acceptodds
Under review as a conference paper at ICLR 2027

AttriSwarm: Orchestrating Agent Swarms for Cross-Page Visual Attribution

Abstract

Multimodal large language models (MLLMs) increasingly address knowledge-intensive tasks by integrating evidence scattered across multiple document pages into a final answer. Because such answers are synthesized from many pages, the link between the answer and its supporting visual evidence can become obscured, hindering verification and trust. Cross-page visual attribution is therefore essential: it requires identifying the supporting pages and localizing fine-grained evidence regions within them, thereby making multi-page reasoning transparent and auditable. Existing approaches either process multiple pages jointly or iteratively refine evidence within a single trajectory, risking cross-page context pollution and, for interactive methods, substantial sequential overhead. We introduce AttriSwarm, an agent-swarm framework that assigns these concerns to separate roles: a leader handles global cross-page reasoning, while subagents perform page-wise attribution in isolation and in parallel. To support training and evaluation, we construct AttriDoc, a cross-page QA dataset linking supporting pages to their annotated evidence regions, and use it for role-specific training of AttriSwarm, improving answer accuracy, page-level , and region-level IoU over its training-free counterpart by an average of 18.94, 29.34, and 23.04 percentage points, respectively, across three open-source backbones and three datasets.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal, Visual Attribution, Agent Swarm, Cross-Page

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_876XOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AttriSwarm, a multi-agent framework for cross-page visual attribution. It uses a leader agent for query decomposition and global reasoning, while parallel subagents independently identify relevant pages and localize supporting evidence regions. The authors also introduce AttriDoc, a dataset with page- and region-level annotations, and develop role-specific SFT and RL training. Experiments across multiple models and datasets show improvements in answer accuracy, attribution quality, and robustness to distractor pages.

Critical strengths
  1. The paper addresses a meaningful problem of cross-page visual attribution, where both the supporting pages and the relevant evidence regions need to be identified.
  2. The experiments cover multiple backbones and datasets, with encouraging improvements in answer accuracy and attribution quality. The new AttriDoc dataset is also a useful contribution.
Critical weaknesses

1. Insufficient insights into the proposed architectural design.

The main idea seems to be a combination of existing techniques. LongAgent [1] has explored leader–worker decomposition for document reasoning, while ORCA [2] and MACT [3] have investigated specialized multi-agent collaboration for visual document understanding. Fine-grained visual source attribution has also been studied in VISA [4].

Combining existing ideas is not a problem in itself. However, I would expect a deeper discussion of why this particular combination works well for cross-page visual attribution. The paper does not sufficiently explain what challenges are unique to this task, or why the proposed coordination between global reasoning and local attribution is more effective than simpler alternatives. More analysis of the components and their interactions would help clarify the contribution.

2. Insufficiently controlled comparisons to establish the benefits of the proposed architecture.

Although AttriSwarm achieves substantial improvements, the current comparisons make it difficult to determine how much of the gain comes from the architecture itself. In Table 2, the trained AttriSwarm is compared mainly against training-free baselines or existing checkpoints without task-specific adaptation. Interestingly, the training-free AttriSwarm does not consistently outperform the single-pass baseline.

Table 3 shows that role-specific training contributes substantially to the improvements, but it remains unclear whether a simpler architecture could achieve similar results with the same training data. Without a training-matched comparison, the advantages of the proposed architecture are not fully established.

[1] Zhao et al. LONGAGENT: Achieving Question Answering for 128k-Token-Long Documents through Multi-Agent Collaboration. EMNLP 2024.

[2] Lassoued et al. ORCA: Orchestrated Reasoning with Collaborative Agents for Document Visual Question Answering. CVPR 2026.

[3] Yu et al. Visual Document Understanding and Reasoning: A Multi-Agent Collaboration Framework with Agent-Wise Adaptive Test-Time Scaling. CVPR 2026.

[4] Ma et al. VISA: Retrieval Augmented Generation with Visual Source Attribution. ACL 2025.

