Research map / Architectures and efficiency / Mixture of experts
$\phi$-Balancing for Mixture-of-Experts Training
Lizhang Chen, Jonathan Li, Qi Wang, Runlong Liao, Shuozhe Li, Chen Liang, Ni Lao, qiang liu
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 |
|---|---|---|---|
| use phi-balancing (negative entropy) instead of Shannon-free l_p p=1 potential (ablation) | phi-balancing · Mirror map ablation | validation loss: 3.142 -> 3.084 | Improved |
| use negative Shannon entropy instead of l_p p=2 (ablation) | phi-balancing · Mirror map ablation | MaxVio_global: 0.610 -> 0.104 | Improved |
| use negative Shannon entropy instead of softplus potential (ablation) | phi-balancing · Mirror map ablation | MaxVio_global: 0.810 -> 0.104 | Improved |
| use negative Renyi (alpha=0.95) instead of Shannon (ablation) | phi-balancing · Mirror map ablation | validation loss: 3.091 vs 3.084 | No clear effect |
| use phi-balancing instead of ST-MoE | DeepSeek-MoE-16B-Chat · 7-benchmark fine-tuning average | average accuracy: 51.95 -> 53.92 | Improved |
| use phi-balancing instead of ST-MoE | DeepSeek-V2-Lite · 7-benchmark fine-tuning average | average accuracy: 49.29 -> 51.86 | Improved |
| use phi-balancing instead of ST-MoE | Moonlight-16B-A3B-Instruct · 7-benchmark fine-tuning average | average accuracy: 73.37 -> 76.07 | Improved |
| use phi-balancing instead of ST-MoE | Moonlight-16B-A3B-Instruct · Math500 | accuracy: 64.80 -> 66.20 | Hurt |
| use phi-balancing instead of ST-MoE at global batch 32 | Gemma-MoE 986M · MMLU | accuracy: 41.96 -> 42.88 | Improved |
| increase global batch size from 32 to 512 (ablation) | phi-balancing · MMLU | accuracy: 42.88 -> 43.18 | Improved |
| track EMA of routing probabilities instead of selection frequencies (ablation) | phi-balancing · Validation loss 111M | validation loss: 3.089 -> 3.0847 | No clear effect |
| track EMA of routing probabilities instead of selection frequencies (ablation) | phi-balancing · Validation loss 986M | validation loss: 2.598 -> 2.6019 | No clear effect |
Closest papers by meaning
- Mixture-of-Experts with Expert Choice Routing (NeurIPS 2022 · similarity 0.90)
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.89)
- Dense Backpropagation Improves Training for Sparse Mixture-of-Experts (NeurIPS 2025 · similarity 0.89)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.87)
- Drop-Upcycling: Training Sparse Mixture of Experts with Partial Re-initialization (ICLR 2025 · similarity 0.87)
- Autonomy-of-Experts Models (ICML 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.
