betaFeedback

WorksComputer Science

On-Policy Morphing into Hybrid Attention Models

Abstract

Hybrid attention models combine full and linear attention to achieve a better quality–efficiency trade-off. Converting pretrained Transformers into such models requires identifying which layers should retain full attention under a prescribed budget. We identify two off-policy mismatches in FlashMorph's layer selection: (i) continuous branch interpolation and teacher-forced hidden states differ from discrete hybrid computation; and (ii) frozen model weights fail to account for parameter adaptation during subsequent recovery. We introduce FlashMorph2, an on-policy layer-selection method that (I) evaluates sampled discrete hybrid architectures end to end using their own hidden states, and (II) jointly optimizes layer-selection scores and model weights under the recovery objective to reflect post-adaptation performance. Experiments across Qwen3 model scales and linear-attention variants show that FlashMorph2 consistently outperforms existing layer-selection methods, improving average scores over FlashMorph by +2.42+2.42 points on RULER, +1.08+1.08 on LongBench, and +0.76+0.76 on recall-intensive tasks, while completing layer selection in only 0.720.72–3.573.57 GPU hours across 0.6B–8B models, demonstrating high efficiency and scalability.

Paper details

vlnc:W_x6k5szta
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…