Research map / Architectures and efficiency / Mixture of experts
MoDES: Accelerating Mixture-of-Experts Multimodal Large Language Models via Dynamic Expert Skipping
Yushi Huang, Zining Wang, Zhihang Yuan, Yifu Ding, Ruihao Gong, Jinyang Guo, Xianglong Liu, Jun Zhang
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 GMLG (global layer importance) to single-threshold skipping (ablation) | Thresholding · ChartQA, skip 83% experts | accuracy: 76.74 -> 79.28 | Improved |
| add GMLG to DMT (ablation) | DMT · MMBench, skip 83% experts | accuracy: 79.42 -> 81.44 | Improved |
| use separate text/vision thresholds (DMT) instead of single threshold (ablation) | Thresholding · MMBench, skip 83% experts | accuracy: 65.48 -> 79.42 | Improved |
| add GMLG and DMT together (ablation) | Thresholding · VideoMMMU, skip 83% experts | accuracy: 40.33 -> 47.11 | Improved |
| add GMLG to DMT (ablation) | DMT · MME, skip 83% experts | score: 2081 -> 2162 | Improved |
| use frontier search instead of naive search (ablation) | MoDES · 13 benchmarks, skip 83% | Avg. relative score (%): 96.24 -> 96.25 | No clear effect |
| calibrate and search on COCO instead of GQA (ablation) | MoDES · ChartQA, skip 83% experts | accuracy: 84.20 -> 86.56 | Mixed |
| use MoDES instead of MC-MoE at 83% skipping | Kimi-VL-A3B-Instruct · 13 benchmarks | Avg. relative score (%): 88.32 -> 96.25 | Improved |
| use MoDES instead of MC-MoE at 88% skipping | Qwen3-VL-MoE-30B-A3B · 13 benchmarks | Avg. relative score (%): 86.66 -> 97.33 | Improved |
| use MoDES instead of DiEP at 50% skipping | Kimi-VL-A3B-Instruct · 13 benchmarks | Avg. relative score (%): 98.17 -> 99.91 | Improved |
| combine MoDES with 1.5-bit quantization instead of MC-MoE | Kimi-VL-A3B-Instruct · ChartQA, skip 67% | accuracy: 69.46 -> 72.28 | Improved |
| skip experts with MoDES instead of default top-k | Kimi-VL-A3B-Instruct · RealWorldQA, skip 50% | accuracy: 65.36 -> 65.62 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DiEP — DiEP: Adaptive Mixture-of-Experts Compression through Differentiable Expert Pruning (NeurIPS 2025) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- CuMo: Scaling Multimodal LLM with Co-Upcycled Mixture-of-Experts (NeurIPS 2024 · similarity 0.90)
- Routing Experts: Learning to Route Dynamic Experts in Existing Multi-modal Large Language Models (ICLR 2025 · similarity 0.90)
- BAM! Just Like That: Simple and Efficient Parameter Upcycling for Mixture of Experts (NeurIPS 2024 · similarity 0.90)
- MoME: Mixture of Multimodal Experts for Generalist Multimodal Large Language Models (NeurIPS 2024 · similarity 0.90)
- Solving Token Gradient Conflict in Mixture-of-Experts for Large Vision-Language Model (ICLR 2025 · similarity 0.90)
- Accelerating Multimodal Large Language Models by Searching Optimal Vision Token Reduction (CVPR 2025 · similarity 0.89)
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.
