AI Paper Atlas

Research map / Architectures and efficiency / Mixture of experts

Preventing Expert Collapse in MoE-dVLMs via Modality-Wise Norm Alignment

Zongkai Liu, Zhen Cao, Hui Zhang, Chao Yu, liqiang niu, Fandong Meng

ECCV 2026 · Architectures and efficiency › Mixture of experts · approach: Expert routing and specialization · 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 TowerAlign post-projector norm alignment (ablation)MoE-dVLM · ChartQAaccuracy: 4.7 -> 59.0Improved
add TowerAlign post-projector norm alignment (ablation)MoE-dVLM · DocVQA (val)accuracy: 13.8 -> 67.7Improved
add TowerAlign post-projector norm alignment (ablation)MoE-dVLM · AI2Daccuracy: 58.7 -> 70.5Improved
add TowerAlign post-projector norm alignment (ablation)MoE-dVLM · MMMU-Pro (vision)accuracy: 14.2 -> 15.1Improved
use global affine compression instead of token-wise RMSNorm (ablation)TowerAlign · ChartQAaccuracy: 19.4 -> 59.0Improved
use global affine compression instead of token-wise RMSNorm (ablation)TowerAlign · MMMU-Pro (vision)accuracy: 16.1 -> 15.1Hurt
raise load-balancing loss coefficient to 10x (0.1) (ablation)MoE-dVLM · DocVQA (val)accuracy: 13.8 -> 9.3Hurt
add Gumbel-Softmax noise to routing (ablation)MoE-dVLM · DocVQA (val)accuracy: 13.8 -> 6.8Hurt
add Gumbel-Softmax noise to routing (ablation)MoE-dVLM · AI2Daccuracy: 58.7 -> 28.9Hurt
train full multi-turn dialogue as one sequence instead of per-turn SFT (ablation)TowerAlign · ChartQAaccuracy: 43.8 -> 59.0Improved
train full multi-turn dialogue as one sequence instead of per-turn SFT (ablation)TowerAlign · DocVQA (val)accuracy: 55.3 -> 67.7Improved
use sparse MoE TowerAlign instead of dense LLaDA-VLM at matched dataLLaDA-VLM · DocVQA (val)accuracy: 61.2 -> 67.7Improved

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