Emphasis-Based Extraction of Patent Differences Using Small Language Model

This study investigates the extraction of differences unique to the application invention using small LLMs under resource-constrained environments. Given technically similar invention pairs annotated with emphasis markers(**), the task is explicitly defined as selectively identifying only those differences that are emphasized and unique to the Application Invention. We compare Baseline, Prompt Engineering, and Fine-tuning approaches using Qwen2.5-14B. Experimental results show that Prompt Engineering and Fine-tuning outperform the Baseline, with the fine-tuned small model slightly exceeding the baseline performance of GPT-OSS-120B, indicating that task-specific adaptation can offset model scale limitations.