AI Paper Atlas

Research map / Architectures and efficiency / Mixture of experts

Attention Sink Forges Native MoE in Attention Layers: Sink-Aware Training to Address Head Collapse

Zizhuo Fu, Wenxuan Zeng, Runsheng Wang, Meng Li

ICML 2026 · Architectures and efficiency › Mixture of experts · approach: Efficient MoE language model systems · 0 citations (OpenAlex)

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
add auxiliary head load balancing loss (ablation)Vanilla attention 0.6B · from scratch 11-benchmark averageaccuracy: 39.54 -> 40.82Improved
add auxiliary head load balancing loss (ablation)Gated attention 2B · from scratch 11-benchmark averageaccuracy: 46.30 -> 47.70Improved
add auxiliary head load balancing loss (ablation)Sink attention 2B · from scratch 11-benchmark averageaccuracy: 46.35 -> 46.50Improved
add auxiliary head load balancing loss (ablation)Vanilla attention 2B · FineWeb-Edu validationBPB: 0.7243 -> 0.7198Improved
use Gated attention instead of Vanilla attentionbase loss, 2B · from scratch 11-benchmark averageaccuracy: 45.46 -> 46.30Improved
use Sink attention instead of Vanilla attentionbase loss, 0.6B · from scratch 11-benchmark averageaccuracy: 39.54 -> 40.31Improved
add auxiliary load balancing loss in fine-tuning (ablation)Qwen3-8B · reasoning averageaccuracy: 74.31 -> 76.01Improved
add auxiliary load balancing loss in fine-tuning (ablation)LLaMA3.1-8B · reasoning average ProcessBenchaccuracy: 23.63 -> 31.14Improved
add auxiliary load balancing loss in fine-tuning (ablation)Qwen3-8B · LongBench averageaccuracy: 50.67 -> 55.38Improved
add auxiliary load balancing loss in fine-tuning (ablation)Qwen3-4B · LongBench HotpotQAaccuracy: 60.81 -> 58.61Hurt
add auxiliary load balancing loss in fine-tuning (ablation)LLaMA3.1-8B · LongBench TriviaQAaccuracy: 46.34 -> 41.82Hurt
zero first-token value in all heads (ablation)Qwen3-8B · 9-benchmark averageaccuracy: 70.36 -> 64.10Hurt

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