VISOR: Agentic Visual Retrieval-Augmented Generation via Iterative Search and Over-horizon Reasoning

Researchers introduce VISOR, a new agentic visual retrieval-augmented generation system that improves how AI models reason over multi-page visual documents. By addressing key technical challenges in evidence gathering and context management, VISOR achieves state-of-the-art results on complex visual reasoning tasks.

VISOR represents a meaningful advancement in how vision-language models handle complex visual reasoning tasks that require processing information across multiple pages. The research addresses two fundamental technical problems that have limited previous agentic systems: the difficulty of reasoning across scattered visual evidence and the degradation of performance when processing increasingly large amounts of visual data.

The problem context stems from the rapid evolution of retrieval-augmented generation (RAG) systems. As these models have become more capable, researchers have extended them to handle visually rich documents—a significantly harder task than text-only retrieval. Traditional approaches process each page independently, missing connections between related evidence, while longer search horizons accumulate visual tokens that overwhelm the model's reasoning capacity.

VISOR's technical innovations directly address these pain points through three mechanisms: a structured evidence space for cross-page reasoning, visual action evaluation to ensure retrieval quality, and a dynamic trajectory system with sliding windows to maintain search focus. The use of GRPO-based reinforcement learning with specialized state masking demonstrates sophisticated training methodology.

This advancement matters for the broader AI field because visual document understanding represents a growing practical need. Enterprise applications increasingly require reasoning over documents like financial reports, technical manuals, and research papers. Better performance on benchmarks like ViDoSeek and SlideVQA signals progress toward production-ready systems. The efficiency gains mentioned suggest VISOR could reduce computational costs while improving accuracy, making such systems more deployable in real-world settings.

• →VISOR introduces a structured evidence space to enable cross-page visual reasoning where previous systems processed pages independently.
• →A visual action evaluation and correction mechanism improves retrieval precision and prevents degradation from misused visual actions.
• →Dynamic trajectory with sliding window technology mitigates search drift in long-horizon reasoning tasks.
• →GRPO-based reinforcement learning with state masking enables efficient training for dynamic context reconstruction.
• →Achieves state-of-the-art results on ViDoSeek, SlideVQA, and MMLongBench benchmarks with improved computational efficiency.