AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Distilling to Hybrid Attention Models via KL-Guided Layer Selection

Yanhong Li, Songlin Yang, Shawn Tan, Mayank Mishra, Rameswar Panda, Jiawei Zhou, Yoon Kim

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Long-context KV and memory attention

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
select softmax layers by KL-guided greedy addition instead of uniform interleavinghybrid GDN distillation · RULER, Qwen2.5-3B 25% softmaxRULER: 0.6904 -> 0.8631Improved
select layers by KL-guided greedy addition instead of Uniformhybrid GDN distillation · RULER, Llama-3.2-3B 25% softmaxRULER: 0.461 -> 0.7539Improved
use KL-guided selection instead of the strongest baseline (AR) at 12.5% softmaxhybrid GDN distillation · RULER, Qwen2.5-3B 12.5% softmaxRULER: 0.542 -> 0.662Improved
use KL-guided selection instead of SMART at larger scaleshybrid GDN distillation · RULER, Qwen2.5-1.5B and 7BRULER: 1.5B 25%: 0.5098 -> 0.5408; 7B 25%: 0.8158 -> 0.8584Improved
use stage-2 KL importance instead of stage-1 MSE importance (ablation)GA selection · RULER, Qwen2.5-3B 25% softmaxRULER: 0.5408 -> 0.8631Improved
use greedy addition instead of greedy removal (ablation)KL-based selection · RULER, Llama-3.2-3B 25% softmaxRULER: 0.4950 -> 0.7539Improved
use greedy addition instead of greedy removal (ablation)KL-based selection · RULER, Qwen2.5-3B 25% softmaxRULER: 0.8259 -> 0.8631Improved
average GA and GR rankings instead of GA only (ablation)KL-based selection · RULER, Llama-3.2-3B 25% softmaxRULER: 0.7539 -> 0.5580Hurt
use GLA-specific selection instead of GDN-selected layers (ablation)GLA student · RULER, Qwen2.5-3B 25% softmaxRULER: 0.8407 -> 0.6921Hurt
add more softmax layers (K from 1 to L-1) (ablation)hybrid distillation · RULER vs commonsense averageRULER / commonsenseMixed
stop selection pass early (25-40% of stage-2 training) (ablation)layer selection · RULERRULER: difference under 0.01No clear effect
use a larger sliding window in an SWA studentsliding-window distillation · RULER vs commonsenseRULERMixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy