1 / 4
先に読む文献
ステップは読む順序です。研究上の因果関係や、問題の解決を意味しません。
RESEARCH / 65 PAPERS · 4 THEMES
Where the questions came from.
How far they have got.
Same text, different look — does a machine still read the same structure? Four questions, followed from the literature to my own experiments.
- RAG
- Retrieve documents, then answer using their content as evidence.
- evidence
- Which document and which passage an answer rests on.
- status
- Information beyond the words, such as whether a rule is current or superseded.
Theme 1 has development observations. Themes 2 and 3 are in preparation: images exist, but there are no model results. Theme 4 remains a candidate and its experiment has not started. There are no formal conclusions.
Read the research note
I study where necessary evidence can be lost when a system answers from documents. Retrieval-augmented generation, or RAG, finds relevant material and uses it to produce an answer. Documents mix current rules, obsolete clauses and exceptions. Reading the words is not enough if their validity or conditions are lost. My aim is to vary state representations or typography while keeping the wording fixed, and examine retrieval, reading and evidence use separately.
- CandidateKept as a question; no empirical work started.
- In preparationDesign or inputs exist, but no usable observation for this question yet.
- Dev. observationsLimited development results or analysis of saved outputs; not an independent formal validation.
- Formal resultPassed formal review. Currently 0.
研究マップを読み込んでいます…
研究マップを読み込んでいます…
01 上流の文献・関連
02 下流の文献・関連
03 つながる研究の問い
関係の根拠・限界・確かさ
出典と表示に使った情報
02Candidate themes
Per theme: progress, papers, experiments, my partTHEME 01 / 31 PAPERS · 5 Research questions
Semantic & Structural Evidence Fidelity in Text/Structure RAG
- Question
- When a superseded rule and the current rule sit in the same document, can retrieval pick the current one?
- Now
- 100 synthetic guidelines, 500 chunks, comparing kept vs. flattened status marks. Small development observations exist; no formal result yet.
A guideline says "documents due June 10 (formerly June 4)". Ask "this year's deadline?" — if June 4 ranks higher, the answer is wrong despite having a source.
Evidence does not fit on one page.
FIG. 03 / CROSS-PAGEFind the reference in the text."Appendix B" on p.04 points to the next place to read.
Read the figure as text
- Page 4 of the text says "see Appendix B".
- The target is page 9; page number and reference name shown together.
- Read the table and note on page 9.
- The relation between p.04 and p.09 is kept as a traceable path to the source.
A hand-made explanatory animation. No autoplay; play, pause and frame-by-frame. Not a model run or an experimental result. Sample text is Japanese. The real screen: J-Grad’s "back to the evidence" →
Read the research note
Why Reading the words correctly is not enough; lose "which one is valid" and you get a plausible wrong answer.
Investigate whether correct states and relationships remain intact through representation, chunking, and retrieval. Limited development-stage observations are available. (The Japanese title was translated by R.)
- ○ P021An Approach of Strike-Through Text Identification from Handwritten Documents 2014取り消し線で消された文字を分けて認識する研究。根拠として使えるかは未検証。
- ○ P026A study of handwritten text recognition with cross-out words 2026取り消し線の見た目が文字認識に与える影響を比較。正しく読めても、有効な根拠とは限らない。
- ◇ T1-RQ1给定正确状态时，保留/抹平是否改变有效/失效证据的检索排名？無効な状態を残すと検索順位が変わるかを問う。この段階では正式検証前。
- □ E06state-ablation / attempt-005状態情報を残す場合と省く場合を小規模に比較。この時点では開発段階の観察。
- ○ P022WebSRC: A Dataset for Web-Based Structural Reading Comprehension 2021Webページの構造を含めて質問応答を評価する研究。行内の状態情報を使う検索までは扱っていない。
- ○ P024MarkupLM: Pre-training of Text and Markup Language for Visually Rich Document Understanding 2022マークアップ構造をモデルの表現に取り込む研究。行内の状態情報が検索に役立つかは未検証。
- ○ P025Intelligent Predictive Maintenance RAG framework for Power Plants: Enhancing QA with StyleDFS and Domain Specific Instruction Tuning 2024文書のブロック階層を、分割や質問応答に使う研究。より細かな意味関係は未検証。
- ◇ T1-RQ4切块后，限定/引用/状态与正文的支持关系是否完整保留并被检出？分割後も、条件・参照・状態の関係を漏れなく検索できるかを問う。未検証。
- □ E11oracle/关系来源见证文脈全体で関係を確認した資料はある。関係を検索できるかは未検証。
Direction & progress
Theme 1 starts with correctly extracted text and known rule states. It asks whether those states and relationships survive representation, chunking and retrieval, so that current rules can be distinguished from obsolete ones and necessary conditions remain complete. Small candidate sets provide development observations; full-collection effects and cross-chunk completeness remain unverified.
