Research map / Architectures and efficiency / Mixture of experts
GeMoE: Gating Entropy is All You Need for Uncertainty-aware Adaptive Routing in MoE-based Large Vision-Language Models
Chaoxiang Cai, Minghe Weng, Jie Li, Yibo Jiang, Longrong Yang, Zequn Qin, Xi Li
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 gating-entropy guidance via monotonic loss (ablation) | GeMoE · MMB/SQA/GQA | average performance: 60.03 -> 64.77 | Improved |
| enforce positive instead of negative entropy-k correlation (ablation) | GeMoE · MMB/SQA/GQA | average performance: 51.57 -> 64.77 | Improved |
| set L_mono weight alpha from 0.0 to 1.0 (ablation) | GeMoE · MolmoE four-benchmark average | average performance: 65.79 -> 71.74 | Improved |
| increase alpha from 1.0 to 1.5 (ablation) | GeMoE · MolmoE four-benchmark average | average performance: 71.74 -> 71.27 | No clear effect |
| replace static Top-8 with GeMoE adaptive routing | MolmoE-1B-7B · Six-benchmark average | average performance: 63.99 -> 63.61 | No clear effect |
| replace static Top-6 with GeMoE adaptive routing | DeepSeek-VL2-Tiny · Six-benchmark average | average performance: 71.29 -> 70.90 | No clear effect |
| replace static Top-8 with GeMoE | MolmoE-1B-7B · Inference FLOPs | GFLOPs per token: 80.65 -> 68.45 | Improved |
| replace static Top-8 with GeMoE | MolmoE-1B-7B · Throughput | tokens/s: 1709 -> 1828 | Improved |
| replace AdaMoE with GeMoE | MolmoE-1B-7B · Six-benchmark average | average performance: 61.77 -> 63.61 | Improved |
| replace Top-p dynamic routing with GeMoE | MolmoE-1B-7B · Six-benchmark average | average performance: 61.18 -> 63.61 | Improved |
| reduce static Top-k from 8 to 5 | MolmoE-1B-7B · Six-benchmark average | average performance: 63.99 -> 60.07 | Hurt |
| replace Top-5 static routing with GeMoE (5.43 experts) | MolmoE-1B-7B · Six-benchmark average | average performance: 60.07 -> 63.61 | Improved |
Closest papers by meaning
- Routing Experts: Learning to Route Dynamic Experts in Existing Multi-modal Large Language Models (ICLR 2025 · similarity 0.89)
- MoE-GRPO: Optimizing Mixture-of-Experts via Reinforcement Learning in Vision-Language Models (CVPR 2026 · similarity 0.89)
- Layerwise Recurrent Router for Mixture-of-Experts (ICLR 2025 · similarity 0.88)
- Ada-K Routing: Boosting the Efficiency of MoE-based LLMs (ICLR 2025 · similarity 0.87)
- Solving Token Gradient Conflict in Mixture-of-Experts for Large Vision-Language Model (ICLR 2025 · similarity 0.87)
- Oracle-MoE: Locality-preserving Routing in the Oracle Space for Memory-constrained Large Language Model Inference (ICML 2025 · 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.
