1 paper across 1 session
We enable dynamic inference in next-scale autoregressive generation by restructuring intermediate representations with frequency-aware supervision.