Papers
Start with ParseBench to separate text and semantic-format evaluation, then the power-plant paper containing StyleDFS for the structure-to-retrieval pipeline. Next, read the cross-out handwriting study, which remains at the recognition level: readable text is not necessarily valid evidence. MarkupLM and FontHalu provide additional context.
Experiments & reports
Small development comparisons cover preserving, removing and swapping state information. Analysis of saved outputs separates candidate availability, state eligibility and complete evidence selection. These observations do not establish better full-collection retrieval, generalization to unseen documents or cross-chunk completeness.
- ◇ T1-RQ1 给定正确状态时，保留/抹平是否改变有效/失效证据的检索排名？Dev. observations □ E01 E02 E03 E04 E06 E08 E10
- ◇ T1-RQ2 状态交换与等义表达后，模型是否跟随问题要求的状态，而非固定正文/数值？Dev. observations □ E04 E07 E09
- ◇ T1-RQ3 如何区分状态资格、正文相关性与0/1/2条完整选证，解释误收/漏答？Dev. observations □ E05 E06 E07 E09
- ◇ T1-RQ4 切块后，限定/引用/状态与正文的支持关系是否完整保留并被检出？In preparation □ E11
- ◇ T1-RQ5 现行、失效、日期资格与直接上一版应怎样分别定义？Dev. observations □ E12
My part
The agreed division gives the user direction and disclosure decisions, with artificial intelligence assisting literature organization and saved-output analysis. Reports do not fully identify who completed each design, implementation and scoring task; these are not presented as the user's solo work.
THEME 02 / 25 PAPERS · 2 Research questions
Mechanistic Decomposition of Visual-Text Failure
- Question
- Shown the same text as an image, did the model fail to read the characters, or read them and still answer wrong?
- Now
- 10 cases, 20 images, 50 inputs prepared. The model has not been run.
Give the same paragraph as text and as an image; if answers differ, ask only for a transcription to check whether it could read it.
Read the research note
Why Without separating the two, you cannot tell whether to fix OCR or reasoning.
Investigate the difference between failing to read text and answering incorrectly after reading it correctly. Inputs are prepared, but there are no model results yet. (The Japanese title was translated by R.)
- ○ P008OCRBench: on the hidden mystery of OCR in large multimodal models 2024画像内の文字の読み取りを切り出して評価する研究。書体の変化は統制していない。
- ○ P009Exploring the Capabilities of Large Multimodal Models on Dense Text 2024文字の多い画像で質問応答を評価する研究。同じ内容のテキスト版との比較はない。
- ○ P011ReadBench: Measuring the Dense Text Visual Reading Ability of Vision-Language Models 2025同じ内容のテキストと画像で理解の差を比べる研究。差の原因はまだ切り分ける必要がある。
- ○ P012VISTA-Bench: Do Vision-Language Models Really Understand Visualized Text as Well as Pure Text? 2026内容をそろえたまま描画条件を変える研究。生じた差がどの誤りに由来するかは要検討。
- ○ P013Reading, Not Thinking: Understanding and Bridging the Modality Gap When Text Becomes Pixels in Multimodal LLMs 2026文字の読み取り・計算・推論の誤りを分ける研究。読み取りと回答の誤りの量的な関係は未解明。
- ◇ T2-RQ1读字失败与回答失败的交叉关系是什么？文字の読み取りの失敗が、回答の誤りにどうつながるかを問う。モデルの回答はまだない。
- □ E13T2视觉诊断开发入力画像と質問は準備済み。モデルの結果はまだ得られていない。
Direction & progress
Theme 2 is a supporting diagnostic theme. It separates failure to read the words from answering incorrectly after reading them. The plan compares matched text and image inputs while controlling typography and visual processing cost. Inputs have been prepared, but there are no model answers. Its role is to explain failures rather than lead a broad new-method project.
Papers
Read ReadBench for matched text-image comparisons, then VISTA-Bench for rendering differences, followed by Reading, Not Thinking for failure categories. Glyph adds background on why visual processing cost also needs control.
Experiments & reports
Inputs include the same text at normal and smaller sizes, with answer and reading-check prompts. Automated checks cover glyph coverage and overflow; human readability has not been attested. No model has run, so there are no reading or answering scores.
- ◇ T2-RQ1 读字失败与回答失败的交叉关系是什么？In preparation □ E13
- ◇ T2-RQ2 控制视觉预算后，字体/渲染扰动的残差是否仍存在？In preparation □ E13
My part
Artificial intelligence organized the literature map and preparation status. These reports do not establish who specifically generated the images or implemented the preparation. The user's completed design, implementation and evaluation work requires task-level records.
