Research map / Architectures and efficiency / Mixture of experts
Synergistic Intra- and Cross-Layer Regularization Losses for MoE Expert Specialization
Rizhen Hu, Yuan Cao, Boao Kong, Mou Sun, Kun Yuan
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 intra-layer specialization and cross-layer coupling losses | vanilla MoE small · C4 validation | perplexity: 14.01 -> 13.75 | Improved |
| add sp and cp losses on top of load balance and z-loss | DeepSeek-style MoE medium · C4 validation | perplexity: 12.07 -> 11.99 | Improved |
| add orthogonality and logit-variance auxiliary loss (Guo et al.) | load-balance MoE · C4 validation | perplexity: 14.01 -> 14.27 | Hurt |
| add sp and cp losses | DeepSeek-MoE-16B LoRA SFT · MMLU | accuracy: 0.4143 -> 0.4586 | Improved |
| add sp and cp losses | DeepSeek-V2-Lite LoRA SFT · MMLU-Pro | accuracy: 0.2605 -> 0.3108 | Improved |
| add sp and cp losses | Ling-mini-2.0 with AuxLossFree · GSM8K | accuracy: 0.8173 -> 0.8559 | Improved |
| add sp and cp losses | Qwen3-30B-A3B full-parameter FT · HumanEval | pass@1: 92.07 -> 95.73 | Improved |
| add z-loss to load balance (ablation) | DeepSeek-MoE-16B LoRA SFT · MMLU | accuracy: 0.4143 -> 0.3717 | Hurt |
| add sp and cp losses; fewer activated experts | medium MoE · C4 validation | perplexity: 12.13 (N=6) vs 12.15 (baseline N=10) | Improved |
| add sp and cp losses; smaller expert pool | medium MoE · C4 validation | perplexity: 11.93 (E=96) vs 11.94 (baseline E=112) | Improved |
| add specialization loss to load balance (ablation) | 1.1B MoE pretraining · training perplexity | perplexity | Improved |
| add sp loss to model with cp loss (ablation) | 0.4B MoE · coupling loss | Rcp | Improved |
Closest papers by meaning
- Dense Backpropagation Improves Training for Sparse Mixture-of-Experts (NeurIPS 2025 · similarity 0.90)
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.90)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.89)
- From Sparse to Soft Mixtures of Experts (ICLR 2024 · similarity 0.89)
- On the Expressive Power of Mixture-of-Experts for Structured Complex Tasks (NeurIPS 2025 · similarity 0.89)
- Domain-Specific Pruning of Large Mixture-of-Experts Models with Few-shot Demonstrations (NeurIPS 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.
