SimpleSearch-VL: A Simple Recipe for Multimodal Agentic Deep Search
SimpleSearch-VL employs FAR for efficient sampling, integrates evidence verification for reliability, and maintains lightweight tools, significantly boosting multimodal agentic search performance.
Key Findings
Methodology
This approach centers on Factorized Adaptive Rollout (FAR), which dynamically allocates sampling based on signal quality, combined with explicit chain-of-thought evidence verification. FAR expands prompt groups adaptively, prioritizing hard samples and avoiding tail delays. Evidence verification uses image reverse search to retrieve thumbnails, titles, and URLs, enabling the model to verify matches before utilizing evidence. The model performs webpage self-summaries internally, reducing external dependencies. Training on only 5K supervised trajectories and 2K RL data, it achieves significant improvements over baselines, with 15.8 and 16.0 point gains on 8B and 30B models respectively, outperforming many open-source counterparts.
Key Results
- On six benchmarks, SimpleSearch-VL-8B and 30B improve average scores by 15.8 and 16.0 points, surpassing most open-source models, demonstrating high data efficiency and robustness.
- The model exhibits strong evidence verification, especially in visual and web retrieval tasks, outperforming traditional methods in relevance and trustworthiness.
- Limited training data combined with FAR's dynamic sampling effectively enhances learning efficiency and tail latency mitigation.
Significance
This work addresses core challenges in multimodal search: efficiency bottlenecks, evidence reliability, and system complexity. FAR's signal-aware sampling reduces training and inference costs, while integrated evidence verification enhances trustworthiness. The lightweight design simplifies deployment, making the system suitable for real-world applications. Achieving state-of-the-art results on multiple benchmarks, it paves the way for more reliable and efficient multimodal AI systems, impacting fields like knowledge retrieval, visual reasoning, and human-computer interaction.
Technical Contribution
The paper introduces FAR, a novel signal-aware sampling strategy that adaptively prioritizes hard samples and reduces tail delays. It combines this with an explicit evidence verification mechanism based on image reverse search, enabling the model to validate retrieved evidence directly within the context. The webpage self-summary module further simplifies the system architecture, removing external summarization dependencies. These innovations collectively improve training efficiency, model reliability, and system practicality, representing a significant advancement over existing fixed-budget or heuristic sampling methods.
Novelty
This is the first integration of signal-aware, dynamic sampling (FAR) with explicit evidence verification in multimodal deep search agents. Unlike prior works relying on fixed sampling or implicit verification, FAR actively allocates resources based on reward signals, focusing on hard samples. The evidence verification via image reverse search ensures evidence authenticity, a feature rarely explored in multimodal settings. Additionally, the embedded webpage self-summary mechanism simplifies system design, making the approach more practical for deployment. These innovations collectively set a new standard for efficiency and reliability in multimodal agentic search.
Limitations
- While effective, the evidence verification relies on the accuracy of reverse image search, which may falter with ambiguous or complex visuals, affecting trustworthiness.
- Training with limited supervised trajectories may limit generalization to unseen or more complex tasks, requiring larger datasets for broader applicability.
- Tail latency, although mitigated, can still pose challenges in extremely long or multi-source retrieval scenarios, especially under resource constraints.
Future Work
Future directions include enhancing multi-source evidence fusion, integrating more sophisticated verification techniques, and scaling training data. Exploring unsupervised or semi-supervised methods could further improve generalization. Additionally, extending the framework to handle more complex multi-turn interactions and multi-modal reasoning tasks will broaden its applicability. The authors also plan to explore real-world deployment scenarios, optimizing for latency and robustness in dynamic environments.
AI Executive Summary
SimpleSearch-VL introduces a highly efficient and reliable framework for multimodal agentic search, driven by the innovative Factorized Adaptive Rollout (FAR). Traditional multimodal search systems often suffer from tail delays, high system complexity, and unreliable evidence verification. To address these issues, the authors propose FAR, which dynamically allocates sampling efforts based on reward signals, prioritizing difficult samples and reducing tail latency. This mechanism ensures that the model focuses on the most informative samples, significantly improving training efficiency.
Complementing FAR, the system employs explicit evidence verification through image reverse search, allowing the model to directly check whether retrieved visual evidence matches the query. This step enhances the trustworthiness of the retrieved information, a crucial aspect in knowledge-intensive tasks. Additionally, the model incorporates a self-summary mechanism for webpages, eliminating the need for external summarization tools and simplifying the overall architecture.
Experimental results demonstrate that, with only 5K supervised trajectories and 2K RL data, SimpleSearch-VL achieves remarkable improvements over baseline models, with 15.8 and 16.0 point gains on 8B and 30B variants, respectively. It outperforms most open-source multimodal search agents and reaches performance levels comparable to top-tier models like Gemini-3-Pro. The framework’s efficiency, combined with its verification reliability and practical design, makes it a promising approach for real-world applications in visual reasoning, web search, and knowledge retrieval.
Overall, this work advances the field by integrating signal-aware sampling, explicit evidence validation, and lightweight system design, setting new standards for multimodal agentic search. Future work will focus on scaling data, improving multi-source verification, and deploying in complex, real-time environments, further pushing the boundaries of intelligent multimodal systems.
Deep Dive
Abstract
We present SimpleSearch-VL, an efficient, reliable, and practical framework for multimodal agentic search. Its core idea is to improve the agent's own search-and-verification process rather than scaling data, tools, or auxiliary model components. For efficiency, Factorized Adaptive Rollout (FAR) improves sampling efficiency by forming more informative training groups while using redundant samples to mitigate long-tail latency and expose hard samples. For reliability, SimpleSearch-VL performs evidence-verified reasoning, explicitly using chain-of-thought verification to assess the relevance of retrieved visual and textual cues to the original context. For practicality, SimpleSearch-VL keeps a lightweight tool interface and performs webpage self-summary within the agent, requiring no additional external dependencies. With only 5K supervised tool-interleaved trajectories and 2K RL data, SimpleSearch-VL improves Qwen3-VL agentic baselines by 15.8 and 16.0 average points for the 8B and 30B-A3B variants, respectively. The SimpleSearch-VL-30B-A3B model further achieves performance competitive with agentic Gemini-3-Pro.