THEME 03 / 29 PAPERS · 2 Research questions
Page-Level Visual Retrieval Stability under Typography
- Question
- Same page, typeface changed and nothing else — does the retrieval rank move?
- Now
- 10 pages rendered in 3 typefaces. The rank comparison is next.
Render the same "eligibility" passage in Mincho, Gothic and UD, then search with the same question.
Read the research note
Why To check whether a typeface difference unrelated to content affects retrieval stability. Corporate documents vary in typeface by author.
Investigate whether changing the typography of the same page affects retrieval stability. Images are prepared, but there are no retrieval results yet. (The Japanese title was translated by R.)
- ○ P038Unifying Multimodal Retrieval via Document Screenshot Embedding 2024ページのスクリーンショットを使う検索の研究。書体の変化は系統的に統制していない。
- ◇ T3-RQ1同页换字体是否改变检索score/rank/Top-K保留？同じページの書体を変えると順位が変わるかを問う。検索結果はまだない。
- □ E14T3字体检索开发この段階では、ページと入力の準備まで。埋め込みや順位の結果はまだない。
Direction & progress
Theme 3 asks whether changing only the font on a page changes retrieval scores and ranks. Content, line breaks, layout, readability and visual processing cost are to be matched as closely as possible. The plan compares whole-page representations with finer-grained matching. Only images and inputs are ready; there are no retrieval results. Little or no change would also be recorded with its conditions.
Papers
Start with Document Screenshot Embedding for page-image retrieval, then ColPali for fine-grained matching, and Vision-Text Augmentation for visual presentation variables. This is a reading order, not an assertion of undocumented lineage between the papers.
Experiments & reports
Pages with matched content in multiple fonts and their retrieval inputs are prepared, with shared line breaks and positions. Actual model processing cost has not been measured, and no page encoding or ranking has run. Font effects and relative robustness between retrieval architectures remain unknown.
- ◇ T3-RQ1 同页换字体是否改变检索score/rank/Top-K保留？In preparation □ E14
- ◇ T3-RQ2 差异是否超过同输入噪声，dense与late-interaction是否不同？In preparation □ E14
My part
Artificial intelligence assisted with literature mapping and describing preparation status. Available material does not establish the implementation split between the user and artificial intelligence. Building a retrieval model or completing a comparative evaluation is not claimed as the user's achievement.
THEME 04 / 38 PAPERS · 2 Research questions
Passive Typography and Evidence Arbitration
- Question
- When every needed piece of evidence is readable, does the model pick the bolder, larger one as "the evidence"?
- Now
- Candidate theme. No experiment started, no observations yet.
Set the main rule "TOEIC 600+" in bold and the exception "Mathematics: 700" in small type. The comparison: when size and weight change, does the answer for a maths applicant change too? Not yet run (illustrative example).
Read the research note
Why Exceptions tend to live in small notes. A system that grabs the prominent main rule answers wrong for exactly the people the exception covers.
Investigate whether passive typographic prominence changes reliance on evidence. The experiment has not started. (The Japanese title was translated by R.)
- ○ P047Magnifying What Matters: Attention-Guided Adaptive Rendering for Visual Text Comprehension 2026選んだ部分を拡大して読解を助ける研究。文書にもともとある目立ち方の影響は未検証。
- ◇ T4-RQ1被动显著性是否改变规则/例外证据依赖与答案？もともとの文字の目立ち方で、規則と例外の使い方が変わるかを問う。実験は未着手。
- □ E15T4被动显著性文書にもともとある目立ち方を比べる実験案。観察や結論はまだない。
Direction & progress
Theme 4 remains a candidate: when all evidence is readable, can passive visual prominence change which rules, conditions or exceptions are used? This differs from locating and enlarging evidence in response to a question. Content, position and visual cost need control, and attention must be distinguished from actual reliance. No experiment or observation has started.
Papers
Read Seeing but Not Believing to distinguish attending to evidence from answering correctly. Then read Magnifying What Matters for active enlargement, followed by DocAttriBench for examining reliance by masking evidence.
Experiments & reports
The experiment has not started. There is only a proposal to swap the prominence of rules and exceptions and examine answers and evidence reliance. Controls for readability and visual processing cost, and the reliance measurement itself, are not yet established. Neither an effect nor the absence of one can be claimed.
- ◇ T4-RQ1 被动显著性是否改变规则/例外证据依赖与答案？Candidate □ E15
- ◇ T4-RQ2 遮蔽反事实能否区分实际必要证据依赖与仅注意到/读到？Candidate □ E15
My part
Artificial intelligence organized prior work, the candidate question and unperformed work. Experimentation and evaluation have not started. Available material does not identify which specific design tasks the user has completed; additional work records are needed.
