Multimodal Deep Search · 2026

DeepVoyager-VL

Incentivizing Vision-in-the-Loop Search for Long-Horizon Multimodal Agents

Huanyao Zhang^1◇Jiepeng Zhou^2◇Runhao Zhao^3◇Yanzhe Shan⁴Jiaoyang Chen⁵Bowen Zhou¹Bo Li¹Fang Wang¹Jialong Wu¹Zhengwei Tao¹Lang Mei⁶Xiaohan Yu⁶Liyan Liu⁶Chong Chen^6†Wentao Zhang^1†

◇ Equal contribution Project leader† Corresponding authors

¹Peking University

²HKUST (GZ)

³NUDT

⁴OUC

⁵HITSZ

⁶Huawei Cloud BU

Vision in the LoopLong-Horizon SearchSupervised Fine-TuningMultimodal Event Graph

Paper

Code

Dataset

DeepVoyager-VL-8B

DeepVoyager-VL-30B

01 / Overview

Vision in the loop, not only at the input.

The supervision gap. As Figure 1(a)–(c) shows, existing methods build the reasoning chain in text first and place vision at the input or near the answer, so acquired images rarely drive later retrieval.

EventVoyage-VL is our structure-before-language synthesis pipeline and resulting vision-in-the-loop data source. As Figure 1(d)–(e) shows, it creates long-horizon questions with intermediate visual dependencies; the curated trajectories are then distilled into DeepVoyager-VL.

Comparison of four multimodal search data synthesis paradigms and a long-horizon vision-in-the-loop example — Figure 1. Comparison of multimodal search data synthesis paradigms. Prior methods place vision at the input through entity substitution (a), concentrate visual reasoning before text-based search (b), or graft visual evidence near the answer (c). We instead synthesize vision-in-the-loop questions from a visually enriched multimodal event graph (d), as illustrated by a representative long-horizon example (e).

Five-stage overview of DeepVoyager-VL from multimodal event graph construction through supervised fine-tuning — Figure 2. Overview of DeepVoyager-VL, encompassing vision-in-the-loop data synthesis, difficulty-aware trajectory curation, and supervised agent training.

02 / Performance

Leading multimodal search performance at both scales.

Across ten multimodal search benchmarks, DeepVoyager-VL-8B and DeepVoyager-VL-30B-A3B outperform scale-matched multimodal search agents on eight and nine benchmarks, respectively, while the same framework also transfers consistently across three different base models.

DeepVoyager-VL-8B54.8Average score · best on 8 of 10 benchmarks

DeepVoyager-VL-30B-A3B58.6Average score · best on 9 of 10 benchmarks

Performance across ten multimodal search benchmarks

Accuracy or task-native score; higher is better.

