Researchers weigh token-drop trick: variable-length sequences make it far harder

TimDarcet · x · 2026-09-09

Replying to a question about whether a tokendrop-style permute+slice trick could fully skip a layer, TimDarcet concedes it doesn't work with packed sequences: it would require variable-shape kernels and hit heterogeneous distributed compute — still interesting, but clearly harder than the fixed-seqlen case.

Related event: Meta researcher shares batch-ratio token dropping training trick(3 posts)→

Original post →

More from Research

Research channel →