Research map / Architectures and efficiency / Mixture of experts
Mining Tensor/Neuron-Level Sparsity to Maximize Mixture-of-Experts Potential in Post-Training and Inference
Weilin Cai, Le Qin, Shwai He, Junwei Cui, Ang Li, Jiayi Huang
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 complete expert partition P=4 (8/32 experts) before fine-tuning (ablation) | Mixtral-8x7B · 9-benchmark average (fine-tuned) | average accuracy (%): 70.53 -> 71.12 | Improved |
| add complete expert partition P=2 (4/16 experts) before fine-tuning (ablation) | Mixtral-8x7B · 9-benchmark average (fine-tuned) | average accuracy (%): 70.53 -> 70.86 | Improved |
| add complete expert partition P=4 without fine-tuning (ablation) | Mixtral-8x7B · 9-benchmark average (pretrained) | average accuracy (%): 70.18 -> 70.19 | No clear effect |
| add complete expert partition P=4 before fine-tuning (ablation) | Mixtral-8x7B · GSM8K (fine-tuned) | accuracy (%): 60.73 -> 62.85 | Improved |
| add 1T-Drop with threshold 0.05 to the partitioned fine-tuned model | Mixtral-8x7B 8/32 · 9-benchmark average | average accuracy (%): 71.12 -> 71.43 | Improved |
| add 1T-Drop with threshold 0.05 to the partitioned fine-tuned model | Mixtral-8x7B 8/32 · GSM8K | accuracy (%): 62.85 -> 64.52 | Improved |
| add neuron-level expert reconstruction to 2T-Drop (ablation) | Mixtral-8x7B 8/32 · 9-benchmark average, 24.0% drop rate | average accuracy (%): 70.90 -> 71.04 | Improved |
| replace 1T-Drop with 2T-Drop (Reconstruct) at similar drop rate | OLMoE-Instruct · 9-benchmark average | average accuracy (%): 65.10 -> 65.63 | Improved |
| replace 1T-Drop with 2T-Drop (Reconstruct) at similar drop rate | DeepSeek-V2-Lite-Chat · 9-benchmark average | average accuracy (%): 67.27 -> 67.65 | Improved |
| apply 1T-Drop with no fine-tuning (inference-only, 21.7% drop) | OLMoE-Instruct · 9-benchmark average | average accuracy (%): 65.91 -> 65.10 | Hurt |
| use accumulated absolute gate-up instead of accumulated gate-up importance profiling (ablation) | 2T-Drop (Reconstruct) · 9-benchmark average | average accuracy (%): 66.79 -> 67.65 | Improved |
| apply 2T-Drop to long-context and generation benchmarks | OLMoE · HumanEval/MATH500/RULER average | average score: 35.3 -> 35.1 | No clear effect |
Closest papers by meaning
- Parameters vs FLOPs: Scaling Laws for Optimal Sparsity for Mixture-of-Experts Language Models (ICML 2025 · similarity 0.91)
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · similarity 0.90)
- Exploiting Activation Sparsity with Dense to Dynamic-k Mixture-of-Experts Conversion (NeurIPS 2024 · similarity 0.90)
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.90)
- Mixture of Experts Made Intrinsically Interpretable (ICML 2025 · similarity 0.90)
- Mixture-of-Experts Meets Instruction Tuning: A Winning Combination for Large Language Models (ICLR 2024 · 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.
