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
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:10653-10668, 2026.

Abstract

Mixture of Experts (MoE) has emerged as a mainstream architecture for Large Language Models (LLMs), balancing computational efficiency with model scalability. While prior work has explored increasing tensor-level sparsity via finer-grained expert configurations during pre-training, we identify significant unexploited sparsity at both the tensor and neuron levels during post-training and inference. To leverage this, we propose complete expert partition for post-training and threshold-based token-expert dropping for inference. These techniques improve the Mixtral-8$\times$7B model’s average accuracy by 1% across nine downstream benchmarks (notably 4% on GSM8K). To further optimize the accuracy-efficiency trade-off for inference, we introduce dual-threshold token-expert dropping with partial expert partition and reconstruction. Our approach yields a 1.19$\times$ MoE speedup and a 0.5% accuracy gain on Mixtral-8$\times$7B when combining post-training and inference optimizations. For inference-only optimization on OLMoE-Instruct and DeepSeek-V2-Lite-Chat, we achieve up to 1.41$\times$ MoE speedup with a negligible accuracy loss ($<$0.5%).

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-cai26i, title = {Mining {T}ensor/{N}euron-Level Sparsity to Maximize Mixture-of-Experts Potential in Post-Training and Inference}, author = {Cai, Weilin and Qin, Le and He, Shwai and Cui, Junwei and Li, Ang and Huang, Jiayi}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {10653--10668}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/cai26i/cai26i.pdf}, url = {https://proceedings.mlr.press/v306/cai26i.html}, abstract = {Mixture of Experts (MoE) has emerged as a mainstream architecture for Large Language Models (LLMs), balancing computational efficiency with model scalability. While prior work has explored increasing tensor-level sparsity via finer-grained expert configurations during pre-training, we identify significant unexploited sparsity at both the tensor and neuron levels during post-training and inference. To leverage this, we propose complete expert partition for post-training and threshold-based token-expert dropping for inference. These techniques improve the Mixtral-8$\times$7B model’s average accuracy by 1% across nine downstream benchmarks (notably 4% on GSM8K). To further optimize the accuracy-efficiency trade-off for inference, we introduce dual-threshold token-expert dropping with partial expert partition and reconstruction. Our approach yields a 1.19$\times$ MoE speedup and a 0.5% accuracy gain on Mixtral-8$\times$7B when combining post-training and inference optimizations. For inference-only optimization on OLMoE-Instruct and DeepSeek-V2-Lite-Chat, we achieve up to 1.41$\times$ MoE speedup with a negligible accuracy loss ($<$0.5%).} }
Endnote
%0 Conference Paper %T Mining Tensor/Neuron-Level Sparsity to Maximize Mixture-of-Experts Potential in Post-Training and Inference %A Weilin Cai %A Le Qin %A Shwai He %A Junwei Cui %A Ang Li %A Jiayi Huang %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-cai26i %I PMLR %P 10653--10668 %U https://proceedings.mlr.press/v306/cai26i.html %V 306 %X Mixture of Experts (MoE) has emerged as a mainstream architecture for Large Language Models (LLMs), balancing computational efficiency with model scalability. While prior work has explored increasing tensor-level sparsity via finer-grained expert configurations during pre-training, we identify significant unexploited sparsity at both the tensor and neuron levels during post-training and inference. To leverage this, we propose complete expert partition for post-training and threshold-based token-expert dropping for inference. These techniques improve the Mixtral-8$\times$7B model’s average accuracy by 1% across nine downstream benchmarks (notably 4% on GSM8K). To further optimize the accuracy-efficiency trade-off for inference, we introduce dual-threshold token-expert dropping with partial expert partition and reconstruction. Our approach yields a 1.19$\times$ MoE speedup and a 0.5% accuracy gain on Mixtral-8$\times$7B when combining post-training and inference optimizations. For inference-only optimization on OLMoE-Instruct and DeepSeek-V2-Lite-Chat, we achieve up to 1.41$\times$ MoE speedup with a negligible accuracy loss ($<$0.5%).
APA
Cai, W., Qin, L., He, S., Cui, J., Li, A. & Huang, J.. (2026). Mining Tensor/Neuron-Level Sparsity to Maximize Mixture-of-Experts Potential in Post-Training and Inference. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:10653-10668 Available from https://proceedings.mlr.press/v306/cai26i.html.

Related Material