Research map / Vision-language and multimodal / Cross-modal learning
Symbiotic-MoE: Unlocking the Synergy between Generation and Understanding
Xiangyue Liu, Zijian Zhang, Miles Yang, Zhao Zhong, Liefeng Bo, Ping Tan
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 generative training with shared-expert bridge | Symbiotic-MoE vs Only_LM_MMU · MMLU | accuracy: 0.405 -> 0.507 | Improved |
| add generative training with shared-expert bridge | Symbiotic-MoE vs Only_LM_MMU · OCRBench | score: 662 -> 768 | Improved |
| add shared expert (ablation) | original VLM experts · MMLU at iteration 0 | accuracy: 0.460 -> 0.697 | Improved |
| merge Text and ViT into one understanding group (bimodal) instead of tripartite split (ablation) | expert grouping · MMLU at iteration 0 | accuracy: 0.244 -> 0.601 | Improved |
| assign 96 experts to understanding instead of 32 (ablation) | bimodal split · MMLU at iteration 0 | accuracy: 0.285 -> 0.601 | Improved |
| assign 96 experts to understanding instead of 64 (ablation) | bimodal split · MMLU at iteration 0 | accuracy: 0.453 -> 0.601 | Improved |
| assign 96 experts to understanding instead of 86 (ablation) | bimodal split · MMLU at iteration 0 | accuracy: 0.561 -> 0.601 | Improved |
| add warmup gradient shielding on shared experts (ablation) | Symbiotic-MoE · MMLU and OCRBench | accuracy | Improved |
| use 1e-6 instead of 1e-4 learning rate for understanding experts (ablation) | Symbiotic-MoE · MMLU and OCRBench | accuracy | Improved |
| add shared experts and routing disentanglement versus MoT | Symbiotic-MoE vs MoT · COCO-30K | FID: 19.87 -> 13.65 | Improved |
| replace standard MoE with MoT structural isolation | MoT vs Standard MoE · MMLU | accuracy: 0.308 -> 0.392 | Improved |
| use independent QKV (Bagel) instead of MoT | Bagel vs MoT · T2I-CompBench | score: 0.43 -> 0.45 | Improved |
Closest papers by meaning
- Rosetta: Composable Native Multimodal Pretraining (NeurIPS 2026 · similarity 0.90)
- Breaking Dual Bottlenecks: Evolving Unified Multimodal Models into Self-Adaptive Interleaved Visual Reasoners (ICML 2026 · similarity 0.88)
- R2-T2: Re-Routing in Test-Time for Multimodal Mixture-of-Experts (ICML 2025 · similarity 0.87)
- $\texttt{I$^2$MoE}$: Interpretable Multimodal Interaction-aware Mixture-of-Experts (ICML 2025 · similarity 0.87)
- DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models (ICML 2026 · similarity 0.87)
- FLAME: Adaptive Mixture-of-Experts for Continual Multimodal Multi-Task Learning (NeurIPS 2026 · 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.