Questions for authors

I suggest the author to provide more detailed architectural ablations and analysis to explain the contributions of context isolation, query decomposition, and leader-based evidence aggregation, particularly their benefits for cross-page reasoning. Also, it's also suggested to include a single-pass or shared-context baseline trained with the same AttriDoc supervision and comparable training budgets to better distinguish architectural improvements from the effects of task-specific training. I would like to reconsider the rate if the rebuttal is sufficient.

5 comments
AuthorsOpenReview ↗

Rebuttal for Weakness 1

W1: Insufficient insights into the proposed architectural design. Discussion in why this particular combination works well for cross-page visual attribution. Why the proposed coordination between global reasoning and local attribution is more effective than simpler alternatives.

Thank you for asking us to clarify the architectural rationale and why these components are combined for cross-page visual attribution. We agree that leader-worker coordination, agent specialization, and visual source attribution have each been studied before. Our design is motivated by the need to satisfy two requirements together: each evidence region must be grounded on the correct page, while the answer may depend on complementary facts distributed across several pages. A page can therefore be relevant even when it cannot answer the question on its own. Missing a necessary page prevents complete attribution and may leave insufficient evidence for answering.

AttriSwarm assigns these requirements to complementary roles. The leader identifies the evidence needed for the question; page-wise attribution agents assess whether each page supports any part of it and localize the corresponding regions in separate contexts. Each attribution agent returns the relevant regions and findings from its assigned page. The leader combines these relevant regions to answer questions that require information from multiple pages. This allows each agent to focus on local evidence while the leader handles cross-page reasoning.

This also explains the choice relative to simpler alternatives: (1) Jointly processing all pages allows cross-page reasoning, but similar content across pages may interfere with local relevance assessment and region localization. Figure 4 in the submission provides system-level evidence of robustness to such distractors. (2) Separating the leader and attribution roles alone does not remove this shared context if the attribution agent still reads all pages together. To examine this distinction, we additionally evaluate a shared-context attribution variant trained to assess and localize evidence across multiple pages in one response; this comparison is reported below in response for Q1. (3) A single attribution agent could process pages one at a time in separate contexts, but sequential processing increases latency as the number of pages grows. AttriSwarm executes these independent page-wise calls in parallel. Figure 5 in the submission directly supports this choice, showing approximately a 4.1× speedup over the serial variant with 32 distractors on successful post-retrieval runs.

To sum up, the contribution of AttriSwarm lies in coordinating isolated, parallel page-wise attribution with global answer synthesis, allowing local evidence to remain linked to its source page while being combined for cross-page reasoning.

AuthorsOpenReview ↗

Rebuttal for Weakness 2 and Question 2

W2&Q2: Insufficiently controlled comparisons to establish the benefits of the proposed architecture. Table 3 shows that role-specific training contributes substantially to the improvements, but it remains unclear whether a simpler architecture could achieve similar results with the same training data. Also, it's also suggested to include a single-pass or shared-context baseline trained with the same AttriDoc supervision and comparable training budgets to better distinguish architectural improvements from the effects of task-specific training.

Thank you for highlighting the need for a trained simpler baseline. We agree that comparisons against training-free baselines alone cannot establish whether a simpler architecture could achieve similar results with task-specific supervision. We therefore added a single-pass baseline using the same Qwen3-VL-4B-Instruct backbone, trained with SFT and GRPO on AttriDoc supervision to generate the answer and visual attributions together.

AttriSwarm relies on page-wise agents to select relevant evidence before global reasoning. Without task-specific training, these agents may incorrectly filter out necessary pages, leaving the leader unable to cite their evidence and potentially lacking information needed to answer. This is one reason why training-free AttriSwarm does not consistently outperform single-pass, which retains all retrieved candidate pages in its input. Role-specific training helps the agents make more reliable relevance and localization decisions.

DatasetMethodAcc.Page F1Reg. IoUJAS
AttriDoc-PaperSingle-pass (trained)4.7836.1123.3114.04
AttriDoc-PaperAttriSwarm (trained)25.8662.5347.9137.40
AttriDoc-SlideSingle-pass (trained)54.3228.7717.9514.62
AttriDoc-SlideAttriSwarm (trained)64.7371.9146.5740.71
CiteVQASingle-pass (trained)34.5531.5716.7511.03
CiteVQAAttriSwarm (trained)45.6153.8031.3024.07

Trained AttriSwarm outperforms the trained single-pass baseline in answer accuracy and all three attribution metrics across the three datasets. Single-pass uses the original query with Top10 retrieval, whereas AttriSwarm uses its full retrieval and coordination pipeline. This comparison supports the complete trained system; the architectural ablations separately assess the contributions of individual design choices.

