Rosinality

Attention-FFN disaggregation. This was explored by ByteDance previously (arxiv.org/abs/2504.02263), but StepFun vastly improved latency.

창작물의 내용이나 작가의 윤리성에 집착하는 것도 특이한 현상. 한국이 유독 심한 것도 맞는 듯?

arxiv.org/abs/2507.16812 Extracting QA data from textbooks.

arxiv.org/abs/2507.16577 Expand the state size by N times, then select the top-K among them and update using softmax weights.