65Papers
Bibliographic data only; no PDFs or figures. Reading and verification status are tracked separately.| ID | Title | Authors | Year | Themes | |
|---|---|---|---|---|---|
| P001 | A Study of Style Effects on OCR Errors in the MEDLINE Database · Proceedings of SPIE-IS&T Electronic Imaging, Vol. 5676 Notes on the relationship between font styles and text recognition errors in medical literature. | Penny Garrison; Diane Davis; Tim Andersen; Elisa Barney Smith | 2005 | T1 T2 | [DOI] →1 |
| P002 | Text Degradations and OCR Training · Eighth International Conference on Document Analysis and Recognition, ICDAR 2005 Notes on how degradation from printing and scanning affects training for text recognition. | Elisa H. Barney Smith; Tim Andersen | 2005 | T1 T2 | [DOI] |
| P003 | A Study on the Accuracy of OCR Engines for Source Code Transcription from Programming Screencasts · 17th International Conference on Mining Software Repositories, MSR 2020 Notes comparing fonts, font sizes, and text recognition engines for screenshots of source code. | Abdulkarim Khormi; Mohammad Alahmadi; Sonia Haiduc | 2020 | T1 T2 | [DOI] →2 |
| P004 | Texture or Semantics? Vision-Language Models Get Lost in Font Recognition · Conference on Language Modeling, COLM 2025 Notes on how vision-language models distinguish texture from textual meaning when recognizing fonts. | Zhecheng Li; Guoxian Song; Yujun Cai; Zhen Xiong; Junsong Yuan; Yiwei Wang | 2025 | T1 T2 T3 T4 | [arXiv] →1 |
| P005 | OmniDocBench: Benchmarking Diverse PDF Document Parsing with Comprehensive Annotations · IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2025；材料给出的论文页码为 24838–24848（待原文核实）。 Notes on evaluation and annotation issues across PDF document parsing tasks. | Linke Ouyang（仅材料提供的第一作者；完整作者名单未提供） | 2025 | T1 T2 T4 | [DOI] →1 |
| P006 | How does Watermarking Affect Visual Language Models in Document Understanding? · Conference on Language Modeling, COLM 2025 Notes on evaluating document understanding in the presence of watermarks. | Chunxue Xu; Yiwei Wang; Bryan Hooi; Yujun Cai; Songze Li | 2025 | T1 T2 T3 T4 | [arXiv] |
| P007 | When Good OCR Is Not Enough: Benchmarking OCR Robustness for Retrieval-Augmented Generation · ACL 2026, Industry Track Notes on how text recognition errors carry through to retrieval and answers. | Lin Sun; Wangdexian; Jingang Huang; Linglin Zhang; Change Jia; Zhengwei Cheng; Xiangzheng Zhang | 2026 | T1 T2 T4 | [DOI] |
| P008 | OCRBench: on the hidden mystery of OCR in large multimodal models · Science China Information Sciences, Volume 67, Article 220102 Notes on evaluating how multimodal models recognize text in images. | Yuliang Liu, Zhang Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xu-Cheng Yin, Cheng-Lin Liu, Lianwen Jin, Xiang Bai | 2024 | T2 T3 T4 | [DOI] →2 |
| P009 | Exploring the Capabilities of Large Multimodal Models on Dense Text · arXiv Notes on multimodal models' ability to read densely packed text. | Shuo Zhang, Biao Yang, Zhang Li, Zhiyin Ma, Yuliang Liu, Xiang Bai | 2024 | T2 T3 T4 | [arXiv] →1 |
| P010 | OCRBench v2: An Improved Benchmark for Evaluating Large Multimodal Models on Visual Text Localization and Reasoning · NeurIPS 2025, Datasets and Benchmarks Track Notes on extending evaluation to text localization and reasoning. | Ling Fu, Zhebin Kuang, Jiajun Song, Mingxin Huang, Biao Yang, Yuzhe Li, Linghao Zhu, Qidi Luo, Xinyu Wang, Hao Lu, Zhang Li, Guozhi Tang, Bin Shan, Chunhui Lin, Qi Liu, Binghong Wu, Hao Feng, Hao Liu, Can Huang, Jingqun Tang, Wei Chen, Lianwen Jin, Yuliang Liu, Xiang Bai | 2025 | T2 T3 T4 | [DOI] |
| P011 | ReadBench: Measuring the Dense Text Visual Reading Ability of Vision-Language Models · arXiv Notes on reading long, dense text presented as images. | Benjamin Clavié, Florian Brand | 2025 | T2 T3 T4 | [arXiv] →1 |
