TY  - RPRT
TI  - Rethinking Length-Based Training: Batch Composition and Loss Normalization in Speech Token Language Models
AU  - Hongjin Song
AU  - Runwu Shi
AU  - Weiqiao Shan
AU  - Jiale Luo
AU  - Yujin Wang
AU  - Yifei Wu
AU  - Chunxiang Jin
PY  - 2026
UR  - https://arxiv.org/abs/2609.25890
ID  - 2609.25890
ER  -