Improving Financial Document Parsing Accuracy

Summary & Methodology Analysis

FinixDoc uses the Qwen3-VL-4B architecture, a vision-language transformer that processes visual and textual inputs through multimodal fusion and autoregressive decoding, as its foundation. The authors extended this model using a custom evaluation suite called FinixDocBench, which tests performance across digital, camera-captured, and high-complexity workflows. Empirical results show that FinixDoc-VL provides substantial improvements over the base model on the OmniDocBench dataset, proving effective for specialized financial tasks where standard models often fail. The system is specifically optimized for internal financial workflows, where it achieved a score of 84.08 compared to 78.73 for the next best alternative. To handle ultra-large pages, the system employs a split-then-merge strategy that breaks down massive document segments for individual parsing before reintegrating them, significantly boosting parsing success for extreme document scales. Despite these gains, the current iteration lacks an automated mechanism to refuse uncertain queries, relying instead on a design principle that prioritizes safe omission over potentially incorrect, ungrounded outputs in ambiguous scenarios. One technical bottleneck is that the split-then-merge approach can cause complex tables to exceed the TEDS evaluator 6-minute timeout threshold, resulting in a TEDS score of 0 for those cases. Approximately 25% of the model's table cases encounter this timeout during evaluation.