Ant Group proposes a Densing Law for billion-scale user representation learning

antgroup · hf · 2026-08-25

This study proposes a scaling law linking data size to tokenization capacity for user behavior modeling at billion-scale capacity. Based on this, it introduces an adaptive tokenization method to improve the efficiency of user representation learning.

Related event: Ant Group proposes densification law for billion-scale user representation(2 posts)→

Original post →

More from Research

Research channel →