SimpleSearch-VL: A Simple Recipe for Multimodal Agentic Deep Search

2026-06-30

AI Summary

This work addresses the reliance of multimodal agents on large-scale data and complex external tools in deep search scenarios by proposing a lightweight, end-to-end solution. It introduces Factorized Adaptive Rollout (FAR) to enhance sampling efficiency and incorporates an evidence-verified chain-of-thought mechanism to improve reasoning reliability. The agent autonomously generates webpage summaries internally, eliminating external dependencies. By integrating tool-interleaved trajectory supervision with reinforcement learning, the method achieves significant performance gains using only 5K supervised and 2K reinforcement samples. On Qwen3-VL 8B and 30B-A3B baselines, it improves average scores by 15.8 and 16.0 points, respectively, with the latter matching the performance of Gemini-3-Pro.

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.

Problem

• multimodal agentic search
• search efficiency
• evidence-verified reasoning
• tool-interleaved trajectories
• deep search

Innovation

• multimodal agentic search
• Factorized Adaptive Rollout
• evidence-verified reasoning
• lightweight tool interface
• chain-of-thought verification