Implicit attribute value extraction remains an underexplored challenge in multimodal learning, particularly in the context of Chinese e-commerce. To address this gap, we introduce ChiImpAVE, the first open-source Chinese multimodal dataset designed for implicit attribute value extraction. Additionally, we propose ImpliLLM, a novel multimodal large language model optimized for implicit attribute extraction. ImpliLLM leverages contrastive learning to enhance the semantic alignment between image and text modalities, while incorporating reinforcement learning to refine the model’s decision-making process. These strategies effectively address challenges such as ambiguous attribute boundaries and hallucinated outputs. Comprehensive experiments on ChiImpAVE and ImplicitAVE datasets show that ImpliLLM outperforms state-of-the-art models, achieving a score of 68.79 on ChiImpAVE and 86.89 on ImplicitAVE. The contributions of this work include the introduction of the ChiImpAVE dataset, which enables implicit attribute value extraction in Chinese e-commerce, and the development of the ImpliLLM model, which effectively addresses challenges such as ambiguous attribute boundaries and hallucinated outputs, advancing the field of multimodal attribute extraction. The dataset are available in https://huggingface.co/datasets/chenqi-205/ChilmpAVE .