AuthorsOpenReview ↗

Rebuttal for Question 1

Q1: I suggest the author to provide more detailed architectural ablations and analysis to explain the contributions of context isolation, query decomposition, and leader-based evidence aggregation, particularly their benefits for cross-page reasoning.

Thank you for this suggestion. We agree that more detailed ablations help to explain the contributions of different components. Here, we add three ablation studies.

Context isolation. Context isolation allows each attribution subagent to assess and localize partial evidence on its assigned page without competing page content, while the Leader connects these findings for cross-page reasoning. To examine shared-context attribution as an alternative to page-wise processing, we trained a 4B attribution subagent to jointly assess relevance and localize evidence across candidate pages in one response, rather that isolating different page contexts to attribution agent swarms. We evaluated it on all 669 AttriDoc-Paper questions, retaining the trained leader and the retrieval pipeline with the same retrieval and interaction limits.

Attribution ContextAcc.Page F1Reg. IoUJAS
Trained shared-context attribution8.9734.5511.167.64
AttriSwarm (trained page-wise attribution)25.8662.5347.9137.40

Page-wise attribution achieves higher scores across all four metrics, supporting its effectiveness.

Query decomposition. Query decomposition targets complementary facts required by cross-page questions, helping retrieval collect evidence for different parts of the answer. On the AttriDoc-Paper test set, we restrict retrieval to the original question's Top10 results while retaining the trained Leader and attribution subagents.

Retrieval settingAcc.Page F1Reg. IoUJAS
Original query, Top1022.4258.9848.0336.37
Full AttriSwarm (with query decomposition)25.8662.5347.9137.40

Executing decomposed queries improves answer accuracy, page selection, and JAS. Region IoU is similar in the two settings. Because the retrieved candidate pages also change, this experiment measures the contribution of decomposed retrieval within the pipeline, rather than question decomposition in isolation.

Leader-based evidence aggregation. Individual pages may provide only partial facts needed for a cross-page answer. To examine whether independent local answers sufficent, we replay the archived retrieval and attribution schedules, preserving the original boxed overviews and detail-crop configurations. The same trained Leader answers each retained page in a separate branch with its own history. We then concatenate these answers without a final cross-page synthesis step.

DatasetPer-page answers + concatenation: Acc.Full AttriSwarm: Acc.
AttriDoc-Paper2.0925.86
AttriDoc-Slide35.9864.73
CiteVQA32.3445.61

Concatenation collects local answers but does not explicitly compare, reconcile, or combine their facts into a joint conclusion. Independent page-level answers with concatenation perform substantially worse across all three datasets, suggesting that collecting local answers alone without further selection and global understanding is insufficient for these questions. These results support the value of jointly interpreting complementary evidence to produce a coherent answer. The reference scores are from the reported full pipeline, while the concatenation variant replays its archived evidence acquisition schedule.

Reviewer_oqrROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AttriSwarm separates cross-page answer synthesis from visual attribution within individual candidate pages. A leader integrates evidence, while subagents with separate contexts retrieve and ground candidate pages in parallel. The authors introduce AttriDoc-Paper and AttriDoc-Slide, then train both roles with SFT and GRPO. Evaluation includes these datasets and CiteVQA, using answer accuracy, page F1, region IoU, and joint attribution score (JAS). Ablations examine role training, while additional studies consider distractors and parallel execution.

Critical strengths
  1. The paper addresses a useful distinction between producing an answer and showing the visual evidence behind it. On AttriDoc-Slide with Qwen3-VL-8B-Instruct, accuracy rises from 57.14% for Single-pass to 68.43%, while JAS rises from 6.51% to 45.86% (Table 2). These are improvements of 11.29 and 39.35 percentage points, respectively. The JAS gain is particularly encouraging for the attribution task, although JAS itself does not measure answer correctness.

