Pith. sign in

REVIEW 4 major objections 6 minor 36 references

AdaDocVQA: Adaptive Framework for Long Document Visual Question Answering in Low-Resource Settings

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read An adaptive pipeline of retrieval, self-generated QA data, and ensemble voting sets a new state of the art for Japanese document VQA.

desk verdict A coherent Japanese document VQA pipeline with plausible numbers, but the SOTA claim is underwritten: no zero-shot baseline on JDocQA, and the LAVA test set is too small to support the headline gaps. read the letter →

arxiv 2508.13606 v1 pith:IVGIQP76 submitted 2025-08-19 cs.CL

Haoxuan Li , Wei Song , Aofan Liu , Peiwu Qin This is my paper
classification cs.CL
keywords documentvisualquestionansweringlow-resourcelanguageJapaneseQAretrieval-augmentedgenerationdataaugmentationensembleinferencevision-languagemodelLoRAfine-tuning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that a particular combination of techniques—hybrid retrieval, automatically generated reasoning questions, and ensemble voting—can push a large vision-language model to state-of-the-art accuracy on Japanese document VQA, a low-resource setting with scarce training data and long documents. The authors build AdaDocVQA around a LoRA-fine-tuned Qwen2.5-VL-72B model and report 83.04% accuracy on yes/no questions, 52.66% on factual questions, and 44.12% on numerical questions on JDocQA, plus 59% accuracy on the LAVA benchmark. If these results hold, the recipe offers a template for adapting document VQA to other low-resource languages without new model architectures. The paper's central move is to treat data generation and passage retrieval as part of the model training and inference loop, not as preprocessing details.

What carries the argument

The load-bearing mechanism is a three-stage closed loop. First, a hybrid retriever fuses TF-IDF scores with multilingual-e5-large semantic similarity via $S_{\mathrm{final}}(d) = \alpha S_{\mathrm{tfidf}}(d) + \beta S_{\mathrm{semantic}}(d)$ and selects passages adaptively (minimum $m$, maximum $n$, threshold $\tau$). Second, a VLM-based data engine generates and verifies reasoning QA pairs in five categories, filtering by answer feasibility. Third, ensemble inference runs multiple decoding configurations, stops early when confidence exceeds 0.8 after at least 10 runs, and aggregates answers by majority vote. The same retriever feeds both training and inference, making long-document QA compu

What would settle it

One concrete test: fine-tune a different base model (for example InternVL3-78B) on the same generated QA data and evaluate on JDocQA; if the accuracy gain is much smaller than what the paper reports for Qwen2.5-VL, that would indicate the gains are partly an artefact of the generator and the fine-tuned model sharing the same pretrained base. A complementary check: evaluate on JDocQA test questions whose wording never appears in the generated training set; if accuracy drops sharply, the augmentation may have taught answer-matching rather than document reasoning.

Watch

Extended reading notes

Core claim

AdaDocVQA establishes that a 72-billion-parameter vision-language model, Qwen2.5-VL, fine-tuned with LoRA on a mix of filtered original data and automatically generated reasoning QA pairs, can answer questions about long Japanese documents at levels that surpass published baselines, provided the model's input is limited to relevant retrieved segments. The retrieval stage fuses TF-IDF with multilingual-e5-large semantic embeddings, with a dynamic threshold selecting 3 to 7 pages. The inference stage samples 20 decoding configurations, stops early when confidence reaches 0.8 after at least 10 runs, and aggregates answers by majority vote. On JDocQA the framework achieves 83.04% on Yes/No, 52.6

Load-bearing premise

The load-bearing premise is that the VLM-generated question-answer pairs, and the answer-feasibility filter, produce training data whose distribution matches the held-out JDocQA and LAVA test questions; if the generated data instead reinforce the generator's own answer patterns, the fine-tuning gains may reflect self-consistency with the generator rather than improved document understanding, and the reported improvements would not transfer to other models or languages.

