data-ordering

This paper systematically analyzes length-based training in speech token language models, showing that short-to-long ordering has no independent benefit when batch composition and token exposure are fixed, and evaluates the impact of loss normalization.