| P012 | VISTA-Bench: Do Vision-Language Models Really Understand Visualized Text as Well as Pure Text? · arXiv Notes on differences in understanding when the same content is supplied as an image or as text. | Qing'an Liu, Juntong Feng, Yuhao Wang, Xinzhe Han, Yujie Cheng, Yue Zhu, Haiwen Diao, Yunzhi Zhuge, Huchuan Lu | 2026 | T2 T3 T4 | [arXiv] →1 |
| P013 | Reading, Not Thinking: Understanding and Bridging the Modality Gap When Text Becomes Pixels in Multimodal LLMs · arXiv Notes on differences between reading and reasoning abilities when text is rendered as pixels. | Kaiser Sun, Xiaochuang Yuan, Hongjun Liu, Chen Zhao, Cheng Zhang, Mark Dredze, Fan Bai | 2026 | T2 T3 T4 | [arXiv] |
| P014 | OCR Hinders RAG: Evaluating the Cascading Impact of OCR on Retrieval-Augmented Generation · ICCV 2025 Notes on the cascading effects of text recognition errors on retrieval and answers. | Junyuan Zhang, Qintong Zhang, Bin Wang, Linke Ouyang, Zichen Wen, Ying Li, Ka-Ho Chow, Conghui He, Wentao Zhang | 2025 | T2 T3 T4 | |
| P015 | FADE: Probing the Limits of VLMs on fine-grained OCR · 4th Workshop on Advances in Language and Vision Research (ALVR), ACL 2026 Notes on the limits of fine-grained text recognition. | Deep Shah, Nehal Kathrotia, Sanket Badhe | 2026 | T2 T3 T4 | [DOI] |
| P016 | Reading ≠ Seeing: Diagnosing and Closing the Typography Gap in Vision-Language Models · arXiv Notes on the gap between font perception and reading in vision-language models. | Heng Zhou, Ao Yu, Li Kang, Yuchen Fan, Yutao Fan, Xiufeng Song, Hejia Geng, Yiran Qin | 2026 | T2 T3 T4 | [arXiv] |
| P017 | Glyph: Scaling Context Windows via Visual-Text Compression · ACL 2026, Main Conference, Long Papers Notes on compressing inputs by rendering long text as images. | Jiale Cheng, Yusen Liu, Xinyu Zhang, Yulin Fei, Wenyi Hong, Ruiliang Lyu, Weihan Wang, Zhe Su, Xiaotao Gu, Xiao Liu, Yushi Bai, Jie Tang, Hongning Wang, Minlie Huang | 2026 | T2 T3 T4 | [DOI] →2 |
| P018 | FocusVTC: Efficient and High-Performance Visual Text Compression with Adaptive Resolution · arXiv Notes on compressing text in images through adaptive resolution. | FangZhi Zhong, Xuerui Qiu, Yuqi Pan, Ya Liu, Shaowei Gu, Bo Xu, Guoqi Li | 2026 | T2 T3 T4 | [arXiv] |
| P019 | Same Semantics, Different Outcome: On the Modality Robustness of Multimodal LLMs under Knowledge Conflict · Findings of EMNLP 2026 Notes on consistency across input modalities for the same content when knowledge conflicts arise. | Jungyeon Lee, Yejin Yoon, Taeuk Kim | 2026 | T2 T3 T4 | [arXiv] |
| P020 | Identifying Subscripts and Superscripts in Mathematical Documents · Mathematics in Computer Science Notes on recognizing superscripts and subscripts in mathematical documents. | Walaa Aly, Seiichi Uchida, Masakazu Suzuki | 2008 | T1 T4 | [DOI] |
| P021 | An Approach of Strike-Through Text Identification from Handwritten Documents · 14th International Conference on Frontiers in Handwriting Recognition, ICFHR 2014；材料给定论文整体出版页码为 643–648（待原文核实） Notes on recognizing struck-through text in handwritten documents. | Chandranath Adak, Bidyut B. Chaudhuri | 2014 | T1 T4 | [DOI] →1 |
| P022 | WebSRC: A Dataset for Web-Based Structural Reading Comprehension · EMNLP 2021 Notes on data and tasks for reading comprehension of webpage structure. | Xingyu Chen, Zihan Zhao, Lu Chen, Jiabao Ji, Danyang Zhang, Ao Luo, Yuxuan Xiong, Kai Yu | 2021 | T1 T4 | [DOI] →1 |
| P023 | Information Extraction from Visually Rich Documents with Font Style Embeddings · 26th International Conference on Pattern Recognition, ICPR 2022 Notes on the role of font-style representations in extracting information from documents. | Ismail Oussaid, William Vanhuffel, Pirashanth Ratnamogan, Mhamed Hajaiej, Alexis Mathey, Thomas Gilles | 2022 | T1 T4 | [DOI] [arXiv] |
| P024 | MarkupLM: Pre-training of Text and Markup Language for Visually Rich Document Understanding · ACL 2022 Notes on jointly modeling text and webpage markup. | Junlong Li, Yiheng Xu, Lei Cui, Furu Wei | 2022 | T1 T4 | [DOI] →1 |
