Extracting information from invoices proves to be a difficult task. The problem exists because it serves as a fundamental element for document comprehension. The economic system together with healthcare services and governmental operations receive influence from it. The main solution for this task now consists of pre-trained language models (PLMS) which include large language models (LLMS) that operate on deep learning systems. The wide range of available methods produces a difficult environment which researchers and practitioners must navigate. The research analyzes eight fine-tuning methods through their architectural design and task formulation and computational efficiency to create a complete comparison. The research traces the development of token classification models from BERT which used OCR output to modern multimodal learning systems including LayoutLMv2 and Donut and instruction-tuned LLMs. The finding demonstrate that multimodal architectures achieve higher performance than text-only models when working with intricate layout structures. The implementation of parameter-efficient fine-tuning (PEFT) methods enables performance maintenance while reducing computational expenses. The selection of extraction strategies will be guided by our structured decision framework which bases its choices on operational constraints and performance requirements