Document-level event extraction (DocEE) is a challenging task that requires identifying structured event arguments spread across an entire document. While Large Language Models (LLMs) possess strong language comprehension abilities, they often lack inherent knowledge of the specific schemas of different event types. A common approach to align LLM outputs is to concatenate predefined event schemas and descriptions with the input document, a strategy we term a “passive” structure-learning approach. This method, however, can lead to excessively long prompts, increasing computational costs and the risk of model hallucination. In this paper, we propose an Active Structure-Learning Strategy (ASLS), where the LLM is directly fine-tuned on structured JSON-based outputs. This enables the model to implicitly “understand” the event structure without requiring explicit schema descriptions in the input. Our experiments achieve state-of-the-art performance on two evaluation datasets and in the online testing.