Qifu Technology's Paper Accepted by IJCAI 2025, Using MLLM to Pave New Path in Fintech

With an acceptance rate of merely 19.3% — selecting 1,042 out of 5,404 submissions — IJCAI is recognized as one of the most prestigious AI conferences globally and is listed as a Class-A event by the China Computer Federation （CCF）.

The research focuses on Compositional Zero-Shot Learning, aiming to identify novel combinations of attributes and objects using existing knowledge. Overcoming traditional challenges such as background interference, limited semantic capture in word embeddings, and overconfidence in known combinations, the team introduced an innovative framework named TRIDENT. By integrating Multimodal Large Language Model (MLLM) embeddings and attribute smoothing, TRIDENT utilizes modules like Feature Adaptive Aggregation (FAA) to mitigate background noise, learns conditional masks for detailed feature extraction, and leverages MLLM's hidden states to enhance semantic representation. This approach has demonstrated state-of-the-art performance across multiple datasets, offering new solutions for fields including image recognition and content understanding.