Review Reports

Round 1

Reviewer 1 Report (New Reviewer)

Comments and Suggestions for Authors

The authors of this study propose an approach called dynamically fusing visual context (DFVC), which trains a lightweight multi-layer perceptron (MLP) adapter to incorporate contextual information from neighboring document pages into the encoding of a target page. Through experiments on the benchmark datasets of MMLongBench-Doc and MP-DocVQA, the authors demonstrate that the proposed DFVC significantly outperforms baseline methods in retrieval performance. While the paper reports improved retrieval performace with minimal computational cost, the manuscript lacks sufficient clarification on several key points and would benefit from further revision.

1.    Could the authors clarify how the losses defined in equation (6), the loss over K samples for a query, and (7), the loss over a batch of size B, are combined in equation (8), and how K and B are connected?
2. Could the authors clarify how the benchmark datasets of MMLongBench-Doc and MP-DocVQA are used for training the adapter and evaluation? For example, are the datasets split into subsets for training and evaluation, or the full datasets are used for both training and evaluation? The manuscript would benefit from a clearer description of the datasets and how they are used.
3. Could the authors provide the specific values of K and Tau for equation (5), and details on the text encoder used for embedding the query in the experiments?
4. Although the authors claimed "the proposed DFVC can be applied to existing vector databases without re-computing the base embeddings", applying the adapter to existing embedding database appears to require creating new embeddings. It would be useful if the authors can elaborate on applying the adapter to existing embedding database in more details.
5. Clarification of several minor issues would further improve the clarity of the manuscript:
   - The proposed adapter architecture appears inconsistent (e.g., two linear layers in Line 178 vs. a 3-layer MLP in Line 239).
   - The statements "we initialize the bias of the final MLP layer to a negative value (e.g., -5.0) and α to 0" (line 220) and "Replacing the residual connection with a direct feed-forward output (direct MLP) ..." (line 330) are unclear and would benefit from further explanation.

Comments on the Quality of English Language

A thorough grammar check and language polishing would further improve clarity of the manuscript.

Author Response

Comments 1: Could the authors clarify how the losses defined in equation (6), the loss over K samples for a query, and (7), the loss over a batch of size B, are combined in equation (8), and how K and B are connected?

Response 1:Following this comment, we extended the paragraph (in blue text) around (8) for detailing the contribution of the ranking loss  in (6) and the drift loss  in (7).  constrains the adapter to generate proper weights for neighboring pages and hence lets the fused embedding vector for a positive sample have a higher similarity with the query. K is the number of samples used in (5) and (6) for computing the ground-truth and predicted probability; and B is the batch size for computing the drift between the original vector and the fused embedding vector. In this work, B is 64. 

Comments 2: Could the authors clarify how the benchmark datasets of MMLongBench-Doc and MP-DocVQA are used for training the adapter and evaluation? For example, are the datasets split into subsets for training and evaluation, or the full datasets are used for both training and evaluation? The manuscript would benefit from a clearer description of the datasets and how they are used.

Response 2: Following this comment, we added the use of the two datasets and extended the description for them. Both datasets aim to test the retrieving performance of the models on multipage documents. MMLongBench-Doc and MP-DocVQA were split into two parts: 80% for training the adapter in the proposed DFVC, and 20% for validating the performance of the adapter.

Comments 3: Could the authors provide the specific values of K and Tau for equation (5), and details on the text encoder used for embedding the query in the experiments?

Response 3: We added the values of K and Tau following equation (5). In general, K is set to 8, i.e., totally 8 samples were used for training the adapter. Tau is set to 0.1, serving as the hyper-parameter. For the encoder of the query (text), this work used the multimodal large language model (MLLM), which was also added in the manuscript, following equation (1).

Comments 4: Although the authors claimed "the proposed DFVC can be applied to existing vector databases without re-computing the base embeddings", applying the adapter to existing embedding database appears to require creating new embeddings. It would be useful if the authors can elaborate on applying the adapter to existing embedding database in more details.

