Biomedical Event Extraction (BEE) aims to extract structured information from biomedical texts to support downstream applications. While Large Language Models (LLMs) have shown promise in this task, current methods suffer from hallucinations due to the generative nature and the limited size of BEE datasets for effective instruction tuning. To this end, we propose Dual-path Biomedical Event Extraction (DBEE), a novel framework that mitigates these issues via dual-path consistent extraction, and discrepancy retention. Specifically, DBEE consists of two components: (1) a Chain-of-Thought (CoT) enriched instruction design and data augmentation that reforms the event extraction process into two paths with cross-verification, and (2) discrepancy retention module to recall plausible events from inconsistent predictions. Extensive experiments on Chinese and English BEE datasets show that DBEE achieves state-of-the-art performance. The code is open-sourced and available at https://github.com/zengjianjun-ecust/DBEE .