  2. The role ablation helps explain where the grounding improvements come from. On the same Slide setting, training only the attribution agent raises JAS from 11.47% to 38.43%; training both roles reaches 45.86% (Table 3). This is useful evidence that attribution supervision matters, alongside the out-of-domain CiteVQA results.

Critical weaknesses
  1. Role-specific training is part of the proposed method, and Table 3 measures its contribution. However, the controls do not isolate context separation under matched supervision. In the 8B Slide setting, training-free AttriSwarm scores 45.50%, Single-pass 57.14%, and trained AttriSwarm 68.43%. The 22.93-point training-associated gain and 11.29-point baseline advantage do not decompose the causal contributions. Training-free GPT-5.1 also improves Slide accuracy from Single-pass's 65.90% to 72.44%, so architectural utility is not uniformly absent. A matched-supervision comparison would clarify the contribution in the trained system.

  2. Some evaluation populations still need clarification, especially source overlap and failed outputs. Section 3 provides dataset counts but does not quantify source-document or slide-deck overlap between training and test. Section 5.3 mentions excluding empty/error outputs, without clearly identifying each affected denominator. Appendix D.4 explicitly limits latency to successful post-retrieval runs on frozen page sets. The remaining issue is how differing failure rates affect comparisons within that diagnostic setting.

Questions for authors
  1. Is there an existing shared-context or Single-pass control with comparable supervision and inference budgets? If not, please distinguish evidence for the complete trained system from evidence specifically isolating context separation and filtering.

  2. Are the splits disjoint by source paper and slide deck? Could you report source overlap, handling of repeated tables or related questions, and retrieval coverage?

  3. For the distractor study, which metrics exclude empty/error outputs, and what are the eligible and valid-output counts per method? Given the disclosed successful-run latency setting, are failure rates and failed-attempt costs available to assess whether the comparison changes when failures are included?

I will further consider my score in light of the other reviewers’ comments and whether the authors’ response adequately addresses my concerns.

4 comments
AuthorsOpenReview ↗

Rebuttal for Weakness 1 and Question 1

W1&Q1: Role-specific training is part of the proposed method, and Table 3 measures its contribution. However, the controls do not isolate context separation under matched supervision. A matched-supervision comparison would clarify the contribution in the trained system. Is there an existing shared-context or Single-pass control with comparable supervision and inference budgets?

Thank you for highlighting the need to distinguish the effectiveness of the complete trained system from the contribution of context separation. We added two complementary comparisons using Qwen3-VL-4B-Instruct.

Trained single-pass baseline. We trained a single-pass model with SFT and GRPO on AttriDoc supervision to generate answers and visual attributions jointly.

DatasetMethodAcc.Page F1Reg. IoUJAS
AttriDoc-PaperSingle-pass (trained)4.7836.1123.3114.04
AttriDoc-PaperAttriSwarm (trained)25.8662.5347.9137.40
AttriDoc-SlideSingle-pass (trained)54.3228.7717.9514.62
AttriDoc-SlideAttriSwarm (trained)64.7371.9146.5740.71
CiteVQASingle-pass (trained)34.5531.5716.7511.03
CiteVQAAttriSwarm (trained)45.6153.8031.3024.07

AttriSwarm's advantage persists when the simpler baseline also receives task-specific training. This comparison supports the complete framework: single-pass uses original-query Top10 retrieval, whereas AttriSwarm uses its full retrieval and coordination pipeline.

Trained shared-context attribution. To examine context organization more directly, we trained an attribution subagent to jointly assess relevance and localize evidence across multiple candidate pages in one response. On all 669 AttriDoc-Paper questions, we retained the trained Leader and retrieval pipeline with the same retrieval and interaction limits.

Attribution contextAcc.Page F1Reg. IoUJAS
Shared-context (trained)8.9734.5511.167.64
Page-wise, isolated (AttriSwarm)25.8662.5347.9137.40

Page-wise attribution performs better across all four metrics. This provides complementary evidence for its effectiveness within the trained pipeline, although training and visual-input differences, together with shared-context format failures, prevent attributing the entire gap to context separation alone.

