# Training Matryoshka Mixture-of-Experts for Elastic Inference-Time Expert Utilization

arXiv: https://arxiv.org/abs/2509.26520 · HTML: https://mlsanity.com/paper/2509.26520/

## Key facts

- Cited by 7 papers in the map (data through 14 Oct 2026).
- Placed in MoE large language models, one of 12 directions in Mixture-of-experts.
- Builds on 28 of its 31 references that are in the map.
- First posted on arXiv 30 Sep 2025.

## Placed in

- Mixture-of-experts / [MoE large language models](https://mlsanity.com/moe/moe-large-language-models/)

## Builds on (most cited first, 28 of 28)

- [Qwen3 Technical Report](https://mlsanity.com/paper/2505.09388/) (2025)
- [Measuring Massive Multitask Language Understanding](https://mlsanity.com/paper/2009.03300/) (2020)
- [DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning](https://mlsanity.com/paper/2501.12948/) (2025)
- [Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge](https://mlsanity.com/paper/1803.05457/) (2018)
- [HellaSwag: Can a Machine Really Finish Your Sentence?](https://mlsanity.com/paper/1905.07830/) (2019)
- [Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities](https://mlsanity.com/paper/2507.06261/) (2025)
- [Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer](https://mlsanity.com/paper/1701.06538/) (2017)
- [BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions](https://mlsanity.com/paper/1905.10044/) (2019)
- [Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answering](https://mlsanity.com/paper/1809.02789/) (2018)
- [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://mlsanity.com/paper/1909.08053/) (2019)
- [The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale](https://mlsanity.com/paper/2406.17557/) (2024)
- [DeepSpeed-MoE: Advancing Mixture-of-Experts Inference and Training to Power Next-Generation AI Scale](https://mlsanity.com/paper/2201.05596/) (2022)
- [DataComp-LM: In search of the next generation of training sets for language models](https://mlsanity.com/paper/2406.11794/) (2024)
- [Matryoshka Representation Learning](https://mlsanity.com/paper/2205.13147/) (2022)
- [Retrieve, Program, Repeat: Complex Knowledge Base Question Answering via Alternate Meta-learning](https://mlsanity.com/paper/2010.15875/) (2020)
- [Nemotron-CC: Transforming Common Crawl into a Refined Long-Horizon Pretraining Dataset](https://mlsanity.com/paper/2412.02595/) (2024)
- [MatFormer: Nested Transformer for Elastic Inference](https://mlsanity.com/paper/2310.07707/) (2023)
- [Matryoshka Multimodal Models](https://mlsanity.com/paper/2405.17430/) (2024)
- [Harder Tasks Need More Experts: Dynamic Routing in MoE Models](https://mlsanity.com/paper/2403.07652/) (2024)
- [Dynamic Mixture of Experts: An Auto-Tuning Approach for Efficient Transformer Models](https://mlsanity.com/paper/2405.14297/) (2024)
- [Matryoshka Query Transformer for Large Vision-Language Models](https://mlsanity.com/paper/2405.19315/) (2024)
- [Flextron: Many-in-One Flexible Large Language Model](https://mlsanity.com/paper/2406.10260/) (2024)
- [Towards Greater Leverage: Scaling Laws for Efficient Mixture-of-Experts Language Models](https://mlsanity.com/paper/2507.17702/) (2025)
- [XMoE: Sparse Models with Fine-grained and Adaptive Expert Selection](https://mlsanity.com/paper/2403.18926/) (2024)
- [Federated Model Heterogeneous Matryoshka Representation Learning](https://mlsanity.com/paper/2406.00488/) (2024)
- [Expert Race: A Flexible Routing Strategy for Scaling Diffusion Transformer with Mixture of Experts](https://mlsanity.com/paper/2503.16057/) (2025)
- [Train Once, Deploy Anywhere: Matryoshka Representation Learning for Multimodal Recommendation](https://mlsanity.com/paper/2409.16627/) (2024)
- [Matryoshka Representation Learning for Recommendation](https://mlsanity.com/paper/2406.07432/) (2024)

## Built on by (most cited first, 7 of 7)

- [Variational Inference, Entropy, and Orthogonality: A Unified Theory of Mixture-of-Experts](https://mlsanity.com/paper/2601.03577/) (2026)
- [m3BERT: A Modern, Multi-lingual, Matryoshka Bidirectional Encoder](https://mlsanity.com/paper/2605.19568/) (2026)
- [PARCEL: Pool-Anchored Resampling with Conditioned Elastic Queries for Efficient Vision-Language Understanding](https://mlsanity.com/paper/2605.30126/) (2026)
- [Do Vision Transformers Need All-to-All Attention? Global Communication Through Elastic Learned Cores](https://mlsanity.com/paper/2605.12491/) (2026)
- [FlexMoE: One-for-All Nested Intra-Expert Pruning for MoE Language Models](https://mlsanity.com/paper/2606.27866/) (2026)
- [Training-Free Halving of Activated Experts in Fine-Grained Mixture-of-Experts Models](https://mlsanity.com/paper/2609.04575/) (2026)
- [Elastic Selective Spectral Hybrids for Train-Once, Export-Many Budgeted Inference](https://mlsanity.com/paper/2609.32486/) (2026)

## Abstract (arXiv)

Mixture-of-Experts (MoE) has emerged as a promising paradigm for efficiently scaling large language models without a proportional increase in computational cost. However, the standard training strategy of Top-K router prevents MoE models from realizing their full potential for elastic inference. When the number of activated experts is altered at inference time, these models exhibit precipitous performance degradation. In this work, we introduce Matryoshka MoE (M-MoE), a training framework that instills a coarse-to-fine structure directly into the expert ensemble. By systematically varying the number of activated experts during training, M-MoE compels the model to learn a meaningful ranking: top-ranked experts collaborate to provide essential, coarse-grained capabilities, while subsequent experts add progressively finer-grained detail. We explore this principle at multiple granularities, identifying a layer-wise randomization strategy as the most effective. Our experiments demonstrate that a single M-MoE model achieves remarkable elasticity, with its performance at various expert counts closely matching that of an entire suite of specialist models, but at only a fraction of the total training cost. This flexibility not only unlocks elastic inference but also enables optimizing performance by allocating different computational budgets to different model layers. Our work paves the way for more practical and adaptable deployments of large-scale MoE models.