AI Paper Atlas

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

ICML 2026 · Architectures and efficiency › Mixture of experts · approach: Expert routing and specialization

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 changedWhereMeasuredEffect
add complete expert partition P=4 (8/32 experts) before fine-tuning (ablation)Mixtral-8x7B · 9-benchmark average (fine-tuned)average accuracy (%): 70.53 -> 71.12Improved
add complete expert partition P=2 (4/16 experts) before fine-tuning (ablation)Mixtral-8x7B · 9-benchmark average (fine-tuned)average accuracy (%): 70.53 -> 70.86Improved
add complete expert partition P=4 without fine-tuning (ablation)Mixtral-8x7B · 9-benchmark average (pretrained)average accuracy (%): 70.18 -> 70.19No clear effect
add complete expert partition P=4 before fine-tuning (ablation)Mixtral-8x7B · GSM8K (fine-tuned)accuracy (%): 60.73 -> 62.85Improved
add 1T-Drop with threshold 0.05 to the partitioned fine-tuned modelMixtral-8x7B 8/32 · 9-benchmark averageaverage accuracy (%): 71.12 -> 71.43Improved
add 1T-Drop with threshold 0.05 to the partitioned fine-tuned modelMixtral-8x7B 8/32 · GSM8Kaccuracy (%): 62.85 -> 64.52Improved
add neuron-level expert reconstruction to 2T-Drop (ablation)Mixtral-8x7B 8/32 · 9-benchmark average, 24.0% drop rateaverage accuracy (%): 70.90 -> 71.04Improved
replace 1T-Drop with 2T-Drop (Reconstruct) at similar drop rateOLMoE-Instruct · 9-benchmark averageaverage accuracy (%): 65.10 -> 65.63Improved
replace 1T-Drop with 2T-Drop (Reconstruct) at similar drop rateDeepSeek-V2-Lite-Chat · 9-benchmark averageaverage accuracy (%): 67.27 -> 67.65Improved
apply 1T-Drop with no fine-tuning (inference-only, 21.7% drop)OLMoE-Instruct · 9-benchmark averageaverage accuracy (%): 65.91 -> 65.10Hurt
use accumulated absolute gate-up instead of accumulated gate-up importance profiling (ablation)2T-Drop (Reconstruct) · 9-benchmark averageaverage accuracy (%): 66.79 -> 67.65Improved
apply 2T-Drop to long-context and generation benchmarksOLMoE · HumanEval/MATH500/RULER averageaverage score: 35.3 -> 35.1No clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy