Visual Retrieval-Augmented Generation (VRAG) empowers Vision-Language Models to retrieve and reason over visually rich documents. To tackle complex queries requiring multi-step reasoning, agentic VRAG systems interleave reasoning with iterative retrieval. However, existing agentic VRAG faces two critical bottlenecks. (1) Visual Evidence Sparsity: key evidence is scattered across pages yet processed in isolation, hindering cross-page reasoning; moreover, fine-grained intra-image evidence often requires precise visual actions, whose misuse degrades retrieval quality; (2) Search Drift in Long Horizons: the accumulation of visual tokens across retrieved pages dilutes context and causes cognitive overload, leading agents to deviate from their search objective. To address these challenges, we propose VISOR (Visual Retrieval-Augmented Generation via Iterative Search and Over-horizon Reasoning), a unified single-agent framework. VISOR features a structured Evidence Space for progressive cross-page reasoning, coupled with a Visual Action Evaluation and Correction mechanism to manage visual actions. Additionally, we introduce a Dynamic Trajectory with Sliding Window and Intent Injection to mitigate search drift. They anchor the evidence space while discarding earlier raw interactions, preventing context from being overwhelmed by visual tokens. We train VISOR using a Group Relative Policy Optimization-based Reinforcement Learning (GRPO-based RL) pipeline with state masking and credit assignment tailored for dynamic context reconstruction. Extensive experiments on ViDoSeek, SlideVQA, and MMLongBench demonstrate that VISOR achieves state-of-the-art performance while maintaining reasonable and controllable computational costs for long-horizon visual reasoning tasks. The source code is available at https://github.com/syc1336/VISOR.
Figures & tables
Figure 1. Two critical bottlenecks in agentic VRAG: (Top) Visual Evidence Sparsity , with relevant clues scattered across pages or small image regions; (Bottom) Long-Horizon Search Drift , where growing visual context obscures prior evidence and the original goal.
Figure 2. Overview of VISOR . At each step i , the agent produces a ⟨ think ⟩ … ⟨ action ⟩ response. Evidence extracted from the reasoning trace is accumulated in a structured Evidence Collection space E . The action space comprises three operations: search , crop , and answer . The agent loop context is reconstructed at each turn: the user query and E are always pinned at the top, while only the last W turns of raw interactions are retained via a Sliding Window ; each turn’s observation also includes an Intent Injection prompt to re-anchor the agent’s focus.
Method
SlideVQA
ViDoSeek
MMLongBench
Single-hop
Multi-hop
Overall
Extraction
Logic
Overall
Text
Table
Chart
Figure
Layout
Overall
Qwen2.5-VL-7B
Vanilla RAG ⋆ ( Faysse et al., 2024 )
29.10
17.40
26.10
26.40
41.30
32.88
13.10
14.70
15.90
4.30
7.60
–
ReAct ⋆ ( Yao et al., 2022 )
34.80
20.40
31.11
27.50
42.10
33.85
10.10
12.40
10.20
6.20
7.10
–
ViDoRAG † ( Wang et al., 2025a )
72.15
39.86
63.88
66.05
72.83
69.00
24.40
23.96
21.91
24.14
20.34
25.50
M3RAG †⋆ ( Du and Li, 2026 )
–
–
65.82
–
–
69.36
–
–
–
–
–
–
Table 1. Main results on SlideVQA, ViDoSeek, and MMLongBench. We report accuracy (%). † denotes multi-agent architectures. ‡ denotes fine-tuned models. ⋆ denotes results taken from published papers under the same or comparable experimental settings. The best result in each column is bolded and the second-best is underlined .
Method
ViDoSeek
SlideVQA
Vanilla
Fine-tuned
Vanilla
Fine-tuned
VISOR (Full)
57.88
74.87
54.00
72.37
w/o II
49.04
69.00
43.52
69.07
w/o ES
39.58
47.81
39.10
51.33
w/o SW
53.94
70.93
47.47
67.54
w/o ES & SW
42.11
68.48
40.81
64.61
Table 2. Ablation study of VISOR on Qwen2.5-VL-7B on SlideVQA and ViDoSeek. We report overall accuracy (%). II = Intent Injection, ES = Evidence Space with Visual Action Evaluation and Correction, SW = Sliding Window.
Figure 3. Accuracy of Qwen2.5-VL-7B on SlideVQA (500 samples) as a function of the number of retrieved reference images. The upper bound (67.0%) is the accuracy when correct reference images are directly provided to the model.
Method
Completeness (%)
Avg. Retrieved
EVisRAG ( Sun et al., 2025 )
80.2
3 (fixed)
VRAG-RL ( Wang et al., 2025b )
76.8
1.78
ViDoRAG ( Wang et al., 2025a )
91.3
10 (fixed)
VISOR
84.2
2.34
Table 3. Retrieval completeness (%) and average number of retrieved images per trajectory on SlideVQA (2,215 samples). Completeness = fraction of questions where all ground-truth evidence pages are retrieved.
Figure 4. Breakdown of average per-sample inference latency on ViDoSeek. VISOR introduces a verification step yet remains faster than the multi-agent baseline ViDoRAG and competitive with VRAG-RL.
Method
Invalid Action Rate ↓
Finish Rate ↑
Acc ↑
Vanilla
0.61
74.08
57.88
SFT
0.53
79.42
63.57
SFT+RL
0.09
96.23
74.87
Table 4. Effect of training stage on the model.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Name
Value
Finetuning type
Full
Freeze vision tower
True
Freeze multi-modal projector
True
Freeze language model
False
Cutoff length
16384
Epochs
3
Appendix
Table 5. Key hyperparameters for SFT.
Name
Value
Number of agent groups
5
Warmup steps ratio
0.285
Train batch size
8
Mini batch size per GPU
1
Micro batch size per GPU
1
Learning rate (Actor)
1.0e-6
Appendix
Table 6. Key hyperparameters for RL.
Figure 5. The prompt template used for LLM-as-Judge evaluation.
Judge
Extraction
Logic
Overall
Qwen-max-latest
73.49
76.66
74.87
DeepSeek-V3.2
71.47
78.67
74.61
Appendix
Table 7. Comparison of evaluation scores on ViDoSeek under two judge models.
Figure 6. A representative case illustrating redundant crop usage in VRAG-RL.
Figure 7. A success case of VISOR on a multi-hop SlideVQA question. Retrieved page images are shown at the top for layout reasons; the corresponding reasoning steps are presented below. Step 1 retrieves an irrelevant page from a different document. Step 2 identifies the year 2010 from an Amazon fulfillment center growth chart. Step 3 retrieves a category evolution slide providing context. Step 4 locates the Media vs. Others sales breakdown chart, from which the answer is read. The final verification search (Step 5) returns a page with an unrelated chart from a different document, yet VISOR correctly produces the answer from accumulated evidence.
Figure 8. A failure case of VISOR on a multi-hop MMLongBench question. Both reference pages are retrieved correctly, but the model misreads the legend of a pie chart, confusing the iOS 8 slice (gray, 41%) for iOS 9 (blue, 51%), leading to an incorrect final answer despite successful retrieval.
Figure 9. All prompt templates used in VISOR. Each template corresponds to a distinct interaction event in the agent loop.
Window Size W
SlideVQA
1
70.43
2
72.37
3
71.56
4
68.17
Appendix
Table 8. Sensitivity of VISOR to sliding window size W on the fine-tuned model (SlideVQA). We report overall accuracy (%).