Notes and summaries of research papers.

• Attention Residuals — Kimi Team technique replacing fixed residual weights with learned softmax attention over layer outputs.