Response 4: Following this comment, we added a paragraph at the end of Section 3.1.1 to explain the “without re-computing the base embeddings”. The adapter operates on the embedding vectors output of the backbones, and consequently can be directly applied to the embedding vectors without calling the large language models (as backbones) to generate base vectors. This significantly reduces the demand for the computational resource.

Comments 5: Clarification of several minor issues would further improve the clarity of the manuscript:
   - The proposed adapter architecture appears inconsistent (e.g., two linear layers in Line 178 vs. a 3-layer MLP in Line 239).
   - The statements "we initialize the bias of the final MLP layer to a negative value (e.g., -5.0) and α to 0" (line 220) and "Replacing the residual connection with a direct feed-forward output (direct MLP) ..." (line 330) are unclear and would benefit from further explanation.

Response 5: -We changed “3-layer MLP” to “two linear layers” to make the manuscript easier to follow. A MLP (multi-layer perceptron) comprising three layers contains two linear mappings, the first is from the input layer to the hidden layer, and the second is from the hidden layer to the output layer. 

-We revised the text to make the manuscript more legible in the part ‘Contribution of residual connection’. Specifically, “the residual connection” was used to fuse the embedding vectors in equation (3). To fuse the embedding vectors, one can predict the coefficients from scratch, i.e., predict the coefficients for , , and , but this is more difficult to obtain proper weights than the residual connection since a small initial value for  may result in context dilution and hard convergence.

Reviewer 2 Report (New Reviewer)

Comments and Suggestions for Authors

This paper addresses an important limitation in recent visual document retrieval systems: the independent encoding of document pages. The authors argue that many semantic cues in real-world documents span multiple pages, and treating pages independently causes retrieval errors. To address this, the paper proposes a method called Dynamically Fusing Visual Context (DFVC), which uses a lightweight adapter to fuse embeddings from neighboring pages. The adapter predicts fusion weights for the previous and next pages, and the fused embedding is trained using a listwise contrastive ranking loss with an additional drift regularization term. The method is designed to be parameter-efficient and compatible with frozen visual-language backbones.

Experiments on two benchmarks (MMLongBench-Doc and MP-DocVQA) demonstrate consistent improvements over page-independent baselines and simple fusion methods. The method is simple, practical, and easy to integrate into existing pipelines. However, the novelty is somewhat incremental, and several aspects of the experimental design, analysis, and clarity of presentation need improvement.

Major Comments

1. Novelty and Positioning Need Stronger Justification

The core idea of the paper is to fuse embeddings from neighboring pages using learned weights predicted by a lightweight adapter. While this is a reasonable and practical idea, it appears to be conceptually close to several existing approaches in both text and multi-modal retrieval. For example, sliding window retrieval, contextual embedding methods, late fusion strategies, and graph-based document models all attempt to incorporate cross-segment or cross-page context. The proposed method essentially learns adaptive weights for neighboring embeddings, which can be seen as a learned variant of window-based context aggregation.

The paper currently positions DFVC as a novel solution to the context dilution problem, but it does not sufficiently explain how it differs in principle from other learned fusion or contextualization techniques. The related work section mentions several approaches, but the comparisons remain high-level and descriptive rather than analytical.

To improve this, the authors should explicitly articulate what is fundamentally new about DFVC. For example, they could explain whether the novelty lies in the lightweight post-hoc adapter design, the specific drift regularization, the listwise contrastive objective, or the combination of these elements. A short subsection that directly compares DFVC with sliding window approaches, sequence models, and graph-based methods would help clarify the unique contribution.

2. Performance Gains Are Modest and Need Statistical Validation

The experimental results show consistent improvements across both datasets and backbones. However, the absolute gains are relatively small. For example, on MMLongBench-Doc using the VisRAG backbone, the improvement is approximately 0.4959 to 0.5250 in nDCG@5. On MP-DocVQA, the gain is from 0.1873 to 0.2050. While these improvements are consistent, they are not large, especially considering the simplicity of the fusion baselines.

Moreover, the paper reports only single-run results without any measure of statistical variability. There are no confidence intervals, standard deviations, or statistical significance tests. Without these, it is difficult to assess whether the improvements are robust or due to randomness in training or sampling.

