Agentic data science with the Data Agent Kit
cloud.google.com·8h ago
AI Research
arXiv CS.AI·Aug 17·5 min read
VISOR: Agentic Visual Retrieval-Augmented Generation via Iterative Search and Over-horizon Reasoning
TL;DR
VISOR is a new framework for Visual Retrieval-Augmented Generation that enhances multi-step reasoning over visual documents. It addresses challenges like visual evidence sparsity and search drift, achieving state-of-the-art performance in long-horizon visual reasoning tasks.
✦ Why It Matters
Explore VISOR's capabilities to enhance your visual reasoning applications today.
Key Takeaways
- Implement VISOR for improved visual reasoning in your projects.
- Utilize the structured Evidence Space for better cross-page analysis.
- Adopt the Dynamic Trajectory to prevent search drift.
- Leverage GRPO-based RL for dynamic context reconstruction.
How It Works
VISOR integrates a structured Evidence Space that allows for progressive reasoning across multiple pages, addressing the issue of visual evidence sparsity. It also incorporates a Visual Action Evaluation and Correction mechanism to ensure that visual actions are effectively managed, enhancing retrieval quality.
The Dynamic Trajectory with Sliding Window and Intent Injection helps maintain focus during long searches by anchoring the evidence space and filtering out irrelevant interactions.