expert pruning
Expert pruning is a way to make large language models run faster by removing some of their internal "expert" modules. These models, called mixture-of-experts (MoE) models, use many small specialized networks and only activate a few for each task. Pruning permanently deactivates some of these experts, trading a bit of accuracy for speed or memory savings. The key challenge is deciding which experts to remove and how to adjust the remaining ones so the model still works well.
The papers show that careful selection matters more than simply ranking experts individually. One method uses differentiable math to find a complementary set that reconstructs the original output, keeping 97% of math performance after removing 75% of experts. Another prunes only during the decoding phase while keeping full experts for prefill, achieving up to 1.81 times faster throughput. A third approach, which requires no additional training, measures how replaceable each expert is and outperforms other methods on reasoning tasks, though it notes that better task scores don't always mean better generation behavior.
- Papers
- 3
- Released code
- 0
- First seen
- Oct 2026
- Latest
- Oct 2026
3 papers in the last two months, against 0 in the two before.
Start here
Three papers to read first, in the order they came out- 1Pruning experts by replacement damage better preserves MoE reasoning performance
RAZOR prunes replaceable experts in MoE models using consensus residuals, surpassing baselines on reasoning tasks.
Oct 2026 - 2Expert pruning in MoE models can be tailored to decode phase, preserving prefill quality while boosting throughput
SlimWise improves MoE decode throughput by pruning experts only during decode while keeping full prefill.
Oct 2026 - 3Joint expert selection preserves MoE performance under aggressive pruning
Joint reconstruction-based expert selection and weight adaptation preserve MoE accuracy better than independent importance rankings, especially under aggressive pruning.
Oct 2026
How fast it is moving
Often studied with
mixture of expertsA neural architecture with many specialized sub-networks, only a few activated per input via a router, balancing capacity and efficiency.3 togethermodel routingDirecting each query or subtask to the most suitable model to optimize cost, accuracy, or quality.1 togetherprefill-decode disaggregationSplitting LLM serving into separate prefill and decode phases to optimize GPU resource allocation and batch sizing.1 togetherThe papers
Most central to this idea first, not most recent.
- Big Techcs.LG
Joint expert selection preserves MoE performance under aggressive pruning
Joint reconstruction-based expert selection and weight adaptation preserve MoE accuracy better than independent importance rankings, especially under aggressive pruning.
LinkedIn, MIT · Oct 2026
- Chinese Techcs.LG
Pruning experts by replacement damage better preserves MoE reasoning performance
RAZOR prunes replaceable experts in MoE models using consensus residuals, surpassing baselines on reasoning tasks.
Tencent, Code Models · Oct 2026
- Top Universitycs.LG
Expert pruning in MoE models can be tailored to decode phase, preserving prefill quality while boosting throughput
SlimWise improves MoE decode throughput by pruning experts only during decode while keeping full prefill.
a2sys, KAIST · Oct 2026
Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.