Organizations: Nanjing University of Science and Technology · Singapore Management University · Nanjing Pami Intelligent Technology Co., Ltd. · University of Wisconsin - Madison · Nanjing Forestry University
Document Question Answering (DQA) involves generating answers from a document based on a user's query, representing a key task in document understanding. This task requires interpreting visual layouts, which has prompted recent studies to adopt multimodal Retrieval-Augmented Generation (RAG) that processes page images for answer generation. However, in multimodal RAG, visual DQA struggles to utilize a large number of images effectively, as the retrieval stage often retains only a few candidate pages (e.g., Top-4), causing informative but less visually salient content to be overlooked in favor of common yet low-information pages. To address this issue, we propose a Multi-Armed Bandit-based DQA framework (MAB-DQA) to explicitly model the varying importance of multiple implicit aspects in a query. Specifically, MAB-DQA decomposes a query into aspect-aware subqueries and retrieves an aspect-specific candidate set for each. It treats each subquery as an arm and uses preliminary reasoning results from a small number of representative pages as reward signals to estimate aspect utility. Guided by an exploration-exploitation policy, MAB-DQA dynamically reallocates retrieval budgets toward high-value aspects. With the most informative pages and their correlations, MAB-DQA generates the expected results. On four benchmarks, MAB-DQA shows an average improvement of 5%-18% over the state-of-the-art method, consistently enhancing document understanding. Codes are available at https://github.com/ElephantOH/MAB-DQA.
Figures & tables
Figure 1: Aspect Retrieval Degradation in DQA. (a) Ground-truth evidence and answer. (b-c) Similarity heatmap between query and patches. (d) Aspects of low importance in the question. (e-f) High spurious similarity on irrelevant pages due to frequent terms. (g-h) Aspects of low importance’ aggregated score can exceed crucial evidence.
Figure 2: The overview of our proposed framework MAB-DQA. (a) Decompose a query into aspect-aware subqueries and model the relationships using hyperedges to construct a hypergraph; (b) For the hypergraph’s jumping retrieval pages, model each subquery as an arm and use the Bandit-based method to select the next page based on VLM feedback; (c) Derive the final answer through reflective reasoning.
Dataset
Document
Question
Issue
MMLongBench
134
1073
212 (19.8%)
LongDocURL
396
2325
647 (27.8%)
PaperTab
307
393
-
FetaTab
871
1016
-
Table 1: Statistics of Datasets. "Issue" refers to the number of samples in the labeled dataset that the Colpali model may be ignoring regarding key query conditions.
Model
RAG Method
MMLongBench
LongDocURL
FetaTab
PaperTab
Avg
Qwen-2.5-VL-7B
Direct
0.204
0.398
0.350
0.112
0.266
LLaVA-1.6-7B
Direct
0.176
0.110
0.301
0.102
0.172
MDocAgent
ColPali + ColBert
0.315
0.527
0.598
0.227
0.417
M3DocRAG
ColPali (Top-4)
0.296
0.503
0.537
0.152
0.372
MoloRAG
ColPali (Top-4)
0.371
0.536
0.554
0.157
0.405
MoloRAG+
MoloRAG+ (Top-4)
0.372
0.528
0.600
0.195
0.424
Table 2: Comparison of our method with DQA approaches. Pure VLM-based, Multi-Agent-based, and RAG-based methods are evaluated using accuracy (%).
Top-K
Method
MMLongBench
LongDocURL
Recall
Precision
NDCG
MRR
Recall
Precision
NDCG
MRR
1
M3DocRAG
45.31
56.80
56.80
56.80
46.82
64.51
64.45
64.51
MDocAgent (ColBert)
30.65
40.50
40.50
40.50
40.72
56.31
56.31
56.31
MDocAgent (Colpali)
46.61
59.86
59.86
59.86
46.90
64.18
64.18
64.18
MoLoRAG
48.93
64.11
64.11
64.11
49.59
67.84
67.84
67.84
MoLoRAG+
50.30
64.82
64.82
64.82
48.70
66.90
66.90
66.90
Table 3: Retrieval Performance Comparison on MMLongBench and LongDocURL Benchmarks under Top-K ( K=1,3,5 ) Settings. All values are in %. The best results are in bold.
Model Variant
QD
MABR
HRRA
MMLongBench
LongDocURL
FetaTab
PaperTab
Avg. Imp.