Model	MMSearch	SimpleVQA	LiveVQA	FVQA	BC-VL	MM-BC	MMSearch+	VDR	BC-V³	VisBrowse	Avg.
Direct Answer
GPT-5.5	68.7	67.0	73.0	66.7	47.9	17.5	20.3	18.6	23.0	36.1	43.9
Gemini-3.1-Pro	64.2	64.1	65.0	58.9	41.4	11.5	26.4	15.6	19.3	23.7	39.0
Claude-Opus-4.6	59.8	71.7	53.1	60.1	43.5	13.2	13.2	15.4	15.0	27.2	37.2
Qwen3-VL-30B-A3B-Instruct	18.7	53.2	42.7	34.7	29.6	4.0	3.2	3.8	6.0	11.2	20.7
Qwen3-VL-8B-Instruct	15.2	44.7	41.0	28.0	25.1	4.9	3.2	2.8	1.0	8.9	17.5
Agentic Workflow
GPT-5.5	82.7	82.3	90.3	84.3	68.2	51.9	48.6	42.0	55.0	66.3	67.2
Gemini-3.1-Pro	82.7	81.0	87.3	81.3	65.9	44.4	51.5	40.8	50.0	62.1	64.7
Claude-Opus-4.6	81.7	81.7	88.0	79.3	63.2	49.4	52.1	36.6	55.0	62.7	65.0
Qwen3-VL-30B-A3B-Instruct	64.7	71.0	73.3	72.3	41.6	9.9	17.7	21.6	11.3	23.1	40.7
Qwen3-VL-8B-Instruct	61.7	60.0	67.7	69.7	34.6	5.8	13.5	16.8	7.7	15.4	35.3
Multimodal Deep Search Agents · 8B
MMSearch-R1-7B	53.8	57.4	48.4	58.4	—	—	—	—	4.0	—	—
WebWatcher-7B	49.1	54.3	51.2	—	21.2	—	—	—	4.7	—	—
DeepEyesV2-7B	63.7	59.4	—	60.6	—	—	—	—	—	—	—
SenseNova-MARS-8B	67.8	70.2	56.2	67.1	—	—	—	—	—	—	—
Vision-DeepResearch-8B	69.6	—	76.7	64.7	42.6	—	20.4	29.2	11.7	—	—
MM-DeepResearch-8B	67.8	65.9	65.0	69.2	37.9	—	—	—	—	—	—
POINTS-Seeker-8B	70.8	68.8	77.7	71.2	44.4	—	25.2	—	—	—	—
OpenSearch-VL-8B	64.5	71.6	59.6	71.5	37.6	—	—	20.8	—	—	—
SimpleSearch-VL-8B	77.1	76.6	75.2	76.8	52.1	—	32.5	—	—	—	—
Visual-Seeker-8B	72.2	—	—	—	47.6	16.1	27.3	—	—	34.7	—
DeepVoyager-VL-8B	72.7	76.3	82.7	82.7	58.4	24.0	37.1	35.0	32.3	47.3	54.8
Multimodal Deep Search Agents · 30B
WebWatcher-32B	55.3	59.0	58.7	—	27.0	—	—	—	8.7	—	—
SenseNova-MARS-32B	74.3	74.1	60.8	72.6	—	—	—	—	—	—	—
Skywork-R1V4-30B-A3B	66.1	—	—	67.2	38.4	—	—	—	—	—	—
Vision-DeepResearch-30B-A3B	69.6	—	77.6	74.2	53.7	—	28.5	37.8	—	—	—
REDSearcher-MM-30B-A3B †	72.9	—	79.3	—	57.2	23.5	26.6	—	—	—	—
MM-DeepResearch-32B	69.0	67.6	68.0	70.1	43.0	—	—	—	—	—	—
LMM-Searcher-30B-A3B †‡	71.0 / 72.3	—	—	—	—	22.3 / 30.1	32.9 / 34.8	—	—	42.0 / 48.3	—
OpenSearch-VL-30B-A3B	68.7	74.9	67.4	73.2	41.1	—	—	33.5	—	—	—
SimpleSearch-VL-30B-A3B	83.6	79.6	81.1	79.0	55.9	—	34.4	—	—	—	—
DeepVoyager-VL-30B-A3B	74.0	81.0	82.7	84.7	64.2	30.5	40.6	39.4	35.0	53.8	58.6

Table 1.Performance comparison across ten multimodal information-seeking benchmarks. Within each comparison group, the best and second-best results are bolded and underlined, respectively. † denotes initialization from a Thinking checkpoint; ‡ denotes LMM-Searcher's 30-turn / 100-turn settings (x / y).

Framework comparison across base models

All framework evaluations use up to 50 interaction turns.

Base model	Evaluation method	MMSearch	MM-BC	MMSearch+	VisBrowse	Avg.
GPT-5	Direct Answer	33.3	10.3	19.1	26.0	22.2
	Vision-DeepResearch workflow	63.7	—	17.2	—	—
	LMM-Searcher agentic search	72.2	23.7	34.8	35.5	41.6
	DeepVoyager-VL (Ours)	79.7	49.8	50.8	59.2	59.8
Gemini-2.5-Pro	Direct Answer	39.8	10.3	14.5	27.2	23.0
	Vision-DeepResearch workflow	69.0	—	22.2	—	—
	LMM-Searcher agentic search	66.3	12.1	28.1	16.0	30.6
	DeepVoyager-VL (Ours)	72.0	19.0	33.4	43.8	42.1
Qwen3-VL-30B-A3B-Thinking	Direct Answer	17.7	7.1	2.7	13.0	10.1
	Vision-DeepResearch workflow	53.2	—	13.6	—	—
	LMM-Searcher agentic search	62.0	9.8	14.4	16.0	25.6
	DeepVoyager-VL (Ours)	70.0	10.0	20.9	20.7	30.4

Table 2. Performance comparison of different multimodal search frameworks across three base models. The average is computed over four benchmarks; “—” denotes unreported results.

03 / Ablation study

Both the data and the agent design matter.

Adding 7K vision-in-the-loop trajectories yields further average gains of +5.4 for 8B and +6.5 for 30B, while removing any core agent component lowers average performance by 2.5–4.0 points.

Training-data ablation

Setting	BC-VL	BC-V³	VisBrowse	Avg.
Qwen3-VL-8B (Agentic)	34.6	7.7	15.4	19.2
+ 20K multi-source trajectories	54.4	26.7	40.8	40.6
+ 7K VIL trajectories	58.4	32.3	47.3	46.0
Qwen3-VL-30B-A3B (Agentic)	41.6	11.3	23.1	25.3
+ 20K multi-source trajectories	62.2	28.7	42.6	44.5
+ 7K VIL trajectories	64.2	35.0	53.8	51.0

Table 3. Cumulative training-data ablation: 20K multi-source trajectories followed by 7K VIL trajectories.

Framework component ablation

Setting	FVQA	VDR	BC-V³	Avg.
Full DeepVoyager-VL	78.0	36.0	39.3	51.1
w/o Summary	74.0	35.0	32.3	47.1
w/o Image Search	77.7	35.4	32.7	48.6
w/o Fetch Image	76.3	35.6	33.3	48.4
w/o Crop Image	75.7	33.0	36.3	48.3

Table 4. DeepVoyager-VL component ablation with Qwen3.6-35B-A3B; each setting removes one component.

04 / Findings

The learned behavior is longer and more visual.

Under unified rollouts, EventVoyage-VL yields 64.3% visual-tool calls and peaks at 16–20 turns, compared with at most 40.6% visual-tool use and 1–10-turn peaks in prior datasets.

Visual tool engagement64.3%

of tool calls are visual, versus 40.6%, 30.3%, and 10.1% in three public datasets.

VIL trajectory gain+6.5

average points from 7K VIL trajectories on the 30B model, beyond a strong 20K mixture.

Training recipeSFT only

achieves 54.8 and 58.6 average scores without an additional reinforcement-learning stage.

Tool-call proportions showing greater visual tool use in EventVoyage-VL — Figure 3. Trajectory statistics from unified Doubao-2.0-Pro rollouts on 1,000 examples per dataset. **(a)** Tool-call proportions by functional category. (b) Trajectory counts in five-turn intervals. OSVL, VDR, and RED-MM denote OpenSearch-VL, Vision-DeepResearch, and REDSearcher-MM, respectively.

Trajectory turn distribution showing a longer interaction horizon for EventVoyage-VL — Figure 3. Trajectory statistics from unified Doubao-2.0-Pro rollouts on 1,000 examples per dataset. **(a)** Tool-call proportions by functional category. (b) Trajectory counts in five-turn intervals. OSVL, VDR, and RED-MM denote OpenSearch-VL, Vision-DeepResearch, and REDSearcher-MM, respectively.

05 / BibTeX

Cite DeepVoyager-VL.

@misc{zhang2026deepvoyagervlincentivizingvisionintheloopsearch,
      title={DeepVoyager-VL: Incentivizing Vision-in-the-Loop Search for Long-Horizon Multimodal Agents}, 
      author={Huanyao Zhang and Jiepeng Zhou and Runhao Zhao and Yanzhe Shan and Jiaoyang Chen and Bowen Zhou and Bo Li and Fang Wang and Jialong Wu and Zhengwei Tao and Lang Mei and Xiaohan Yu and Liyan Liu and Chong Chen and Wentao Zhang},
      year={2026},
      eprint={2608.01827},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.01827}, 
}