Editorial extensions

If this is right

  • Applied as described, the framework makes long PDF documents tractable for 72B vision-language models: retrieval avoids the token-limit and out-of-memory failures that occur when all pages are fed directly.
  • The reported gains on JDocQA and LAVA indicate that automatically generated reasoning QA data can substitute for a large amount of human-annotated data in low-resource languages.
  • The ensemble strategy with early stopping yields accuracy gains while controlling inference cost, suggesting that a fixed decoding configuration is not the best use of a fine-tuned model on this task.
  • If the framework is language-agnostic as claimed, the same pipeline should produce comparable improvements on other low-resource languages, providing a practical path for document VQA beyond English.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The gains may depend on the fact that the same model family (Qwen2.5-VL) is both the data generator and the fine-tuned answerer; a reader should expect that the generated QA pairs carry the generator's stylistic biases, so part of the improvement could be self-consistency rather than general document understanding. A testable check is to fine-tune a different base model (e.g., InternVL3-78B) on th
  • A cross-language transfer test would sharpen the low-resource claim: applying the identical pipeline to Korean or Arabic document VQA, with a retriever and generator suited to that script, would show whether the recipe generalizes or whether Japanese-specific morphological processing (the enhanced TF-IDF tokenization) is the real source of the gains.
  • The dynamic retrieval parameters (m=3, n=7, τ=0.3 on LAVA) are tuned for one benchmark; sweeping these parameters on JDocQA would reveal whether the adaptive selection, rather than the fixed settings, is what drives the reported accuracy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. AdaDocVQA proposes a modular pipeline for Japanese document VQA: VLM-based OCR enhancement and cleaning, Qwen2.5-VL answer-feasibility filtering, generation of 1,808 QA pairs in five reasoning categories, hybrid TF-IDF + dense retrieval with adaptive thresholds, LoRA fine-tuning of Qwen2.5-VL-72B, and ensemble decoding with early stopping. On JDocQA the full pipeline reports 83.04% Y/N, 52.66% factual, and 44.12% numerical accuracy; on LAVA it reports 59%, with an ablation showing retrieval 45%, +LoRA 54%, +ensemble 59%. The paper claims state-of-the-art results for Japanese document VQA.

Significance. The framework is a reasonable engineering synthesis, and the direction—combining retrieval, targeted augmentation, and ensembling for low-resource document VQA—is worthwhile. Strengths include a concrete pipeline, an ablation that removes each component in turn, a public code link, and an explicit accounting of the augmented training set. However, the empirical evidence as reported does not yet support the SOTA claim. The JDocQA comparison is against an older baseline only, the LAVA test set is small enough that the reported gaps are within sampling noise, and no experiment isolates the contribution of the generated data. These issues are fixable with additional baseline and control experiments, so the result is plausible but not demonstrated as stated.

major comments (4)
  1. [§4.3, Table 1] The SOTA claim rests on Table 1, but on JDocQA the only comparison is InstBLIP, a model from a different generation. There is no zero-shot Qwen2.5-VL-72B row and no fine-tuned Qwen2.5-VL without the AdaDocVQA OCR/retrieval/augmentation components. Because the row labeled 'Qwen2.5-VL-72B' appears to be the full pipeline, the 52.66% factual and 44.12% numerical numbers cannot be attributed to the proposed modules rather than to the base model. Please add (a) zero-shot Qwen2.5-VL-72B on the original OCR text/images and (b) a LoRA fine-tuned Qwen2.5-VL-72B without augmentation/retrieval, so that the contribution of each framework component on JDocQA can be separated.
  2. [§4.3, Tables 2–3] LAVA has 177 test samples. The 59% vs. 49% difference over InternVL3-78B is 10 points (≈18 examples) and, under an independent-samples binomial approximation, the standard error is about 5.3 points (p≈0.06). The ablation's 54→59 improvement is 5 points (≈9 examples) and is within noise; no confidence intervals, significance tests, or repeated-seed estimates are reported. Given the SOTA claim, please report confidence intervals/bootstrap intervals and, where possible, run multiple seeds or a larger held-out set. Also specify whether the same 177 samples were used for hyperparameter selection; if so, the intervals should be adjusted.
  3. [§3.3, §4.1.1, Table 3] The data-augmentation claim is not tested. The 1,808 generated QA pairs are folded into the 10,282-sample training set, but the Table 3 ablation only toggles the whole LoRA training condition. One cannot tell whether the gains come from the generated pairs, the filtered original JDocQA data, or the new OCR text. Additionally, the generator and feasibility filter are Qwen2.5-VL, the same model family that is fine-tuned and evaluated; if the generated questions share the generator's answer patterns, the improvement may be self-consistency rather than document understanding. Please ablate the generated data separately and, at minimum, hand-verify a sample of the generated pairs or use an independent judge for filtering.
  4. [§3.4.4, §4.2.2] The adaptive retrieval and early-stopping mechanisms introduce several hand-set parameters (α=0.6, β=0.4, m=3, n=7, τ=0.3, confidence 0.8, minimum 10 inferences). No sensitivity analysis is reported. On a 177-sample test set, the risk that these values were effectively chosen to maximize the reported accuracy is nontrivial, especially since the SOTA claim rests on a 10-point gap. Please report a sensitivity sweep or justify the values on a validation split.
minor comments (6)
  1. [§4.3, Tables 1–2] Clarify row labels: 'Qwen2.5-VL-72B' is the full AdaDocVQA pipeline, not the stock model. This creates confusion about what is being compared.
  2. [Eq. (3)] The set-builder notation is ambiguous; '|R| < m' refers to the cardinality of the set under construction. Please rewrite as a ranked list with explicit thresholds on the ordered documents.
  3. [§4.1.1] State which OCR model was used for re-extraction; 'deep learning models' is not reproducible.
  4. [§4.1.1] Report the number of questions discarded by feasibility filtering (9,290 + 1,134 − 8,474 = 1,950) and how many were unanswerable vs. other failures. This affects comparability with prior JDocQA results.
  5. [Eq. (1)] There is a typographical issue in the denominator: the norm notation and period are run together. Please fix the spacing/formatting.
  6. [Figures and dataset names] The paper uses both 'Lava' and 'LAVA'; use the official capitalization consistently throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all reported results are measured against fixed external benchmarks, with no step reducing a prediction to its own input.

