Training multimodal search agents requires tasks that make visually grounded search necessary for determining the answer. Textual identity clues and unused target evidence allow shortcuts in synthetic questions, weakening supervision for visual search and evidence integration. We propose RACE, an image-conditioned task synthesis framework that constructs dependencies from images to targets and from target-specific evidence to answers. Referential Abstraction (RA) replaces identity-revealing text with specific visual references or ambiguous relational references resolved by searching among depicted candidates. Compositional Evidence (CE) links retrieved facts about multiple entities in one image or entities from different images into joint questions whose answers require every designated contribution. Both operations also apply to images found during search, allowing new visual observations to determine subsequent search targets or supply answer evidence. Fine-tuning Qwen3.5-27B on only 12K teacher-generated trajectories from RACE raises average accuracy across ten multimodal benchmarks from 53.5 to 56.0, a relative improvement of approximately 4.7% over the matched base agent. Our code, data, and model will be publicly available.
Figures & tables
Figure 1: Making visually grounded search necessary. (a) Textual identity clues allow the target to be identified without the image. (b) The question uses the gun-holder’s evidence while leaving the knife-holder’s evidence unused. (c) RA resolves the relational reference and assigns the two candidates’ years to u and v ; CE combines both years to answer the directed comparison u−v . Arrows denote information dependencies.
Figure 2: RACE task synthesis and training. (a) Source annotations or visual analysis and retrieval establish target identities and provenance. (b) RA expresses targets through specific visual or ambiguous relational references. CE combines evidence from multiple targets within or across images. The example compares the years of the actors’ first credited feature-film appearances, yielding ∣tA−tB∣ . Evidence exploration and question-only search feedback refine the task. (c) The teacher receives (I,q) and uses tools to solve the task; trajectories with correct final answers are retained for SFT.
Multimodal search
Factual generalization
Model / setting
MM- BC
MS+
BC-V 3
VisB
BC-VL
VDR
MM Search
Simple VQA
Live VQA
FVQA
Avg.
Frontier API and large-backbone references
Qwen3.5-397B-A17B
46.0
42.1
31.0
52.7
49.4
—
78.3
76.4
85.2
77.1
—
Qwen3.7-Plus
37.1
46.3
42.3
59.8
55.6
30.2
80.7
77.7
84.6
79.7
59.4
Qwen3.8-Max
53.2
59.1
56.3
66.9
—
33.2
84.0
83.8
88.8
85.6
—
Gemini-3.1-Pro
48.7
51.5
50.0
62.1
65.9
40.8
86.1
81.0
87.3
84.0
65.7
Table 1: Search accuracy and factual generalization (%, ↑ ). Results cover four primary search benchmarks and six supplementary benchmarks. Models are grouped by role. The final block compares Qwen3.5-27B before and after SFT under matched settings. Avg. is the unweighted mean over all ten benchmarks.
Figure 4
Figure 4: Training trajectory distributions and behavior. (a) Visual calls per trajectory. (b) Proportion of trajectories containing a visual-web-visual pattern: a visual call, web search or page fetching, and another visual call in serialized order, with intervening calls allowed. (c) The proportion of trajectories with visual calls after the first tool round and mean number of distinct tool types. Colors are standardized within each metric, and the blue outline marks Ours. Abbreviations: OS-VL, OpenSearch-VL; VDR, Vision-DeepResearch; RS-MM, REDSearcher-MM.
Figure 5: Evaluation-time search behavior before and after SFT. Gray and blue bars give base-agent and SFT-agent point estimates, respectively. (a) Mean tool calls per trajectory. (b) Mean within-trajectory fraction of tool calls that use visual tools. (c) Fraction of trajectories that use visual tools after the first tool round. Panels (b) and (c) show percentages. Abbreviations: MM-BC, MM-BrowseComp; VisB, VisBrowse-Bench; BC-V3, BrowseComp-V 3 ; MS+, MMSearch-Plus.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Field
Specification
Teacher and student
Qwen3.7-Plus and Qwen3.5-27B
Main training data
12K synthesized trajectories only
Training stage and supervision
One SFT stage over assistant tool calls and responses
Trajectory-run decoding settings
temperature = 0.7 , top-p = 0.9
Trajectory-run iteration limits
MM-BC (100), MS+ (20), BC-V 3 (20), VisB (15)
Primary benchmark questions
MM-BC (224), MS+ (311), BC-V 3 (300), VisB (169), using full sets
Appendix
Table 3: Training setup and archived evaluation settings. Decoding settings and iteration limits refer to the archived runs used for behavioral analysis.
Table 8
Post-first-round visual use
Visual-search -fetch
Input images
Base
Trained
Base
Trained
1
22.5
34.2
42.3
83.3
2
27.8
51.9
44.4
98.1
3+
31.4
37.1
34.3
88.6
Appendix
Table 6: MMSearch-Plus search behavior (%). Base versus trained by input-image count.
Figure 6: Paired differences across task strata. SFT minus base-agent trajectory proportions, in points, for the question pairs in Tables 5 - 6 . Error bars are pointwise paired-bootstrap 95% intervals. The horizontal line marks zero.
