Published Paper Details:

This study explores Intelligent Invoice Information Extraction in the context of recent progress in Vision-Language Models (VLMs). Conventional OCRbased pipelines frequently encounter recognition errors, domain-specific limitations, weak generalization, and poor performance when processing invoices with varied layouts. To address these challenges, we present a Hybrid Vision- Language Framework that combines the OCR-free Donut document transformer with the Gemini multimodal large language model. The framework enables structured extraction of key financial fields from invoices of multiple templates. Donut performs the visual encoding and sequence generation without relying on OCR, whereas Gemini provides higher-level reasoning, validation, and semantic refinement of the extracted information. The objective is to achieve high-precision identification of invoice numbers, dates, vendor details, tax components, itemized records, and total amounts. A detailed review of the literature indicates that only a few existing systems utilize hybrid VLM architectures that fuse OCR-free models with multimodal reasoning models for invoice extraction. Extensive empirical evaluations on custom datasets and standard benchmarks demonstrate substantial performance gains over traditional OCR-based and transformer-based baselines. The key contributions of this work include a scalable system architecture, an analysis of hybrid reasoning effectiveness, comprehensive experimental results, and practical insights for organizations aiming to automate financial processing workflows. The findings suggest that hybrid VLM frameworks offer a significant advancement for Intelligent Document Processing (IDP), reducing manual effort while improving generalization to previously unseen invoice formats.