full rationale

The paper's central claims are empirical accuracy numbers on JDocQA and LAVA, both of which are external benchmarks with fixed test labels. The training pipeline combines OCR enhancement, hybrid retrieval, LoRA fine-tuning, generated QA pairs, and ensemble voting; none of these components is defined in terms of the test labels, and no equation or stated construction makes a reported number equal to an input by construction. The data-augmentation and answer-feasibility filtering use Qwen2.5-VL to produce and validate training QA pairs, and Qwen2.5-VL-72B is also the fine-tuned base model; this creates a real generalization/validity concern (the generated training distribution may reinforce the generator's own answer patterns), but that is not circularity under the criteria here: the generated pairs are training inputs, not fitted test predictions, and the JDocQA/LAVA test answers are never used in generation, filtering, retrieval, or ensemble selection. There are no load-bearing self-citations: references to Qwen2.5-VL, JDocQA, LAVA, and other models are to external work by other groups. No uniqueness theorem is imported from the authors' prior work, no ansatz is smuggled in via self-citation, and the ensemble/retrieval techniques, while not novel in isolation, are not presented as a renamed version of the benchmark results. The paper is self-contained against external benchmarks, so the circularity score is 0.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The empirical claims rest on a large set of hand-set hyperparameters (retrieval weights, thresholds, ensemble composition, LoRA settings, training schedule) and on unverified assumptions about the generator model's data quality, the absence of train/test overlap, and the value of ensemble voting. No new theoretical entities are introduced.

free parameters (7)
  • Retrieval fusion weights alpha, beta = alpha=0.6, beta=0.4 (LAVA)
    Hand-set in Section 4.1.2; controls the balance between TF-IDF and semantic scores in the fused retrieval score and directly determines which pages are selected.
  • Retrieval selection thresholds m, n, tau = m=3, n=7, tau=0.3 (LAVA)
    Hand-set in Section 4.1.2 and used in Equation 3; controls the minimum and maximum number of document pages passed to the model and the relevance threshold.
  • Ensemble early-stop confidence threshold = 0.8 with at least 10 completed inferences
    Section 4.2.2; halts ensemble sampling when answer consistency reaches 0.8, affecting both accuracy and compute cost.
  • Ensemble decoding configuration count and ranges = 20 configurations, temperature 0.1-1.5, varied top-p and top-k
    Section 4.2.2; the set of decoding configurations is chosen by hand and determines the voting pool.
  • LoRA rank and alpha = rank=32, alpha=64
    Section 4.2.1; LoRA hyperparameters are fixed without a reported search and affect the fine-tuning quality.
  • Training epochs and learning rate = 7 epochs, lr=1e-4
    Section 4.2.1; fixed without reported validation-based selection.
  • TF-IDF n-gram range and maximum features = 1 to 5 grams, 50,000 features
    Section 3.4.2; hand-configured for Japanese text and affects the lexical retrieval scores.
assumptions (4)
  • domain assumption Qwen2.5-VL's OCR, answer-feasibility analysis, and QA generation outputs are accurate enough to produce high-quality training data.
    Invoked throughout Sections 3.2 and 3.3; if the generated or filtered data is mislabeled, the fine-tuned model could be trained on unreliable targets.
  • domain assumption The JDocQA train and validation splits plus the generated QA pairs do not overlap with JDocQA test questions or LAVA test documents.
    Required for a valid evaluation; the paper does not describe any overlap check in Section 4.1.
  • domain assumption multilingual-e5-large embeddings capture Japanese semantic similarity well enough for document retrieval.
    Section 3.4.3 relies on this model without reporting any Japanese retrieval benchmark or comparison to other embedders.
  • domain assumption Majority voting over 20 sampled generations yields the correct answer more reliably than any single decoding configuration.
    Section 3.6 assumes ensemble voting helps; no vote distribution analysis or failure-case study is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AdaDocVQA: Adaptive Framework for Long Document Visual Question Answering in Low-Resource Settings." pith.science (2026). https://pith.science/paper/IVGIQP76

