This study examines the application of Reinforcement Learning (RL) in conjunction with Large Language Models (LLMs) for the multi-label classification of financial news. The research evaluates models of different scales, including GPT-4, Llama3 8B, and Llama3 70B, across two scenarios:(1) classification using only the query and corpus without RL, and (2) classification incorporating RL. A dataset consisting of 111,000 financial news articles was constructed through web scraping, with a sample of 1,000 articles manually labeled into three impact categories: weak, semi-strong, and strong. The experimental results demonstrate that smaller models, such as Llama3 8B, exhibit improved accuracy and coverage when RL is applied. However, larger models like GPT-4 displayed reduced performance under RL, suggesting challenges in adapting larger architectures for this task.