The authors should repeat experiments multiple times with different random seeds and report mean and standard deviation. They should also include a statistical significance test, such as a paired t-test or bootstrap test, to demonstrate that the improvements are meaningful.

3. Limited Context Window Is Not Adequately Justified

The proposed method only considers a context window of three pages: the previous, current, and next page. While this is a reasonable starting point, it assumes that relevant semantic context always lies within immediate neighbors. In many real-world documents, relevant content may be separated by several pages, such as multi-page tables, long sections, or references across chapters.

The paper briefly mentions the possibility of using larger context windows in future work, but it does not justify why a window size of three was chosen. It is unclear whether this choice is based on empirical evidence, computational constraints, or architectural simplicity.

The authors should conduct an experiment evaluating different context window sizes, such as 1, 3, 5, or 7 pages, or even document-level fusion. This would help demonstrate whether the proposed method generalizes to longer context dependencies and whether performance saturates or degrades with larger windows.

4. Lack of Efficiency and Latency Analysis

The paper claims that the method is “highly parameter-efficient” and introduces “negligible computational overhead.” However, no quantitative analysis is provided to support these claims. Since the method requires accessing neighboring page embeddings and performing additional fusion operations, it is important to measure the actual overhead.

The paper should include a comparison of: Query latency, Indexing time, Memory usage, Throughput

between the baseline models and DFVC. Even a simple table showing relative overhead would help validate the efficiency claims.

5. Training Protocol and Data Construction Are Under-Specified

The training setup lacks several important details that are necessary for reproducibility. The paper does not clearly describe how training pairs or lists are constructed. It is unclear: How many training queries are used, How positive pages are defined, How negative pages are sampled, Whether negatives are from the same document or different documents, Whether there is any overlap between training and test documents

Since the method relies on listwise contrastive training, the negative sampling strategy is particularly important. Different sampling strategies can significantly affect performance.

The authors should include a subsection describing the training data construction, including dataset splits, number of training samples, and negative sampling strategy.

6. Lack of Qualitative Analysis and Failure Cases

The paper only presents quantitative metrics, without any qualitative examples. It would be helpful to see specific cases where DFVC improves retrieval compared to the baseline. For example, showing a query where the baseline fails due to page-independent encoding, but DFVC succeeds by incorporating context, would make the contribution more concrete.

Similarly, it would be useful to analyze failure cases where DFVC does not help or even hurts performance. This would help clarify the limitations of the approach.

The authors should include at least one qualitative example showing: The query, The target page, The neighboring pages and the Retrieval scores with and without DFVC

Author Response

Comments 1: Novelty and Positioning Need Stronger Justification.

The core idea of the paper is to fuse embeddings from neighboring pages using learned weights predicted by a lightweight adapter. While this is a reasonable and practical idea, it appears to be conceptually close to several existing approaches in both text and multi-modal retrieval. For example, sliding window retrieval, contextual embedding methods, late fusion strategies, and graph-based document models all attempt to incorporate cross-segment or cross-page context. The proposed method essentially learns adaptive weights for neighboring embeddings, which can be seen as a learned variant of window-based context aggregation.

The paper currently positions DFVC as a novel solution to the context dilution problem, but it does not sufficiently explain how it differs in principle from other learned fusion or contextualization techniques. The related work section mentions several approaches, but the comparisons remain high-level and descriptive rather than analytical.

To improve this, the authors should explicitly articulate what is fundamentally new about DFVC. For example, they could explain whether the novelty lies in the lightweight post-hoc adapter design, the specific drift regularization, the listwise contrastive objective, or the combination of these elements. A short subsection that directly compares DFVC with sliding window approaches, sequence models, and graph-based methods would help clarify the unique contribution.

Response 1: The proposed DFVC can be taken as a learnable variant of the sliding window. The adapter is responsible for predicting the proper weights for neighboring pages, and the contrastive loss constrains the adapter to do this by measuring the difference between the predicted probability for the samples and the ground truth.