@misc{pith2026250813606,
  author       = {Pith},
  title        = {Pith review of: AdaDocVQA: Adaptive Framework for Long Document Visual Question Answering in Low-Resource Settings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IVGIQP76}},
  note         = {Machine review of arXiv:2508.13606}
}
read the original abstract

Document Visual Question Answering (Document VQA) faces significant challenges when processing long documents in low-resource environments due to context limitations and insufficient training data. This paper presents AdaDocVQA, a unified adaptive framework addressing these challenges through three core innovations: a hybrid text retrieval architecture for effective document segmentation, an intelligent data augmentation pipeline that automatically generates high-quality reasoning question-answer pairs with multi-level verification, and adaptive ensemble inference with dynamic configuration generation and early stopping mechanisms. Experiments on Japanese document VQA benchmarks demonstrate substantial improvements with 83.04\% accuracy on Yes/No questions, 52.66\% on factual questions, and 44.12\% on numerical questions in JDocQA, and 59\% accuracy on LAVA dataset. Ablation studies confirm meaningful contributions from each component, and our framework establishes new state-of-the-art results for Japanese document VQA while providing a scalable foundation for other low-resource languages and specialized domains. Our code available at: https://github.com/Haoxuanli-Thu/AdaDocVQA.

Figures

Figures reproduced from arXiv: 2508.13606 by the authors.

Figure 1
Figure 1. Overview of AdaDocVQA Our AdaDocVQA framework operates through an integrated pipeline comprising five core components: enhanced OCR pro￾cessing with data quality control, intelligent data augmentation, hybrid text retrieval, model fine-tuning, and ensemble inference, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. VLM Based OCR System Architecture with Data [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. VLM Based Answer Validation System for OCR Text [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Overview of Retrieval Method jieba-based morphological analysis for proper word segmentation, then employs regex-based fallback patterns to capture overlooked technical terms and numerical expressions. The tokenization process begins with text normalization to stan￾dar…
Figure 6
Figure 6. Figure 6: Model Fine-tuning and Inference Pipeline [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Model Ensemble Inference Pipeline 3.5 Model Fine-tuning Strategy Training data integration combines filtered original data, generated reasoning questions, and synthesized OCR content for augmenta￾tion. LoRA fine-tuning targets key model components with opti￾mized adapt…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 9 canonical work pages

  1. [1]

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. Self-rag: Learning to retrieve, generate, and critique through self-reflection. In The Twelfth International Conference on Learning Representations

    2023

  2. [2]

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-vl: A frontier large vision- language model with versatile abilities. arXiv preprint arXiv:2308.12966 1, 2 (2023), 3

    arXiv 2023

  3. [3]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al . 2025. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

    arXiv 2025

  4. [4]

    Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al . 2024. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271 (2024)

    arXiv 2024

  5. [5]

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al . 2024. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 24185–24198

    work page 2024

  6. [6]

    Hao Cheng, Yelong Shen, Xiaodong Liu, Pengcheng He, Weizhu Chen, and Jianfeng Gao. 2021. UnitedQA: A hybrid approach for open domain question answering. arXiv preprint arXiv:2101.00178 (2021)

    work page Pith review arXiv 2021

  7. [7]

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. 2024. The faiss library. arXiv preprint arXiv:2401.08281 (2024)

    arXiv 2024

  8. [8]

    Soyeong Jeong, Jinheon Baek, Sukmin Cho, Sung Ju Hwang, and Jong C Park

Show all 36 references
  1. [9]

    Xinke Jiang, Ruizhe Zhang, Yongxin Xu, Rihong Qiu, Yue Fang, Zhiyuan Wang, Jinyi Tang, Hongxin Ding, Xu Chu, Junfeng Zhao, et al. 2023. Hykge: A hypothesis knowledge graph enhanced framework for accurate and reliable medical llms responses. arXiv preprint arXiv:2312.15883 (202...

