InFindings of the Association for Computational Linguistics: ACL 2026, pages 35261–35277, San Diego, California, United States, July 2026. Association for Computational Linguistics

    2026

  32. [32]

    Visual merit or linguistic crutch? a close look at deepseek-ocr, 2026

    Yunhao Liang, Ruixuan Ying, Bo Li, Hong Li, Kai Yan, Qingwen Li, Min Yang, Okamoto Satoshi, Zhe Cui, and Shiwen Ni. Visual merit or linguistic crutch? a close look at deepseek-ocr, 2026

    2026

  33. [33]

    Optical context compression is just (bad) autoencoding, 2026

    Ivan Yee Lee, Cheng Yang, and Taylor Berg-Kirkpatrick. Optical context compression is just (bad) autoencoding, 2026

    2026

  34. [34]

    Visual text compression as measure transport, 2026

    Lv Tang, Tianyi Zheng, Yang Liu, Bo Li, and Xingyu Li. Visual text compression as measure transport, 2026

    2026

  35. [35]

    Ocrbench: on the hidden mystery of ocr in large multimodal models.Science China Information Sciences, 67, 2023

    Yuliang Liu, Zhang Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xucheng Yin, Cheng-Lin Liu, Lianwen Jin, and Xiang Bai. Ocrbench: on the hidden mystery of ocr in large multimodal models.Science China Information Sciences, 67, 2023

    2023

  36. [36]

    Revisiting the role of language priors in vision-language models

    Zhiqiu Lin, Xinyue Chen, Deepak Pathak, Pengchuan Zhang, and Deva Ramanan. Revisiting the role of language priors in vision-language models. InProceedings of the 41st International Conference on Machine Learning, volume 235 ofProceedings of Machine Learning Research, pages 29914–29934. PMLR, 21–27 Jul 2024

    2024

  37. [37]

    Tokenization falling short: On subword robustness in large language models

    Yekun Chai, Yewei Fang, Qiwei Peng, and Xuhong Li. Tokenization falling short: On subword robustness in large language models. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 1582–1599, Miami, Florida, USA, November 2024. Association for Computational Linguistics

    2024

  38. [38]

    General ocr theory: Towards ocr-2.0 via a unified end-to-end model.arXiv preprint arXiv:2409.01704, 2024

    Haoran Wei, Chenglong Liu, Jinyue Chen, Jia Wang, Lingyu Kong, Yanming Xu, Zheng Ge, Liang Zhao, Jianjian Sun, Yuang Peng, et al. General ocr theory: Towards ocr-2.0 via a unified end-to-end model.arXiv preprint arXiv:2409.01704, 2024

    Pith/arXiv arXiv 2024

  39. [39]

    Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

    Shuai Bai, Keqin Chen, Xuejing Liu, et al. Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

    Pith/arXiv arXiv 2025

  40. [40]

    Hunyuanocr technical report.arXiv preprint arXiv:2511.19575, 2025

    Hunyuan Vision Team, Pengyuan Lyu, Xingyu Wan, Gengluo Li, Shangpin Peng, Weinong Wang, Liang Wu, et al. Hunyuanocr technical report.arXiv preprint arXiv:2511.19575, 2025. 12 APREPRINT- AUGUST4, 2026 Figure A1: Comparison between original and ZeroSense document images.Top:Full page views demonstrate that our generation pipeline perfectly preserves the doc...

    arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.