AI Paper Atlas

Research map / Architectures and efficiency / Mixture of experts

Prism-MoE: Efficient Dense-to-MoE Conversion for Visual Autoregressive Generation

Ying Li, Zefang Wang, Zhaode Wang, Zhiwen Chen, chengfei lv, Huan Wang

ICML 2026 · Architectures and efficiency › Mixture of experts · approach: Expert routing and specialization

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 trajectory-consistent initialization (ablation)Taylor importance pruning init · VAR-d16 ImageNet, 75% active, no FTFID: 27.5 -> 12.6Improved
add distillation fine-tuning at 37.5% active (ablation)trajectory-consistent init · VAR-d16 ImageNetFID: 12.6 -> 9.00Improved
add normalized cosine router (ablation)distillation fine-tuned MoE · VAR-d16 ImageNet, 37.5% activeFID: 9.00 -> 6.04Improved
add spatially-aware routing loss (ablation)cosine router MoE · VAR-d16 ImageNet, 37.5% activeFID: 6.04 -> 4.26Improved
fine-tune longer with standard load-balancing loss (epoch 5 to 20) (ablation)VAR-d20 MoE, 37.5% active · ImageNet-1K C2IFID: 7.05 -> 10.7Hurt
replace linear softmax router + LB loss with Prism-MoE fine-tuningVAR-d20 MoE, 37.5% active · ImageNet-1K C2IFID: 10.7 -> 2.86Improved
replace linear softmax router + LB loss with Prism-MoE fine-tuningLlamaGen-XL MoE, 37.5% active · MS-COCO T2IFID: 31.5 -> 26.0Improved
increase calibration set size from 100 to 200 (ablation)Prism-MoE init · VAR-d20 FIDFID: 6.04 -> 5.03Improved
increase calibration set size from 200 to 300 (ablation)Prism-MoE init · VAR-d20 FIDFID: 5.03 -> 4.96No clear effect
increase ridge multiplier from 1.0 to 2.0 (ablation)Prism-MoE init · VAR-d20 FIDFID: 5.03 -> 5.28Hurt
use Prism-MoE init at 75% active instead of denseSemanticist · SemanticistFID: 2.49 -> 2.99Hurt
use Prism-MoE (37.5% active) instead of denseVAR-d20 · Inference latencylatency: 45.5 ms -> 37.2 msImproved

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