The novelty of the proposed DFVC lies in its two components: the adapter and the loss function. The designed loss function constrains the adapter to learn proper weights for fusing the information across pages,

The proposed DFVC was compared with the sequence model Bi-LSTM, and showed the advantage over Bi-LSTM. The comparison was reported in Table 1 and Table 2 on the dataset MMLongBench-Doc and MP-DocVQA, respectively.

In Section 4.3 ‘Ablation Studies’, we also investigated the sliding window method with fixed weights, 0.1, 0.8, 0.1, and its results were given in Table 3.

Comments 2: Performance Gains Are Modest and Need Statistical Validation.

The experimental results show consistent improvements across both datasets and backbones. However, the absolute gains are relatively small. For example, on MMLongBench-Doc using the VisRAG backbone, the improvement is approximately 0.4959 to 0.5250 in nDCG@5. On MP-DocVQA, the gain is from 0.1873 to 0.2050. While these improvements are consistent, they are not large, especially considering the simplicity of the fusion baselines.

Moreover, the paper reports only single-run results without any measure of statistical variability. There are no confidence intervals, standard deviations, or statistical significance tests. Without these, it is difficult to assess whether the improvements are robust or due to randomness in training or sampling.

The authors should repeat experiments multiple times with different random seeds and report mean and standard deviation. They should also include a statistical significance test, such as a paired t-test or bootstrap test, to demonstrate that the improvements are meaningful.

Response 2: In the field of multimodal retrieval, the improvement from 0.4959 to 0.5250 in nDCG@5 and 0.1873 to 0.2050 are actually not minor improvement, especially the improvement from 0.4959 to 0.5250, which is about 2.9%.

Following this comment, we ran the experiments 5 times using different random seeds and computed the mean and standard deviation and they were added in Table 1 and Table 2.

Also, we conducted a paired t-test, and its result confirms that the improvements achieved by the proposed DFVC over the baselines are statistically significant (p<0.05 in all major comparisons). A paragraph marked in magenta was added in Section 4.2 to explain this.

Comments 3: Limited Context Window Is Not Adequately Justified.

The proposed method only considers a context window of three pages: the previous, current, and next page. While this is a reasonable starting point, it assumes that relevant semantic context always lies within immediate neighbors. In many real-world documents, relevant content may be separated by several pages, such as multi-page tables, long sections, or references across chapters.

The paper briefly mentions the possibility of using larger context windows in future work, but it does not justify why a window size of three was chosen. It is unclear whether this choice is based on empirical evidence, computational constraints, or architectural simplicity.

The authors should conduct an experiment evaluating different context window sizes, such as 1, 3, 5, or 7 pages, or even document-level fusion. This would help demonstrate whether the proposed method generalizes to longer context dependencies and whether performance saturates or degrades with larger windows.

Response 3: For the case 1, it is just the current page itself, which amounts to the vanilla VisRAG and ColPali. For the case 3, it is the approach proposed in this work.

Following this comment, we conducted the experiment on the window size, and the result was given in Table 4. Setting the window size to 3 gave the best retrieval performance since it effectively encodes the context information. A larger window size (5, 7) may introduce visual noise, which may dampen the retrieval performance. A paragraph was also added close to Table 4, marked in magenta.

Comments 4: Lack of Efficiency and Latency Analysis.

The paper claims that the method is “highly parameter-efficient” and introduces “negligible computational overhead.” However, no quantitative analysis is provided to support these claims. Since the method requires accessing neighboring page embeddings and performing additional fusion operations, it is important to measure the actual overhead.

The paper should include a comparison of: Query latency, Indexing time, Memory usage, Throughput

between the baseline models and DFVC. Even a simple table showing relative overhead would help validate the efficiency claims.

Response 4:  Following this suggestion, we conducted an evaluation of query latency, indexing throughput, and GPU memory usage. Section 4.4 and Table 5 were added to explain this. The results showed that the DFVC introduces 4.2 MB of additional trainable parameters (+0.14% memory overhead) and adds only 1.3 ms to the query latency (+1.99% relative overhead). The impact on indexing throughput is roughly unnoticeable. These metrics validate that the proposed DFVC achieves significant retrieval improvement without requiring much computational overhead. Since the DFVC is a lightweight module, it can be easily applied post-hoc to pre-computed 1D page embeddings.