AuthorsOpenReview ↗

Rebuttal for Weakness 2 and Question 2&3

W2&Q2&Q3: Some evaluation populations still need clarification, especially source overlap and failed outputs. Are the splits disjoint by source paper and slide deck? Could you report source overlap, handling of repeated tables or related questions, and retrieval coverage? For the distractor study, which metrics exclude empty/error outputs, and what are the eligible and valid-output counts per method? Given the disclosed successful-run latency setting, are failure rates and failed-attempt costs available to assess whether the comparison changes when failures are included?

Thank you for the opportunity to clarify the statistics regarding the dataset and evaluation populations.

We applied question-level deduplication and confirmed that no identical questions occur between the data actually used for training and the test sets. AttriDoc-Slide has no overlap in source slide decks or evidence pages. For AttriDoc-Paper, some source papers and tables are shared across splits, while the questions target different facts or comparisons.

We report retrieval coverage for the 4B setting below, measured before page filtering over all test questions. Any-GT coverage requires at least one annotated evidence page to be retrieved, while All-GT coverage requires all annotated evidence pages.

DatasetRetrieval settingAvg. retrieved pagesAny-GT coverageAll-GT coverage
AttriDoc-PaperOriginal query, Top1010.0096.86%42.30%
AttriDoc-PaperOriginal query, Top2020.0098.36%58.00%
AttriDoc-PaperAttriSwarm20.0498.95%59.34%
AttriDoc-SlideOriginal query, Top1010.0098.59%73.54%
AttriDoc-SlideOriginal query, Top2020.0098.94%81.66%
AttriDoc-SlideAttriSwarm19.2199.29%83.25%
CiteVQAOriginal query, Top1010.0081.11%54.81%
CiteVQAOriginal query, Top2020.0085.45%62.98%
CiteVQAAttriSwarm20.5091.49%63.91%

Decomposed retrieval provides complementary evidence and improves retrieval coverage. The original-query Top20 control already achieves relatively close All-GT coverage, while the architectural ablations provide separate evidence for the benefits of subsequent page-wise attribution and global evidence integration. Together, these results suggest that the gains reflect both evidence acquisition and downstream evidence processing.

For the distractor study, in the 7B AttriSwarm runs with 8, 16, and 32 distractors, respectively, 1, 1, and 2 cases were excluded before leader inference because grounding outputs unparseable. The corresponding evaluation populations were 244, 244, and 243 questions; all other settings retained all 245 eligible questions.

AuthorsOpenReview ↗

Gentle Follow-up on Rebuttal

Dear Reviewer oqrR,

Thank you again for your thoughtful feedback and the time you have devoted to reviewing our paper. We have carefully addressed your comments in our rebuttal and revised manuscript.

As the discussion period is drawing to a close, we would greatly appreciate your thoughts on whether our responses have addressed your concerns. We would be happy to clarify any remaining points.

Thank you again for your time and consideration.

Best regards,

The Authors

Thank you for the additional experiments and clarifications. The trained single-pass baseline and the MMLongBench-Doc results strengthen the evidence for the complete AttriSwarm system. The reported exclusion counts also address my concern about the distractor-study denominators.

My main concern remains the evidence for page-wise context isolation. Table 16 compares a trained single-pass baseline using Top10 retrieval with AttriSwarm’s full retrieval and coordination pipeline. On AttriDoc-Paper, increasing the original-query retrieval budget to Top20 already raises all-evidence-page coverage to 58.00%, close to AttriSwarm’s 59.34%. The final answer and attribution results for a trained Top20 baseline are still missing. Table 17 holds the trained Leader and retrieval pipeline fixed, but the two variants differ in training and visual inputs, and the shared-context variant has unquantified format failures. These factors leave the contribution of context isolation under comparable inference budgets unresolved.

The authors also confirm that some source papers and tables are shared across the AttriDoc-Paper splits. The extent of that overlap and performance on unseen source papers remain unreported, limiting what can be concluded about generalization from this benchmark.

