Research map / Architectures and efficiency / Mixture of experts
MoE-GRPO: Optimizing Mixture-of-Experts via Reinforcement Learning in Vision-Language Models
Dohwan Ko, Jinyoung Park, Seoung Choi, Sanghyeok Lee, Seohyun Lee, Hyunwoo J. Kim
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 deterministic top-K fine-tuning with MoE-GRPO | InternVL3.5 MoE · 8 multimodal benchmarks avg | accuracy (%): 54.0 -> 56.0 | Improved |
| add stochastic expert sampling (multinomial or noise) without RL | InternVL3.5 MoE · 8 multimodal benchmarks avg | accuracy (%): 54.0 -> 53.7 (multi), 54.3 (noise) | No clear effect |
| apply MoE-GRPO instead of Det-FT on CLIP-MoE | CLIP-MoE · Cross-dataset eval avg (10 datasets) | accuracy (%): 68.8 -> 71.9 | Improved |
| fine-tune CLIP-MoE with Det-FT | CLIP-MoE · Cross-dataset eval avg | accuracy (%): 70.8 -> 68.8 | Hurt |
| apply MoE-GRPO instead of Det-FT for domain generalization | CLIP-MoE · ImageNet variants avg | accuracy (%): 66.0 -> 67.5 | Improved |
| add Gate-GRPO to Token-GRPO (ablation) | MoE-GRPO · 4-benchmark ablation avg | accuracy (%): 53.9 -> 55.7 | Improved |
| add Token-GRPO to Gate-GRPO (ablation) | MoE-GRPO · 4-benchmark ablation avg | accuracy (%): 50.9 -> 55.7 | Improved |
| use modality-aware router guidance instead of Gaussian-noise exploration (ablation) | MoE-GRPO · 4-benchmark ablation avg | accuracy (%): 54.2 -> 55.7 | Improved |
| use modality-aware router guidance instead of multinomial exploration (ablation) | MoE-GRPO · 4-benchmark ablation avg | accuracy (%): 54.8 -> 55.7 | Improved |
| swap GRPO for DAPO or SAPO (ablation) | MoE-GRPO · 4-benchmark ablation avg | accuracy (%): 55.7 -> 55.0 (DAPO), 55.4 (SAPO) | No clear effect |
| use MoE-GRPO instead of Expert Choice or Optimal Transport routing (ablation) | MoE · 4-benchmark ablation avg | accuracy (%): 54.7 / 54.4 -> 55.7 | Improved |
| add load-balancing loss to MoE-GRPO (ablation) | MoE-GRPO · 4-benchmark ablation avg | accuracy (%): 54.8 -> 55.7 | Improved |
Closest papers by meaning
- Solving Token Gradient Conflict in Mixture-of-Experts for Large Vision-Language Model (ICLR 2025 · similarity 0.89)
- Dense Backpropagation Improves Training for Sparse Mixture-of-Experts (NeurIPS 2025 · similarity 0.88)
- Layerwise Recurrent Router for Mixture-of-Experts (ICLR 2025 · similarity 0.88)
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.88)
- Routing Experts: Learning to Route Dynamic Experts in Existing Multi-modal Large Language Models (ICLR 2025 · similarity 0.87)
- Mixture-of-Experts Meets In-Context Reinforcement Learning (NeurIPS 2025 · 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.