Figure 7: Continued visual use across interaction lengths. Bars give the proportion of trajectories with a visual-tool call after the first round, within each round-count range. All panels share a scale. Blue bars denote Ours, and blue annotations give the difference from the highest baseline rate.
Figure 8: Supplementary visual-span statistics for training trajectories. (a) Proportion of trajectories whose normalized visual span exceeds each threshold x ; higher curves indicate more trajectories with widely separated visual calls. (b) Mean normalized visual span for each source. Both panels summarize trajectories with at least one visual call; a single visual call has span zero. Baseline names follow Table 1 .
Construction
Target configuration
Required behavior
Single-target abstraction
One target among the candidates in an image
Resolve a specific visual or ambiguous relational reference, then retrieve the requested information.
Within-image composition
Multiple required targets in one image
Retrieve evidence for each target and derive an answer requiring all designated contributions.
Cross-image composition
One required target per image across multiple images
Resolve each image’s target and combine its evidence with the other images’ contributions.
Appendix
Table 7: Task configurations. Referential Abstraction constructs target references, and Compositional Evidence adds joint information requirements. The same configurations apply to targets in retrieved images.
Multimodal large language models (MLLMs) have demonstrated strong capabilities in visual understanding, yet they remain limited in complex, multi-step reasoning that requires deep searching and integrating visual evidence with external knowledge. In this work, we address this challenge by constructing high-quality, verified multi-hop vision-language training data for multimodal deep-search agents. We propose a Multi-hop Tool-Augmented Agent for Evidence-based QA Synthesis (MTA-Agent), which automatically selects tools and their parameters to retrieve and validate evidence from both visual and textual sources and generates structured multi-hop question-answer trajectories. Starting from diverse VQA seed datasets, our pipeline produces a large-scale training dataset, MTA-Vision-DeepSearch, containing 21K high-quality multi-hop examples. The data is filtered through a multi-stage verification process to ensure factual consistency and answer uniqueness. Using MTA-Vision-DeepSearch, a 32B open-source multimodal search agent achieves state-of-the-art performance, reaching an average of 54.63% across six challenging benchmarks, outperforming GPT-5 (51.86%), Gemini-2.5-Pro (50.98%), and Gemini-3-Pro (54.46%) under the same tool settings. We further show that training on our data improves both reasoning depth and tool-use behavior, increasing the average number of steps from 2.27 to 4.28, and leading to more systematic and persistent search strategies. Additionally, we demonstrate that training can be performed without real-time tool calls by replaying cached interactions, significantly reducing training cost. Importantly, we present MTA-Agent as a fully open recipe for multimodal deep search: we release the entire dataset, training trajectories, and implementation details to enable reproducibility and future research on open multimodal search agents.
Training multimodal agents via reinforcement learning for knowledge-intensive visual reasoning is fundamentally hindered by the extreme sparsity of outcome-based supervision and the unpredictability of live web environments. To resolve these algorithmic and environmental bottlenecks, we introduce ProMMSearchAgent, establishing a novel Sim-to-Real training paradigm for multimodal search. We decouple policy learning into a deterministic, local static sandbox. Crucially, to learn effectively within this constrained environment, we propose an introspective process-oriented reward. By probing the agent's own parametric knowledge boundaries, we generate dense behavioral metadata that explicitly rewards the correct cognitive decision, initiating a multimodal or text search only when visually or factually uncertain. Extensive experiments demonstrate that our locally-trained policy transfers zero-shot to the live Google Search API. ProMMSearchAgent achieves new SOTA performance, outperforming MMSearch-R1 by +5.1% on FVQA-test, +6.3% on InfoSeek, and +11.3% on MMSearch.
Wentao Yan, Shengqin Wang, Huichi Zhou +4
East China Normal University · Huawei Noah’s Ark Lab · University College London +2
Multimodal large language models (MLLMs) have demonstrated impressive capabilities in many visual tasks, but they often struggle with factual grounding when confronted with complex, open-world scenarios. While recent multimodal deep search agents attempt to address this issue by utilizing external tools, the visual-native search paradigm remains underexplored. Existing methods primarily rely on simple images with explicit semantics and text-only evidence trajectories, limiting the agent's ability to perform multi-hop, cross-modal reasoning and search. To address these limitations, we propose Visual-Seeker, a visual-native multimodal deep search agent via active visual reasoning. Rather than treating vision as a static input, our agent actively attends to fine-grained visual details, dynamically harvests visual evidence throughout the search process. To unlock its visual-native potential, we design an active visual reasoning data pipeline and synthesize 5K high-quality multimodal trajectories for model training. Extensive experiments demonstrate the state-of-the-art performance across five challenging multimodal search benchmarks, even surpassing several proprietary models, validating robust visual-native reasoning and search in real-world web environments. The code and data can be accessed at: https://github.com/ZhengboZhang/Visual-Seeker.
Zhengbo Zhang, Changtao Miao, Jinbo Su +10
School of Artificial Intelligence, UCAS · Institute of Automation, CAS · Ant Digital Technologies, Ant Group +2