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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add TowerAlign post-projector norm alignment (ablation) | MoE-dVLM · ChartQA | accuracy: 4.7 -> 59.0 | Improved |
| add TowerAlign post-projector norm alignment (ablation) | MoE-dVLM · DocVQA (val) | accuracy: 13.8 -> 67.7 | Improved |
| add TowerAlign post-projector norm alignment (ablation) | MoE-dVLM · AI2D | accuracy: 58.7 -> 70.5 | Improved |
| add TowerAlign post-projector norm alignment (ablation) | MoE-dVLM · MMMU-Pro (vision) | accuracy: 14.2 -> 15.1 | Improved |
| use global affine compression instead of token-wise RMSNorm (ablation) | TowerAlign · ChartQA | accuracy: 19.4 -> 59.0 | Improved |
| use global affine compression instead of token-wise RMSNorm (ablation) | TowerAlign · MMMU-Pro (vision) | accuracy: 16.1 -> 15.1 | Hurt |
| raise load-balancing loss coefficient to 10x (0.1) (ablation) | MoE-dVLM · DocVQA (val) | accuracy: 13.8 -> 9.3 | Hurt |
| add Gumbel-Softmax noise to routing (ablation) | MoE-dVLM · DocVQA (val) | accuracy: 13.8 -> 6.8 | Hurt |
| add Gumbel-Softmax noise to routing (ablation) | MoE-dVLM · AI2D | accuracy: 58.7 -> 28.9 | Hurt |
| train full multi-turn dialogue as one sequence instead of per-turn SFT (ablation) | TowerAlign · ChartQA | accuracy: 43.8 -> 59.0 | Improved |
| train full multi-turn dialogue as one sequence instead of per-turn SFT (ablation) | TowerAlign · DocVQA (val) | accuracy: 55.3 -> 67.7 | Improved |
| use sparse MoE TowerAlign instead of dense LLaDA-VLM at matched data | LLaDA-VLM · DocVQA (val) | accuracy: 61.2 -> 67.7 | Improved |
Closest papers by meaning
- The Unseen Bias: How Norm Discrepancy in Pre-Norm MLLMs Leads to Visual Information Loss (ICLR 2026 · similarity 0.89)
- Multilinear Mixture of Experts: Scalable Expert Specialization through Factorization (NeurIPS 2024 · similarity 0.88)
- Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE (NeurIPS 2026 · similarity 0.88)
- From Sparse to Soft Mixtures of Experts (ICLR 2024 · similarity 0.88)
- Solving Token Gradient Conflict in Mixture-of-Experts for Large Vision-Language Model (ICLR 2025 · similarity 0.87)
- Machine Vision Therapy: Multimodal Large Language Models Can Enhance Visual Robustness via Denoising In-Context Learning (ICML 2024 · similarity 0.87)
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.
