HierDoc framework enhances long-document visual question answering

Researchers have developed HierDoc, a novel framework for long-document visual question answering that improves evidence retrieval by employing a hierarchical approach. This method first selects relevant pages from a document and then identifies specific regions within those pages for detailed analysis. This two-stage process, optimized using GRPO and structured-set rewards, allows for more precise evidence acquisition compared to previous methods that treated page and region selection separately. HierDoc has demonstrated state-of-the-art performance on relevant benchmarks, outperforming existing open-weight systems and showing significant accuracy improvements. AI

IMPACT Improves accuracy and efficiency in processing and querying long documents, potentially benefiting AI applications that rely on detailed document comprehension.

RANK_REASON Research paper detailing a new method for document analysis. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• alphaXiv
• arXiv
• CatalyzeX
• Connected Papers
• DagsHub
• Gotit.pub
• HierDoc
• Hugging Face
• Litmaps
• LongDocURL
• ScienceCast

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →