News

We release two works investigating how training data for LLMs (slow vs. fast thinking, instruction vs. reasoning) lead to different layer-wise gradient patterns and training dynamics, from which we unified various existing data quality metrics into one: small but diverse gradient directions indicate better data.