Research map / Architectures and efficiency / Mixture of experts
Deconstructing Pre-training: Knowledge Attribution Analysis in MoE and Dense Models
Bo Wang, Junzhuo Li, Hong Chen, Yuanlin Chu, Yuxuan Fan, Xuming Hu
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 MoE (OLMoE) instead of dense (OLMo) architecture | pre-training · FFN neurons | top-1% set stability / Rt: Rt 26.4 -> 48.9 | Improved |
| use MoE instead of dense architecture | pre-training · FFN layers (10K-600K) | layer-distribution consistency rho_avg: 0.54 -> 0.97 | Improved |
| use MoE instead of dense architecture | pre-training · ATTN layers (10K-600K) | sigma_rel: 8.64 -> 0.49 | Improved |
| mask top-1% FFN neurons (ablation) | OLMoE vs OLMo · relational HIT@10 | performance drop: OLMoE 35.47% vs OLMo 96.19% drop | Hurt |
| mask top-10 attention heads (ablation) | OLMoE vs OLMo · relational HIT@10 | performance drop: OLMoE 9.44% vs OLMo 50.43% drop | Hurt |
| mask the single most important attention head (ablation) | OLMoE vs OLMo · relational HIT@10 | performance drop: OLMoE 0.06% vs OLMo 16.46% drop | Mixed |
Closest papers by meaning
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · similarity 0.88)
- Mixture of Experts Made Intrinsically Interpretable (ICML 2025 · similarity 0.88)
- Joint MoE Scaling Laws: Mixture of Experts Can Be Memory Efficient (ICML 2025 · similarity 0.88)
- Drop-Upcycling: Training Sparse Mixture of Experts with Partial Re-initialization (ICLR 2025 · similarity 0.88)
- Domain-Specific Pruning of Large Mixture-of-Experts Models with Few-shot Demonstrations (NeurIPS 2025 · similarity 0.88)
- Mixture-of-Experts Meets Instruction Tuning: A Winning Combination for Large Language Models (ICLR 2024 · similarity 0.88)
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.
