Blog
Research write-ups. All blog posts, including industry blogs, are also listed under Publications on the main page.
Research Blog
In LLMs, one token is both the attention sink and the massive activation. Across SD3.5, Qwen-Image,
Mochi and HunyuanVideo, both move into the text stream: the strongest sinks are the highest-norm text
tokens, while image tokens carry a single shared hot channel written by the adaLN gate.