| P025 | Intelligent Predictive Maintenance RAG framework for Power Plants: Enhancing QA with StyleDFS and Domain Specific Instruction Tuning · EMNLP 2024 Industry Track Notes on chunking by formatting hierarchy and question answering in power plant documents. | Seongtae Hong, Joongmin Shin, Jaehyung Seo, Taemin Lee, Jeongbae Park, Manyoung Cho, Byeongho Choi, Heuiseok Lim | 2024 | T1 T4 | [DOI] →1 |
| P026 | A study of handwritten text recognition with cross-out words · International Journal on Document Analysis and Recognition Notes on recognizing handwritten text that contains crossed-out words. | Gayan H. Pathirage, Simon Corbillé, Filip Wåhlin, Nazrul Islam, Elisa H. Barney Smith | 2026 | T1 T4 | [DOI] |
| P027 | FontHalu: Font-based Hallucinations in Multimodal Large Language Models · OpenReview public manuscript；正式 peer-reviewed venue 待核实。 Notes on evaluating misjudgments by multimodal models caused by font changes. | — | 2026（仅材料归档年份；具体 OpenReview submission / version date 待核实） | T1 T4 | |
| P028 | ParseBench: A Document Parsing Benchmark for AI Agents · arXiv preprint Notes on document parsing evaluation tasks for AI agents. | — | 2026 | T1 T4 | [arXiv] |
| P029 | Leveraging Visual Tokens for Extended Text Contexts in Multi-Modal Learning · NeurIPS 2024, Main Conference Track Notes on using visual text representations to extend text context. | Alex Jinpeng Wang, Linjie Li, Yiqi Lin, Min Li, Lijuan Wang, Mike Zheng Shou | 2024 | T2 T3 T4 | [DOI] →2 |
| P030 | Vision-centric Token Compression in Large Language Model · NeurIPS 2025 Main Conference Notes on compressing language model inputs through visual representations. | Ling Xing, Alex Jinpeng Wang, Rui Yan, Xiangbo Shu, Jinhui Tang | 2025 | T2 T3 T4 | [DOI] →2 |
| P031 | DeepSeek-OCR: Contexts Optical Compression · arXiv / DeepSeek technical release Notes on compression by converting text context into visual representations. | Haoran Wei, Yaofeng Sun, Yukun Li | 2025 | T2 T3 T4 | [DOI] [arXiv] →3 |
| P032 | VTCBench: Can Vision-Language Models Understand Long Context with Vision-Text Compression? · arXiv；作者材料标注 EMNLP 2026，来源材料未核实正式 proceedings 页面，本轮亦未独立核验。 Notes on evaluating long-context understanding after visual text compression. | Hongbo Zhao, Meng Wang, Fei Zhu, Wenzhuo Liu, Bolin Ni, Fanhu Zeng, Gaofeng Meng, Zhaoxiang Zhang | 2025 | T2 T3 T4 | [arXiv] →1 |
| P033 | Fico: Evaluating Vision-Language Models under Visual Fidelity and Compression at Scale · Findings of ACL 2026 Notes on model evaluation under visual fidelity and compression conditions. | Jianhong Tu, Nicholas Crispino, Kyle Montgomery, Chenguang Wang, Dawn Song | 2026 | T2 T3 T4 | [DOI] →1 |
| P034 | Visual Text Compression as Measure Transport · NeurIPS 2026 accepted（来源材料称官方 event / download list 已列出对应标题；实验细节主要依据 arXiv version） Literature note on representations and theoretical questions in visual text compression. | Lv Tang, Tianyi Zheng, Yang Liu, Bo Li, Xingyu Li | 2026 | T2 T3 T4 | [DOI] [arXiv] |
| P035 | Learning Transferable Visual Models From Natural Language Supervision · ICML 2021；Proceedings of Machine Learning Research, Volume 139 Literature note on learning comparable representations through joint image–text learning. | Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark; Gretchen Krueger; Ilya Sutskever | 2021 | T3 | |
| P036 | Disentangling Visual and Written Concepts in CLIP · CVPR 2022 Literature note on distinguishing visual concepts from concepts expressed in written text. | Joanna Materzyńska; Antonio Torralba; David Bau | 2022 | T3 | [DOI] |
| P037 | FontCLIP: A Semantic Typography Visual-Language Model for Multilingual Font Applications · Computer Graphics Forum, Volume 43, Issue 2 Literature note on font semantics and multilingual font representations. | Yuki Tatsukawa; I-Chao Shen; Anran Qi; Yuki Koyama; Takeo Igarashi; Ariel Shamir | 2024 | T3 | [DOI] |
| P038 | Unifying Multimodal Retrieval via Document Screenshot Embedding · EMNLP 2024 Literature note on multimodal retrieval using representations of document screenshots. | Xueguang Ma; Sheng-Chieh Lin; Minghan Li; Wenhu Chen; Jimmy Lin | 2024 | T3 | [DOI] |
