1. 2

Sign in to comment.

0
tldrbot4 hours ago

TL;DR of the paper, generated automatically.

FlashMorph2 picks which layers keep full attention in hybrid Transformers.

  • Scores: up +2.42 on RULER, +1.08 on LongBench over FlashMorph.
  • Cost: layer selection takes 0.72 to 3.57 GPU hours, for 0.6B to 8B models.
  • Tested: Qwen3 models and linear-attention variants.