Research map / Vision-language and multimodal / Human and embodied modeling
PRISM: Synergizing Vision Foundation Models via Self-organized Expert Specialization
Ying Tang, Dong Li, Youjia Zhang, Zikai Song, Junqing Yu, Wei Yang
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 |
|---|---|---|---|
| replace dense multi-teacher distillation by dual-stream MoE (PRISM) over SAK | PRISM ViT-B · PASCAL-Context | delta_m (%): 0.83 -> 2.29 | Improved |
| PRISM vs SAK on indoor scenes | PRISM ViT-B · NYUD-v2 | delta_m (%): 11.11 -> 10.59 | Mixed |
| scale student from ViT-B to ViT-L | PRISM · PASCAL-Context | delta_m (%): SAK 2.30 -> PRISM 3.16 | Improved |
| add MoE to dense baseline (ablation) | PRISM ViT-S · PASCAL-Context (ablation) | Semseg mIoU: 78.04 -> 79.19 | Improved |
| add Universal Anchor stream (ablation) | PRISM ViT-S · PASCAL-Context (ablation) | Normal mErr: 14.65 -> 14.28 | Improved |
| use FiLM-based router instead of task embedding concatenation (ablation) | PRISM ViT-S · PASCAL-Context (ablation) | Semseg mIoU: 78.87 -> 79.19 | Improved |
| add locality-aware decorrelation loss (ablation) | PRISM ViT-S · PASCAL-Context (ablation) | Parsing mIoU: 61.80 -> 69.25 | Improved |
| add locality-aware decorrelation loss (ablation) | PRISM ViT-S · PASCAL-Context (ablation) | Semseg mIoU: 79.80 -> 79.19 | Hurt |
| PRISM vs iso-FLOPs Wide ViT-S (ablation) | PRISM ViT-S · PASCAL-Context (ablation) | Semseg mIoU: 78.15 -> 79.19 | Improved |
| PRISM vs vanilla sparse MoE (ablation) | PRISM ViT-S · PASCAL-Context (ablation) | Semseg mIoU: 72.96 -> 79.19 | Improved |
| reduce expert MLP ratio to 1.0 (PRISM-lite) vs SAK ViT-S | PRISM-lite ViT-S · PASCAL-Context | delta_m (%): 0.43 -> 0.61 | Improved |
| sparse MoE dispatch vs SAK ViT-S | PRISM-lite · Efficiency | Latency (ms): 20.25 -> 47.51 | Hurt |
Closest papers by meaning
- Multilinear Mixture of Experts: Scalable Expert Specialization through Factorization (NeurIPS 2024 · similarity 0.90)
- Swiss Army Knife: Synergizing Biases in Knowledge from Vision Foundation Models for Multi-Task Learning (ICLR 2025 · similarity 0.89)
- Mixture of Nested Experts: Adaptive Processing of Visual Tokens (NeurIPS 2024 · similarity 0.87)
- Hawaii: Hierarchical Visual Knowledge Transfer for Efficient Vision-Language Models (NeurIPS 2025 · similarity 0.87)
- Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE (NeurIPS 2026 · similarity 0.87)
- MoDE: A Mixture-of-Experts Model with Mutual Distillation among the Experts (AAAI 2024 · similarity 0.86)
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.