| P039 | ColPali: Efficient Document Retrieval with Vision Language Models · ICLR 2025 Literature note on retrieving document pages with vision–language representations. | Manuel Faysse; Hugues Sibille; Tony Wu; Bilel Omrani; Gautier Viaud; Céline Hudelot; Pierre Colombo | 2025 | T3 | [arXiv] |
| P040 | VisRAG: Vision-based Retrieval-augmented Generation on Multi-modality Documents · ICLR 2025 Literature note on directly retrieving image evidence and using it to answer questions. | Shi Yu; Chaoyue Tang; Bokai Xu; Junbo Cui; Junhao Ran; Yukun Yan; Zhenghao Liu; Shuo Wang; Xu Han; Zhiyuan Liu; Maosong Sun | 2025 | T3 | [arXiv] |
| P041 | Vision-Text Augmentation: Visual Composition as an Optimization Axis for Language Tasks · OpenReview anonymous manuscript / ACL submission Literature note on how combinations of visual text styling and layout affect language tasks. | — | 2026 | T3 | |
| P042 | Enhancing Multimodal Product Retrieval in E-Commerce by Reversing Typographic Attacks · UniReps: The Third Edition of the Workshop on Unifying Representations in Neural Models Literature note on using typographic interference to improve product retrieval. | Janet Jenq; Hongda Shen | 2026 | T3 | |
| P043 | Reading Between the Pixels: Linking Text-Image Embedding Alignment to Typographic Attack Success on Vision-Language Models · ICLR 2026 Workshop on Agents in the Wild Literature note on the relationship between text–image representation alignment and typographic attacks. | Ravikumar Balakrishnan; Sanket Mendapara; Ankit Garg | 2026 | T3 | [arXiv] |
| P044 | DocVXQA: Context-Aware Visual Explanations for Document Question Answering · ICML 2025 Literature note on context-dependent visual explanations in document question answering. | Mohamed Ali Souibgui; Changkyu Choi; Andrey Barsky; Kangsoo Jung; Ernest Valveny; Dimosthenis Karatzas | 2025 | T4 | |
| P045 | Seeing but Not Believing: Probing the Disconnect Between Visual Attention and Answer Correctness in VLMs · ICLR 2026 Literature note on mismatches between visual attention and answer correctness. | — | 2026 | T4 | |
| P046 | SciEGQA: A Dataset for Scientific Evidence-Grounded Question Answering and Reasoning Literature note on data for grounding answers in evidence in scientific question answering. | — | 2026（来源登记年份；不据 arXiv 标识推定本卡具体实验版本或日期） | T4 | [arXiv] |
| P047 | Magnifying What Matters: Attention-Guided Adaptive Rendering for Visual Text Comprehension · NeurIPS 2026 Literature note on reading visual text through attention-guided rendering of selected regions. | Shenglai Zeng; Qirui Wang; Kai Guo; Xinnan Dai; Xianxuan Long; Hui Liu | 2026 | T4 | [arXiv] |
| P048 | Unveiling Inherent Visual Grounding in Multimodal LLMs for Text-Rich Images · Findings of ACL 2026 Literature note on how models locate visual evidence in text-rich images. | Shijie Zhou; Jihyung Kil; Ming Li; Jiuxiang Gu; Curtis Wigington; Rajiv Jain; Changyou Chen; Ruiyi Zhang | 2026 | T4 | [DOI] |
| P049 | DocVAL: Validated Chain-of-Thought Distillation for Grounded Document VQA · ICML 2026 Literature note on training evidence-grounded document question answering with verified reasoning traces. | Pinaki Prasad Guha Neogi; Ahmad Mohammadshirazi; Ser-Nam Lim; Rajiv Ramnath | 2026 | T4 | |
| P050 | DocAttriBench: Benchmarking Answer Grounding in Document Visual Question Answering · BMVC 2026 Literature note on evaluating whether document question answering outputs are actually supported by evidence. | Luca De Grandis; Silvia Cappelletti; William Raccagni; Marcella Cornia; Lorenzo Baraldi; Rita Cucchiara | 2026 | T4 | [arXiv] |
| P051 | Extracting the Latent Hierarchical Structure of Web Documents · Second International Conference on Signal-Image Technology and Internet-Based Systems（SITIS 2006）；后收入 Springer LNCS revised selected papers Literature note on recovering hidden logical hierarchies from the appearance of web pages. | Michael A. El-Shayeb; Samhaa R. El-Beltagy; Ahmed Rafea | 2006（来源登记的 SITIS conference 年；revised LNCS 实际出版年未知，不补猜） | T1 | [DOI] →2 |