Colpali (Baseline)
×
×
×
0.296
0.554
0.537
0.152
0.0%
w/ MABR
∘
∘
×
0.388
0.543
0.609
0.226
22.8%
w/ HRRA
∘
×
∘
0.395
0.561
0.624
0.236
26.5%
MAB-DQA (Ours)
∘
∘
∘
0.399
0.564
0.638
0.269
33.1%
Table 4: Ablation Study on the Proposed MAB-DQA Modules. Additional ablation in the Appendix.
Backbone
Method
MMLongBench
LongDocURL
Recall
Precision
NDCG
MRR
Recall
Precision
NDCG
MRR
Qwen2.5-VL-7B
MoloRAG
67.40
33.18
39.70
70.66
69.58
35.27
42.30
75.60
LLaVa-13B
65.46
32.23
38.05
67.00
-
-
-
-
Qwen3-30B-A3B
68.41
34.08
40.13
70.78
69.80
36.02
42.47
75.77
Qwen3-32B
72.67
35.46
42.23
73.34
70.57
35.93
43.40
77.95
Qwen2.5-VL-7B
Ours
69.53
34.32
41.05
72.94
70.46
35.73
43.07
77.78
Table 5: Evaluation on MMLongBench and LongDocURL across VLM Backbones and Methods. Retrieval performance was evaluated based on Top-3 settings.
Figure 3: Sensitivity Analysis of Key Hyperparameters of the MAB-DQA Framework under Top-3 Retrieval on LongDocURL. The blue dashed line represents Precision. The red line represents Recall.
Figure 4: Comparison of Average Inference Time Across Different Models on 4 NVIDIA V100 GPUs, with retrieval time (Top-10) in blue and QA time (Top-4) in orange.
Figure 5: Qualitative Analysis of the MAB-DQA Framework.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Qualitative Analysis of the MAB-DQA Framework. We selected a representative query that requires extensive page regression. The evidence for the problem spans 13 pages (specifically required: 10, 12, 14, 15, 20, 21, 22, 23, 24, 25, 26, 30, 37). It is evident that this query contains a large number of correct pages. This example demonstrates that our method, in the presence of positive samples, recalls a broader range of correct pages more comprehensively. As shown in Fig. (a), MAB-DQA, by providing query decomposition, recalls more correct evidence pages. As shown in Fig. (e), when extensive recall is required, distractor pages are usually contained within a small number of hyperedges.
Figure 7: Qualitative Analysis of the MAB-DQA Framework. We selected a representative query that requires a multi-hop answer. The evidence for the problem spans 3 pages (specifically required: 22, 24, 25). This query requires the QA framework to both accurately identify evidence and completely recall it. As shown in Fig. (b), the initial scores provided by Colpali include distractor pages (35, 1). As shown in Fig. (d), through multiple rounds of online evaluation, MAB-DQA correctly excludes the distractor pages. MAB-DQA provides an online reranking algorithm that correctly associates all evidence.
Figure 8: Qualitative Analysis of the MAB-DQA Framework. We selected a representative query that requires avoiding distracting conditions. The evidence for the problem spans 2 pages (specifically required: 7, 8). "Guide book" is likely a high-frequency term that appears frequently throughout the entire document. This causes the retrieval engine to consider every page as directly relevant to the evidence for this problem. As shown in Fig. (d), both Colpali and MoloRAG retrieve incorrectly. MAB-DQA avoids the unimportant condition "guide book," enabling the recall of the correct evidence pages.
Model Variant
MMLongBench
LongDocURL
FetaTab
PaperTab
w/o Query-Agnostic Page Graph
0.394
0.560
0.636
0.247
w/o Atomic and Global Subqueries Set
0.317
0.482
0.543
0.198
w/o Question Reflection
0.401
0.556
0.631
0.219
w/o Answer Reflection
0.397
0.561
0.625
0.247
Ours (Full Model)
0.399
0.564
0.638
0.269
Appendix
Table 6: Individual Component Ablation Study on MAB-DQA Framework
Figure 9: Sensitivity Analysis of Key Hyperparameters of the MAB-DQA Framework under Top-3 Retrieval on MMLongBench. The blue dashed line represents Precision. The red line represents Recall.
Human-inspired AI Research, Korea University · Department of Computer Science and Engineering, Korea University · Computer Science and Engineering, Konkuk University