Reproduce “FinixDoc: Rethinking Financial Document Parsing Beyond Saturated Benchmarks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hang Wang, Jin Zhang, Guoliang Xu, Pengyue Lu, Yao Li, Zijiao Zhang, Tianyu Huang, Weiqi Xiong and 8 more
Ant Group's FinixDoc is an end-to-end agentic parsing system featuring FinixDoc-VL, a specialized vision-language model, designed for real-world financial documents. The system and its model enhance robustness and accuracy by addressing challenges in low-quality visual inputs and large document scales, outperforming existing solutions on a new financial-domain benchmark, FinixDocBench.
Key insights
- Real-world financial document parsing necessitates evaluation beyond general benchmarks, with a focus on visual quality degradation and document scale, as defined by the new Document Parsing Capability Matrix.
- Domain-specific visual adaptation (homoglyph-aware contrastive learning) and business-aligned reinforcement learning with composite rewards are effective for enhancing VLM performance in challenging financial scenarios.
- Complex industrial document parsing problems, such as handling ultra-large pages, require system-level solutions that integrate core models with intelligent preprocessing and post-processing tools.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.