Yuchao Zheng

MLLMs can achieve state-of-the-art multimodal retrieval by learning to compress information into a handful of "bottleneck" tokens, forcing the model to distill relevant semantics.

By unifying feature interaction and sequence modeling in a single Transformer backbone, MixFormer eliminates the co-scaling trade-off that plagues fragmented recommender systems and boosts user engagement in real-world deployments.