DAViD: Domain Adaptive Visually-Rich Document Understanding with Synthetic Insights

I have following concerns that need to be explained:

1. When encountering a new domain, the proposed method requires collecting raw data, tagging a small portion manually, and using a large language model (LLM) to tag the majority. Then, the proposed pipeline is applied for training. This complex process may impede real application.
2. The experiment does not test the method's performance in the original domain after training on new domain data, which is crucial to demonstrate the maintenance of knowledge from other domains.
3. In Table 1, the experimental setup lacks clarity. It is unclear if "Full Training Set" refers to D_n and D_g. If so, the baseline method appears to outperform the proposed method, reducing the contribution of the method. Additionally, the evaluation metrics for Tables 1, 2, and 3 are not mentioned, causing confusion.
4. In Section 6.2, the term "zero-shot" testing is misleading. The method uses some new domain data (a small portion labeled manually) to train the model, enhancing its ability in this domain. This is more suitable to be described as a few-shot application. Furthermore, the baseline settings in Table 3 are unexplained (likely LXMERT and LayoutLMv3) and the extreme low performance of the baseline method is also confusing. Domain knowledge is unlikely to significantly influence the performance, as the trained model should possess basic knowledge of the document understanding (DU) task. Clarification of the experimental setup is required to avoid confusion.
5. The full name of L2V is absent from the paper. Given that an ablation study of L2V is conducted, indicating its importance, a detailed explanation is needed. Additional readability issues are present:
6. Many figures, especially Figures 1, 2, and 10, are blurred.
7. Section 5.1 mentions handwritten (H), digital (D), and printed (P), but uses at least two abbreviations ((\mathcal{H}), (\mathcal{D}), (\mathcal{P})). Are these intended to have different meanings?
8. In Table 2, it is unclear what FST stands for. Is it a typo or an unexplained technique?

1. The paper did not conduct experiments on advanced open-source large models, which limits a comprehensive assessment of the DAViD framework's performance across different types of models.
2. The paper may not have discussed in detail the potential biases in the synthetic data generation process and how these biases could affect model performance.
3. The paper focuses on solving the understanding problems of special domain documents, so the data selection should pay more attention to the diversity and professionalism of the field. The paper's discussion on the breadth of the data selected may not have been sufficiently explored.

• A lot of unclear details: The most significant drawback of the paper is the poor writing that prevents the readers from appreciating the work. The paper introduces many components and creates many new terms. However many of them are unclear and not properly described. The new terms (e.g. L2V, SDS, SIT) could be better explained with figures (potentially Figure 1). But there was no illustration. See the questions below.

• Insufficient baseline comparison: The author should consider two additional baselines [1, 2]:

[1] Kim et al., 2022, OCR-free Document Understanding Transformer.

[2] Lee et al., 2022, Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding

1. The presentation of this paper is poor. The author introduce a lot of terms and abbreviations in Section 3 and 4. The method description is also redundant and complicated, which greatly hinders the readers from understanding this paper. Besides, I strongly suggest the author to rewrite the Introduction section, which is too redundant and repetitive. And I also recommend the author to replot Figure 1 and 3 into vector graphics, because the layout and small words are difficult to read.
2. The author should list the number of parameters for all the method to compare different method’s computation resources. The author takes the document understanding (in this case, it’s key information extraction) into several parts, and uses synthetic data to pre-train each part. In this case where the number of parameters and the amount of training data are higher than the baseline, it is obvious that the final performance is better, especially under low-resource scenarios, which I think is quite trivial.
3. In recent years, there are some MLLMs that are specially trained for document understanding tasks[1-3] and show stronger capability compared with old document analysis pre-trained models (LayoutLMv3, LiLT) and general MLLMs (Llava, Qwen-VL). I strongly suggest the author to add experiments on these MLLMs.

[1] Liu, Yuliang, et al. "Textmonkey: An ocr-free large multimodal model for understanding document." arXiv preprint arXiv:2403.04473 (2024). [2] Wei, Haoran, et al. "Vary: Scaling up the vision vocabulary for large vision-language model." European Conference on Computer Vision. Springer, Cham, 2025. [3] Hu, Anwen, et al. "mplug-docowl 1.5: Unified structure learning for ocr-free document understanding." arXiv preprint arXiv:2403.12895 (2024).