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

2026-08-03

AI Summary

This work addresses the limitation of existing multimodal deep search methods, which largely neglect the role of visual information during intermediate reasoning stages and thus struggle to support long-horizon, multi-turn vision-driven retrieval. To overcome this, the paper introduces a novel agent architecture that deeply integrates visual cues into extended reasoning processes. It constructs multimodal event graphs to synthesize queries requiring intermediate visual dependencies and long reasoning chains, and incorporates mechanisms for active visual acquisition and on-demand image loading. Notably, the proposed approach achieves significant performance gains over current state-of-the-art methods across ten established multimodal search benchmarks through supervised fine-tuning alone—without reliance on reinforcement learning—demonstrating the efficacy of vision-in-the-loop retrieval for complex reasoning tasks.

Abstract

Multimodal large language models (MLLMs) have advanced visual understanding and reasoning, yet their static parametric knowledge limits their ability to address knowledge-intensive and dynamically evolving open-world problems. To move beyond this limitation, multimodal deep search has emerged as a key direction for open-world information access, evolving from single-turn factual retrieval toward long-horizon, multi-turn search guided by visual evidence. However, existing methods typically confine vision to the input or answer stage, overlooking its role in intermediate reasoning, and lack designs tailored to long-horizon interaction. Consequently, visual evidence rarely drives continued retrieval, constraining both interaction depth and reasoning span. To address these limitations, we propose DeepVoyager-VL, a long-horizon multimodal deep-search framework for vision-in-the-loop search. Specifically, we construct a multimodal event graph to drive data synthesis, yielding problems with intermediate visual dependencies and long reasoning chains. We then design an agent framework for active visual acquisition and on-demand image loading. Finally, we fine-tune models on the synthesized data without reinforcement learning. Extensive experiments across ten multimodal search benchmarks demonstrate the effectiveness of our method.

Problem

• multimodal deep search
• vision-in-the-loop
• long-horizon interaction
• visual evidence
• intermediate reasoning

Innovation

• vision-in-the-loop
• multimodal deep search
• long-horizon reasoning
• multimodal event graph
• active visual acquisition