Research map / Architectures and efficiency / Mixture of experts
HEAPr: Hessian-based Efficient Atomic Expert Pruning in Output Space
Ke Li, Zheng Yang, Zhongbin Zhou, Xuefeng, Zhonglin Jiang, Wenxiao Wang
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 |
|---|---|---|---|
| global ranking of atomic experts instead of layer-wise ranking (ablation) | HEAPr · DeepSeekMoE-16B-Base 20% | average accuracy: 60.03 -> 60.68 | Improved |
| global ranking instead of layer-wise ranking (ablation) | HEAPr · DeepSeekMoE-16B-Base 40% | average accuracy: 56.99 -> 57.80 | Improved |
| global ranking instead of layer-wise ranking (ablation) | HEAPr · Qwen1.5-MoE-A2.7B-Chat 25% | average accuracy: 52.52 -> 53.59 | Improved |
| global ranking instead of layer-wise ranking (ablation) | HEAPr · Qwen1.5-MoE-A2.7B-Chat 50% | average accuracy: 48.15 -> 48.57 | Improved |
| Hessian-based importance instead of CAMERA-P heuristic, both layer-wise (ablation) | HEAPr-L vs CAMERA-P · DeepSeekMoE-16B-Base 20% | average accuracy: 59.52 -> 60.03 | Improved |
| Hessian-based importance instead of CAMERA-P at 40% layer-wise (ablation) | HEAPr-L vs CAMERA-P · DeepSeekMoE-16B-Base 40% | average accuracy: 56.87 -> 56.99 | No clear effect |
| atomic expert pruning instead of expert-level pruning (ablation) | HEAPr · DeepSeekMoE-16B-Base 40% | WikiText-2 perplexity: 8.00 -> 6.80 | Improved |
| atomic expert pruning instead of expert-level pruning (ablation) | HEAPr · DeepSeekMoE-16B-Base 20% | WikiText-2 perplexity: 6.90 -> 6.54 | Improved |
| atomic expert pruning instead of expert-level pruning (ablation) | HEAPr · DeepSeekMoE-16B-Base 40% | FLOPs reduction ratio: 0% -> 21% | Improved |
| calibrate on C4 instead of WikiText-2 (ablation) | HEAPr · DeepSeekMoE-16B-Base 20% | average accuracy | No clear effect |
| increase number of calibration samples (ablation) | HEAPr · DeepSeekMoE-16B-Base 20% | average accuracy | Improved |
| HEAPr at 20-25% pruning | HEAPr · Qwen3-30B-A3B 25% | average accuracy: 0.62 -> 0.59 | Hurt |
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.
- MoE-SVD — MoE-SVD: Structured Mixture-of-Experts LLMs Compression via Singular Value Decomposition (ICML 2025) · better on 2 of 2 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
- Mixture Compressor for Mixture-of-Experts LLMs Gains More (ICLR 2025 · similarity 0.89)
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · similarity 0.89)
- LightMoE: Reducing Mixture-of-Experts Redundancy through Expert Replacing (NeurIPS 2026 · similarity 0.88)
- FlexMoE: One-for-All Nested Intra-Expert Pruning for MoE Language Models (NeurIPS 2026 · similarity 0.88)
- Domain-Specific Pruning of Large Mixture-of-Experts Models with Few-shot Demonstrations (NeurIPS 2025 · similarity 0.88)
- DiEP: Adaptive Mixture-of-Experts Compression through Differentiable Expert Pruning (NeurIPS 2025 · 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.
