Researchers have introduced Doc-V*, an OCR-free agentic framework designed for multi-page Document Visual Question Answering. This system tackles the challenge of processing long, visually dense documents by employing a coarse-to-fine interactive approach. Doc-V* begins with a broad overview and then uses semantic retrieval and targeted page fetching to gather evidence, which is then aggregated in a structured working memory for reasoning. The framework has demonstrated improved out-of-domain performance by up to 47.9% compared to retrieval-augmented generation baselines across various benchmarks. AI
IMPACT This framework could improve how AI systems process and reason over lengthy documents, impacting fields like legal tech, research analysis, and document management.
RANK_REASON The cluster describes a new research paper detailing a novel framework for a specific AI task. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- DagsHub
- Doc-V*
- Group Relative Policy Optimization
- Hugging Face
- optical character recognition
- retrieval-augmented generation
- visual question answering
- Yuanlei Zheng
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