    2023 arXiv

  2. [10]

    Zhengbao Jiang, Frank F Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. 2023. Active retrieval augmented generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 7969–7992

    2023

  3. [11]

    Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. 2019. Generalization through memorization: Nearest neighbor language models. arXiv preprint arXiv:1911.00172 (2019)

    2019 arXiv

  4. [12]

    Kentaro Kurihara, Daisuke Kawahara, and Tomohide Shibata. 2022. JGLUE: Japan- ese general language understanding evaluation. In Proceedings of the Thirteenth Language Resources and Evaluation Conference . 2957–2966

    2022

  5. [13]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing s...

    2020

  6. [14]

    Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. 2024. Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895 (2024)

    2024 arXiv

  7. [15]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning . PMLR, 19730–19742

    2023

  8. [16]

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 26296–26306

    2024

  9. [17]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual in- struction tuning. Advances in neural information processing systems 36 (2023), 34892–34916

    2023

  10. [18]

    Minesh Mathew, Viraj Bagal, Rubèn Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. 2022. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision . 1697–1706

    2022

  11. [19]

    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 . 2200–2209

    2021

  12. [20]

    Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty

  13. [21]

    Eri Onami, Shuhei Kurita, Taiki Miyanishi, and Taro Watanabe. 2024. JDocQA: Japanese document question answering dataset for generative language models. arXiv preprint arXiv:2403.19454 (2024)

    2024 arXiv

  14. [22]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  15. [23]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Em- pirical Methods in Natural Language Processing . Association for Computational Linguistics. https://arxiv.org/abs/1908.10084

    2019 arXiv

  16. [24]

    Nils Reimers and Iryna Gurevych. 2020. Making Monolingual Sentence Em- beddings Multilingual using Knowledge Distillation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing . Association for Computational Linguistics. https://arxiv.org/a...

    2020 arXiv

  17. [25]

    Weihang Su, Yichen Tang, Qingyao Ai, Zhijing Wu, and Yiqun Liu. 2024. DRAGIN: Dynamic Retrieval Augmented Generation based on the Information Needs of Large Language Models. arXiv preprint arXiv:2403.10081 (2024)

    2024 arXiv

  18. [26]

    Ryota Tanaka, Kyosuke Nishida, Kosuke Nishida, Taku Hasegawa, Itsumi Saito, and Kuniko Saito. 2023. Slidevqa: A dataset for document visual question an- swering on multiple images. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 13636–13645

    2023

  19. [27]

    Rubèn Tito, Dimosthenis Karatzas, and Ernest Valveny. 2023. Hierarchical multi- modal transformers for multipage docvqa. Pattern Recognition 144 (2023), 109834

    2023

  20. [28]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    2023 arXiv

  21. [29]

    Duc Minh Vo, Hideki Nakayama, Khan Md Anwarus Salam, Jiaxuan Li, Junwen Mo, and Daichi Sato. 2025. LAVA Challenge: Large Vision-Language Model Learning and Applications. ACM MM 2025 Workshop and Grand Challenge. https://lava-workshop.github.io/grandchallenge Dublin, Ireland, O...

    2025

  22. [30]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. 2024. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191 (2024)

    2024 arXiv

  23. [31]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR)

    2023

  24. [32]

    Ori Yoran, Tomer Wolfson, Ori Ram, and Jonathan Berant. 2023. Making retrieval-augmented language models robust to irrelevant context. arXiv preprint arXiv:2310.01558 (2023)

    2023 arXiv

  25. [33]

    Tianyang Zhong, Zhenyuan Yang, Zhengliang Liu, Ruidong Zhang, Yiheng Liu, Haiyang Sun, Yi Pan, Yiwei Li, Yifan Zhou, Hanqi Jiang, et al . 2024. Opportu- nities and challenges of large language models for low-resource languages in humanities research. arXiv preprint arXiv:2412....

    2024 arXiv

  26. [34]

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

    2025 arXiv

  27. [2019]

    In 2019 international conference on document analysis and recognition (ICDAR)

    Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR) . IEEE, 947–952

    2019

  28. [2024]

    arXiv preprint arXiv:2403.14403 (2024)

    Adaptive-rag: Learning to adapt retrieval-augmented large language models through question complexity. arXiv preprint arXiv:2403.14403 (2024)

    2024 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.