Large Multimodal Models (LMMs) excel at visual perception but struggle with real-time, knowledge-intensive queries due to their reliance on static parametric knowledge. While multimodal search agents offer a promising solution, developing them from vanilla LMMs presents two major challenges: the lack of open recipes to cultivate search agency from scratch, and the severe context explosion (attention dilution) that occurs during long-horizon multi-turn investigations. In this paper, we address both challenges by developing a native multimodal search agent. First, we introduce an open training recipe featuring Agentic Seeding, a formative training stage that bootstraps tool-use and planning abilities directly from a non-agentic foundation. Second, to overcome the context bottleneck, we propose V-Fold, an adaptive memory management mechanism. V-Fold retains recent interactions as high-fidelity text while folding stale historical context into the visual space via rendering, exploiting the model's cross-modal alignment to preserve raw evidence without text token redundancy. Combining these innovations, we present POINTS-Seeker-8B, which achieves state-of-the-art performance among models of comparable scale across six multimodal search benchmarks.
Figures & tables
Figure 1: Overview of POINTS-Seeker. Our model adaptively interacts with external tools to solve complex, multi-hop VQA tasks. By leveraging V-Fold compression, stale history is rendered into compact visual tokens, effectively bypassing long-context performance degradation while achieving efficiency and reasoning fidelity.
Figure 2: The framework of our agentic search system. (a) Training stages, including Agentic Seeding, SFT, and RL. (b) V-Fold, which adaptively compresses stale history into visual tokens for efficient long-context reasoning.
Figure 3: Motivation for history-aware compression. (a) Performance degradation in long-context scenarios. (b) The presence of GT in the failure case history indicates that compression must minimize historical information loss.
Model
MMSearch
MMSearch+
FVQA
SimpleVQA
LiveVQA
BC-VL
Direct Answer
GPT-5
33.3
19.1
57.3
61.7
57.5
47.2
Gemini-2.5 Flash
30.4
8.1
47.7
55.5
51.0
37.1
Qwen3-VL-8B-Instruct
15.2
3.2
28.0
42.9
41.0
25.1
RAG Workflow
GPT-4o
24.1
-
-
61.6
34.0
13.4
Table 1: Comparison with state-of-the-art models.
Seeding
SFT
RL
V-Fold
MMSearch
MMSearch+
FVQA
SimpleVQA
LiveVQA
BC-VL
POINTS-Seeker
✘
✔
✘
✘
63.2
18.0
67.9
65.1
70.8
32.3
✔
✔
✘
✘
66.1
23.0
68.9
66.4
76.8
39.8
✔
✔
✔
✘
69.6
23.4
71.2
68.0
77.3
42.6
✔
✔
✔
✔
70.8
25.2
71.2
68.8
77.7
44.4
Qwen3-VL-8B-Instruct
Table 2: Ablation study on our core design.
Method
Dataset
Acc
Resp. Len.
w/o V-Fold
BC-VL
42.6
9530
V-Fold
BC-VL
44.4 (1.8 ↑ )
7353 (22.8% ↓ )
w/o V-Fold
MMSearch+
23.4
9178
V-Fold
MMSearch+
25.2 (1.8 ↑ )
6871 (25.1% ↓ )
Table 3: Impact of V-Fold on length and accuracy.
Method
FVQA-H
LiveVQA-H
BC-VL-H
No Compression
48.0
48.8
21.3
Discard History
52.0
50.2
21.3
Summarization
50.0
48.8
22.7
Image Rendering
48.0
44.4
17.3
Our V-Fold
54.0
50.7
26.7
Table 4: Comparison under long-horizon settings.
Method
MMBench
MMStar
MMMU
MathVista
Hallu.
AI2D
OCRBench
MMVet
Avg.
POINTS-Baseline
74.9
55.5
52.4
75.7
53.2
81.6
85.1
67.8
68.3
POINTS-Seeker
79.1
59.7
52.6
73.7
54.2
84.8
86.7
73.9
70.6
Table 5: Performance on general VQA benchmarks.
Figure 4: Visualization of POINTS-Seeker. Additional qualitative results are in the supplementary materials.
τ
4k
8k
12k
8k
8k
8k
k
2
2
2
0
2
4
Acc.
25.2
25.2
23.4
23.4
25.2
25.2
Table 6: Ablation study on hyperparameters.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Method
MMSearch
MMSearch+
FVQA
SimpleVQA
LiveVQA
BC-VL
V-Fold RL
70.8
24.3
70.6
68.0
78.5
42.1
Our RL
70.8
25.2
71.2
68.8
77.7
44.4
Appendix
Table 7: Ablation study on RL training strategy.
Parameter
Value
page-size
595×842
dpi
72
margin-x
10
margin-y
10
font-path
Verdana.ttf
font-size
9
Appendix
Table 8: Parameter configuration for image rendering.
Figure 5: Case Study 1: SFT vs. RL. Details are simplified for visualization.
Figure 6: Case Study 2: SFT vs. RL. Details are simplified for visualization.
Figure 7: Tool usage analysis for POINTS-Seeker across six diverse evaluation benchmarks.