Ant Group researchers propose the User Behavioral Densing Law, which quantifies how tokenization capacity should scale with data size for user representation learning. Through experiments on billion-scale Alipay data, they found an approximately linear relationship between the logarithms of minimum sufficient tokenization capacity and input data size. The team developed ALGN, an adaptive variable-length tokenization method that improves capacity allocation and outperforms existing approaches across diverse datasets and downstream tasks.