| P052 | Logical Structure Recovery in Scholarly Articles with Rich Document Features · International Journal of Digital Library Systems, 1(4), 1–23 Literature note on recovering the logical structure of scholarly articles from document features. | Minh-Thang Luong; Thuy Dung Nguyen; Min-Yen Kan | 2010 | T1 | [DOI] |
| P053 | Structure Extraction from PDF-based Book Documents · 11th ACM/IEEE-CS Joint Conference on Digital Libraries（JCDL 2011）, pp. 11–20 Literature note on extracting document structure from PDF books. | Liangcai Gao et al.（来源未给完整作者名单，不补全） | 2011 | T1 | [DOI] |
| P054 | Extracting Logical Hierarchical Structure of HTML Documents Based on Headings · Proceedings of the VLDB Endowment, 8(12), 1606–1617 Literature note on recovering HTML document hierarchies from headings. | Tomohiro Manabe; Keishi Tajima | 2015 | T1 | [DOI] |
| P055 | A Machine-Learning Based Approach for Extracting Logical Structure of a Styled Document · KSII Transactions on Internet and Information Systems, 11(2), 1043–1056 Literature note on recovering the logical structure of styled documents with machine learning. | Tae-young Kim et al.（来源未给完整作者名单，不补全） | 2017 | T1 | [DOI] |
| P056 | DocParser: Hierarchical Document Structure Parsing from Renderings · AAAI 2021, 35(5), 4328–4338 Literature note on recovering document hierarchies from rendered pages. | Johannes Rausch; Octavio Martinez; Fabian Bissig; Ce Zhang; Stefan Feuerriegel | 2021 | T1 | [DOI] →1 |
| P057 | Extracting Variable-Depth Logical Document Hierarchy from Long Documents: Method, Evaluation, and Application · Journal of Computer Science and Technology, 37(3), 699–718 Literature note on recovering variable-depth hierarchies in long documents and their applications. | Rong-Yu Cao; Yi-Xuan Cao; Gan-Bin Zhou; Ping Luo | 2022 | T1 | [DOI] |
| P058 | HRDoc: Dataset and Baseline Method toward Hierarchical Reconstruction of Document Structures · AAAI 2023, 37(2), 1870–1877 Literature note on datasets and baseline methods for reconstructing document hierarchies. | — | 2023 | T1 | [DOI] →1 |
| P059 | DSG: An End-to-End Document Structure Generator · IEEE International Conference on Data Mining (ICDM 2023), 518–527 Literature note on end-to-end generation of document structure. | — | 2023 | T1 | [DOI] →1 |
| P060 | DocHieNet: A Large and Diverse Dataset for Document Hierarchy Parsing · EMNLP 2024 Main Conference, 1129–1142 Literature note on diverse data for parsing document hierarchies. | — | 2024 | T1 | [DOI] |
| P061 | Detect-Order-Construct: A Tree Construction Based Approach for Hierarchical Document Structure Analysis · Pattern Recognition, Volume 156, Article 110836 Literature note on recovering document hierarchies through detection, ordering, and tree construction. | Jiawei Wang; Kai Hu; Zhuoyao Zhong; Lei Sun; Qiang Huo | 2024 | T1 | [DOI] →1 |
| P062 | UniHDSA: A Unified Relation Prediction Approach for Hierarchical Document Structure Analysis · Pattern Recognition, Volume 165, Article 111617 Literature note on recovering document hierarchies through unified relation prediction. | Jiawei Wang; Kai Hu; Qiang Huo | 2025 | T1 | [DOI] |
| P063 | Graph-based Document Structure Analysis · International Conference on Learning Representations (ICLR 2025) Literature note on document structure analysis using graph representations. | Yufan Chen et al. | 2025 | T1 | |
| P064 | SuperRAG: Beyond RAG with Layout-Aware Graph Modeling · NAACL 2025 Industry Track, 544–557 Literature note on organizing retrieved evidence with layout-relation graphs. | Chening Yang et al. | 2025 | T1 | [DOI] →1 |
| P065 | LAD-RAG: Layout-aware Dynamic RAG for Visually-Rich Document Understanding · ACL 2026 Main Conference, Long Papers, 15945–15968 Literature note on cross-page relationships and dynamic evidence retrieval. | Zhivar Sourati et al. | 2026 | T1 | [DOI] |
Data: research-map/nodes.jsonl · edges.jsonl (compiled by R, as of 2026-10-24).