I appreciate the authors’ efforts and the additional evidence provided during the rebuttal. I also share Reviewer PxQA’s concern about how much methodological advance AttriSwarm offers over prior work and whether its gains on long-document QA hold under comparable inference budgets. Given these remaining concerns, I will maintain my original score.

Reviewer_PxQAOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies cross-page visual attribution, jointly producing answers, supporting pages, and evidence regions. AttriSwarm uses a leader for query decomposition and answer synthesis, with parallel, context-isolated subagents for page relevance assessment and region localization. The authors introduce AttriDoc-Paper and AttriDoc-Slide and train the two roles separately through SFT and GRPO. Experiments evaluate answer and attribution quality across multiple models and three datasets, alongside analyses of role training, distractor robustness, parallel execution, and evidence support.

Critical strengths

1.A useful task and a coherent system. Providing page-level and region-level evidence makes document answers easier to verify. Separating local attribution from global synthesis is well motivated, and the AttriDoc annotations provide additional training and evaluation resources.

2.Meaningful empirical improvements. Role-specific training improves all four metrics across the nine open-source backbone–dataset combinations. CiteVQA provides external transfer evidence, and some configurations improve without task-specific training. The role-training ablations, fixed-page experiments, and evidence-support analysis usefully complement the main results.

Critical weaknesses
  1. The methodological contribution appears incremental relative to closely related work. MACT [1] already combines specialized document agents with SFT and GRPO; VisRAG 2.0 uses per-page evidence recording, cross-page reasoning, and trajectory-based training; MCite-RL uses teacher-generated visual-citation trajectories followed by SFT and GRPO. WideSeek-R1 also establishes isolated, parallel leader–worker execution. AttriSwarm’s more specific contribution is their integration for cross-page region attribution, with dedicated workers and new annotations. This is a useful extension, but I find the methodological advance and transferable insight insufficiently developed. A clearer positioning against these precedents, including the currently undiscussed MACT, is needed to establish the contribution’s significance.
  2. The practical significance of the additional capability remains insufficiently established. The answer-accuracy gains and CiteVQA transfer are valuable, but the evaluation remains centered on visual attribution. Meanwhile, DocLens [2], which is not discussed, already uses page and region localization to improve general long-document QA. Evidence that AttriSwarm’s particular integration improves such tasks, under their native answer/evidence metrics and comparable resource budgets, would strengthen its practical significance. MMLongBench-Doc [3] or GDP.pdf [4] would be useful examples. The paper’s attribution scope is legitimate; my concern is how much additional capability or generalizable understanding this integration contributes in light of related work.

[1] Yu, Xinlei, et al. "Visual document understanding and reasoning: A multi-agent collaboration framework with agent-wise adaptive test-time scaling." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

[2] Zhu, Dawei, et al. "Doclens: A tool-augmented multi-agent framework for long visual document understanding." Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

[3] Ma, Yubo, et al. "Mmlongbench-doc: Benchmarking long-context document understanding with visualizations." Advances in Neural Information Processing Systems 37 (2024): 95963-96010.

[4] Garre, Suhaas, et al. "GDP. pdf: Benchmarking Grounded Multimodal Reasoning over Professional PDF Documents." arXiv preprint arXiv:2607.11192 (2026).

Questions for authors

1.Relative to MACT, VisRAG 2.0, MCite-RL, and existing parallel agent frameworks, which limitation is addressed specifically by AttriSwarm, and what evidence best demonstrates the resulting methodological or empirical advance?

2.What evidence supports the usefulness of the same framework and trained roles for general long-document tasks whose final outputs need not include bounding boxes? How does its intended capability differ from localization-based document systems such as DocLens?

4 comments
AuthorsOpenReview ↗

Rebuttal for Weakness 1 and Question 1

W1&Q1: The methodological contribution appears incremental relative to closely related work. A clearer positioning against these precedents, including the currently undiscussed MACT, is needed to establish the contribution's significance. Relative to MACT, VisRAG 2.0, MCite-RL, and existing parallel agent frameworks, which limitation is addressed specifically by AttriSwarm, and what evidence best demonstrates the resulting methodological or empirical advance?

