Research and Application of Multimodal LLMs DeepSeek in Power Enterprise Electronicization of Vouchers

This paper discusses the application and research of multimodal LLMs in the field of power enterprise credential electronicization. With the continuous development of artificial intelligence technology, multimodal LLMs show strong generalization ability in handling complex and diverse tasks such as information comprehension, content generation, and multi-round dialogues. In January 2025, DeepSeek quietly launched its model R1, which ignited the entire AI circle with its powerful reasoning and thinking abilities. We choose the open-source LLM DeepSeek R1 7B/8B/14B, Qwen2.5-7B/14B, or GLM-4-9B models. Voucher e-documentation is an important part of power enterprise financial management and tax declaration, and its processing efficiency and accuracy are crucial to the operation of enterprises. Traditional voucher e-enabling methods, such as OCR technology, have improved recognition efficiency to some extent, but still have many limitations. This paper firstly introduces the background, current situation and limitations of voucher electronic, then elaborates the basic principles of multimodal LLMs and its application in voucher electronic, analyzes the innovation points of using multimodal LLMs to recognize financial vouchers, and finally summarizes the research results and looks forward to the future development direction. The project team of the thesis implemented the establishment of 36 metadata models for a certain electricity enterprise, covering documents, vouchers, account books, financial statements, bank statements, invoices, etc. The average daily business processing volume of the personnel in the financial shared service center has increased from 70 items per person per day to 130 items per person per day, achieving a substantial improvement in work efficiency.