HAA-Fin: A Hierarchical Agentic Architecture for Efficient Multimodal Financial Regulation Parsing: HAA-Fin

The deployment of large language models (LLMs) in financial verticals is frequently bottlenecked by computational constraints and the strict imperative for zero-defect accuracy. This paper challenges the prevailing scaling law paradigm by introducing HAA-Fin, a resource-efficient and high-precision framework for regulatory compliance. We construct a Hierarchical Agentic Architecture (HAA) that decomposes complex multimodal auditing tasks, such as detecting layout-dependent loan rate violations or identifying illegal advertising claims, into atomic sub-tasks orchestrated by specialized agents including TimeAgent and RuleAgent. By synergizing a retrieval-augmented generation (RAG) module with layout-aware hybrid search and re-ranking, we enable 8B-parameter models to rival or even surpass the performance of models with over 32B parameters. Validated on a challenging national-level financial benchmark, our method effectively handles multimodal inputs of text and visual layout. It achieves a significant 20.7% accuracy boost via re-ranking and reduces single-document review time to under 1.5 seconds, demonstrating that orchestrating specialized agents is a superior pathway to scalable industrial AI compared to merely increasing model size.