Comments 5: Training Protocol and Data Construction Are Under-Specified

The training setup lacks several important details that are necessary for reproducibility. The paper does not clearly describe how training pairs or lists are constructed. It is unclear: How many training queries are used, How positive pages are defined, How negative pages are sampled, Whether negatives are from the same document or different documents, Whether there is any overlap between training and test documents

Since the method relies on listwise contrastive training, the negative sampling strategy is particularly important. Different sampling strategies can significantly affect performance.

The authors should include a subsection describing the training data construction, including dataset splits, number of training samples, and negative sampling strategy.

Response 5: Following this comment, we added the introduction for the dataset in Section 4.1; and the dataset splits at the beginning of Section 4. 80% of the dataset is used as the training subset, and the rest (20%) is used as the validation. The training subset and the validation subset are not overlapped.

The positive sample is the ground truth page to a query, while the negative samples are sampled randomly, either from the document containing the ground truth page, or from different documents. This explanation was added in the manuscript and marked in magenta.

Comments 6: Lack of Qualitative Analysis and Failure Cases

The paper only presents quantitative metrics, without any qualitative examples. It would be helpful to see specific cases where DFVC improves retrieval compared to the baseline. For example, showing a query where the baseline fails due to page-independent encoding, but DFVC succeeds by incorporating context, would make the contribution more concrete.

Similarly, it would be useful to analyze failure cases where DFVC does not help or even hurts performance. This would help clarify the limitations of the approach.

The authors should include at least one qualitative example showing: The query, The target page, The neighboring pages and the Retrieval scores with and without DFVC

Response 6: A constructive comment! Following this comment, we added a new subsection, "Qualitative Analysis" in which a detailed visual example (Figure 3) in the revised manuscript.

Figure 3 explicitly shows the query, the target page, the neighboring pages, and the retrieval outcomes. We selected a challenging multi-hop query ("cost in 1 day"). The baseline VisRAG fails because the "unit cost" is on Page n but the daily volume ("1 day / daily volume") is on Page n+1. Since neither page independently contain sufficient information to answer the query, the baseline generates a low similarity to either page. In contrast, our DFVC dynamically fuses the context from Page n+1 into Page n, achieving a successful retrieval.

Reviewer 3 Report (New Reviewer)

Comments and Suggestions for Authors

This manuscript proposes a lightweight adapter with contrastive learning to adaptively fuse cross-page semantic information, significantly improving retrieval performance while maintaining parameter efficiency. Experimental results demonstrate consistent gains on multi-page document benchmarks. This is a promising and meaningful research direction.Please provide further explanation regarding the following issues.

1. Please provide a clearer explanation regarding the independent encoding of multi-page documents.

2. Figure 1 is not cited or explained in the main text. Please provide a proper reference and description.

3. Please strengthen the section on research objectives and better connect the study to the topic of this journal.

4. In the Experiments section, please explain how the proposed DFVC was trained for 20 epochs on 240 MP-DocVQA and 5 epochs on MMLongBench-Doc, and clarify how comparisons were made given the different numbers of epochs.

5. In Table 1, please clearly explain how the values under “Evidence Modalities” were calculated.

6. Please explain what nDCG@5 represents.

7. Please add a Discussion section to elaborate on the study findings, compare them with previous research, and address the study limitations.

Author Response

Comments 1:  Please provide a clearer explanation regarding the independent encoding of multi-page documents.

Response 1: Following this comment, we extended two paragraphs for further explaining the independent encoding. One is the paragraph starting with “Despite the advantage”, and the other is the paragraph starting with “Recent advancements have”. The extend paragraphs explain the independent encoding more clearly and are marked in orange.

Comments 2: Figure 1 is not cited or explained in the main text. Please provide a proper reference and description.

