An item is worth one token in Multimodal Large Language Models-based Sequential Recommendation

Researchers propose 'Speeder,' a new AI model for sequential recommendations like predicting a user's next purchase. It solves key problems in current multimodal AI systems: inefficient data representation, bias toward text, and poor memory of long user histories. Speeder compresses item data, learns from images and text progressively, and better tracks sequence order. In tests, it tripled training speed and cut inference time to a quarter on an Amazon dataset.