Thank you for encouraging us to clarify our contribution. We have added a discussion of MACT to the revised manuscript. MACT specializes agents by reasoning function, with verification and adaptive test-time scaling; VisRAG 2.0 records per-image evidence within a unified generation process; and MCite-RL refines citations through sequential retrieval and cropping. Building on these precedents, AttriSwarm addresses a specific challenge in cross-page attribution: isolating local relevance assessment and region localization while preserving the global evidence relationships needed for answering. The Leader specifies complementary evidence targets, page-wise workers independently assess relevance and localize supporting regions in parallel, and the Leader integrates their findings into the answer and final attributions. This coordination limits each worker's exposure to competing page content while reducing sequential processing overhead. Our original experiments support these design objectives: Figure 4 provides system-level evidence of robustness to increasing distractors, while Figure 5 demonstrates the efficiency benefit of parallel over serial attribution on successful post-retrieval runs. The additional component ablations in Appendix F further examine shared-context attribution, decomposed retrieval, and cross-page evidence aggregation.

AuthorsOpenReview ↗

Rebuttal for Weakness 2 and Question 2

W2&Q2: The practical significance of the additional capability remains insufficiently established. Show whether the framework benefits general long-document QA when bounding-box attribution is not required, ideally under comparable resource budgets.

Thank you for highlighting the distinction between attribution performance and broader usefulness in long-document QA. We acknowledge DocLens as closely related work and have added a discussion of it in the revision. DocLens uses OCR-assisted page navigation, layout-based element localization, and answer sampling-adjudication for long-document QA. AttriSwarm emphasizes query-conditioned relevance assessment and region localization in isolated page contexts, coordinated by a Leader that integrates complementary evidence across pages.

To test whether this framework remains useful when bounding-box outputs are not required, we evaluated the existing trained models on MMLongBench-Doc questions, without additional benchmark-specific training. Attribution serves as an internal evidence-processing step; only answer quality is scored using the benchmark’s official answer-extraction prompt and native generalized accuracy and F1.

MethodAcc.F1
Single-pass (trained, Top10)32.2733.85
AttriSwarm (full pipeline)38.6140.76

Both methods use the same Qwen3-VL-4B-Instruct backbone and search within the document specified by each question. Single-pass uses original-query Top10 retrieval, while AttriSwarm retains its full retrieval and coordination pipeline. These results show that the complete framework improves answer quality on an external long-document QA benchmark, supporting its usefulness beyond settings that require visual-attribution outputs.

AuthorsOpenReview ↗

Gentle Follow-up on Rebuttal

Dear Reviewer PxQA,

Thank you again for your constructive comments, which have helped us improve the paper. We have provided detailed responses to your concerns in our rebuttal.

With the discussion period nearing its end, we would be grateful for any further feedback you may have. If any questions remain, we would welcome the opportunity to discuss them.

Thank you for your time and consideration.

Best regards,

The Authors

Thank you for the additional experiments and clarifications. However, I remain unconvinced that the proposed integration offers a substantial methodological advance over MACT, VisRAG 2.0, MCite-RL, and existing parallel agent frameworks. The response clarifies the design, but does not sufficiently establish its distinctive contribution beyond these precedents.

More importantly, the new MMLongBench-Doc evaluation has simpler output requirements: only answers are evaluated, and evidence boxes are not required. Even in this setting, the full trained AttriSwarm pipeline reports 38.61% accuracy, below the original Qwen3-VL-4B-Instruct’s officially reported 43.5%. This raises concerns about its broader usefulness: the additional training and multi-agent processing have not demonstrated an advantage even when explicit evidence localization is not required as an output. Although evaluation settings may differ and a matched comparison is needed, the current evidence does not resolve this concern. I therefore maintain my original rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for Modification and Rebuttal

We sincerely thank all reviewers for the time and care they devoted to evaluating our work. Your comments helped us recognize where our explanations and experimental evidence needed to be clearer.

In response, we have expanded the Related Work discussion and added appendix experiments covering a trained single-pass baseline, architectural ablations, and MMLongBench-Doc evaluation. We also clarify dataset splits, retrieval coverage, and evaluation populations in our responses below. We hope these revisions address the concerns raised, and we would greatly appreciate any further questions or suggestions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.