Response 2:  Following this comment, we referred to Figure 1 in proper paragraphs. Figure 1 (a) was meant to show the independent encoding of existing methods, and Figure 1 (b) was meant to show the structure of the proposed approach, dynamically fusing visual context (DFVC). In last version, we forgot to refer to Figure 1 and in the revised version it was referred to in the proper paragraphs.

Comments 3: Please strengthen the section on research objectives and better connect the study to the topic of this journal.

Response 3:  The main objective of the manuscript is to address the context dilution generated by directly aggregating the neighboring pages. To this end, this work proposed an approach, dynamically fusing visual context (DFVC), to effectively encode the context pages while reducing the context dilution of semantic information.

The paragraph in the revised version starting with “However, for visual document retrievers” was extended to strength the objective of the manuscript. Also, this study falls in the scope of the journal.

Comments 4: In the Experiments section, please explain how the proposed DFVC was trained for 20 epochs on 240 MP-DocVQA and 5 epochs on MMLongBench-Doc, and clarify how comparisons were made given the different numbers of epochs.

Response 4: The number of epochs was determined by the validation retrieval performance. Since MP-DocVQA (20 epochs) consists of scanned documents with significant visual noise and high inter-page similarity, it requies more iterations for the adapter to learn proper coefficients. On MMLongBench-Doc (5 epochs) however, the DFVC reached optimal nDCG performance much faster possibly because of the diversity of the training samples.

For the experimental comparison, in line with machine learning evaluation practices, all comparisons reported in Table 1 and Table 2 were made by using the best-performing checkpoint on the validation set for each specific task. This ensures that we are comparing the peak potential of each method across different benchmarks.

Comments 5: In Table 1, please clearly explain how the values under “Evidence Modalities” were calculated.

Response 5: The values under “Evidence Modalities” are the nDCG@5 that all methods achieved. In last version, we mentioned “nDCG (normalized Discounted Cumulative Gain), was used” but did not explain the meaning of nDCG and nDCG@5. In the revised version, we extended a paragraph to explain the nDCG and nDCG@K. The explaining text was marked in orange in the revised version.

Comments 6: Please explain what nDCG@5 represents.

Response 6:  nDCG is normalized Discounted Cumulative Gain. For the retrieved results, a logarithmic discount is applied to every retrieved result, the higher the ranking of the result, the smaller the discount. By this means, the high-ranked result will be more valued. Then the summation of discounted gain will be computed and normalized by the ideal DCG. The ideal DCG will be the score for the ground truth. The explanation was added in the revised version.

Also see the reply to Question #5.

Comments 7:  Please add a Discussion section to elaborate on the study findings, compare them with previous research, and address the study limitations.

Response 7:  Following this comment, we revised Section 5.2 to be ‘Discussion and Future Work’. In Section 5.1 we summarized the contribution of this work; and Section 5.2 summarized the study finding and the limitation. In addition, the future work is also presented by following the limitation.

Round 2

Reviewer 2 Report (New Reviewer)

Comments and Suggestions for Authors

Accept 

Reviewer 3 Report (New Reviewer)

Comments and Suggestions for Authors

The authors have adequately addressed the concerns raised during the revision process. I believe that the responses provided have satisfactorily clarified the issues identified in the previous review.

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Thank you very much for the opportunity to read your manuscript entitled: “Dynamic Visual Context Fusion for Multi-Page Document Retrieval” submitted for Electronics. This manuscript is valuable and innovative. After revision, it can be considered for acceptance. Specific comments are as follows:

 

1. In the abstract section, the percentage improvement of the proposed model compared with the SOTA methods should be explained.
2. At the end of the abstract, the research value and significance need to be expanded.
3. In research papers, authors are not advised to use too much "We", which is too subjective.
4. At the end of the introduction, the remaining structure should be clarified.
5. It is not recommended to cite a large number of references in the same place.
6. Figure 2 is not academic and professional enough and needs to be redrawn.
7. The comparative experiment is too simple and lacks the advanced model of the latest year. The current research results are not convincing.
8. More discussion should be added about the application value.
9. The limitations of the study are not explicitly mentioned.
10. Language editing is necessary.
11. Authoritative research belonging to this journal in recent three years can be added.