MoME in AI: Clarifying Its Multiple Meanings, from Mixture of Matryoshka Experts to Mixture of Modality Experts

Key points

The acronym MoME appears in AI research with multiple distinct meanings. Context usually disambiguates, but the overlapping naming can confuse researchers and practitioners. This guide clarifies the main usages.

1. MoME in Meta AI: Mixture of Matryoshka Experts

• What it is: A framework for efficient, high-performance audio-visual speech recognition (AVSR), developed collaboratively by Imperial College London (iBUG team), Meta AI, and NatWest AI Research. The paper, "MoME: Mixture of Matryoshka Experts for Audio-Visual Speech Recognition," has been submitted to NeurIPS 2025.
• Naming: "Matryoshka" refers to Russian nesting dolls, describing the model's ability to handle information at multiple levels of compression/granularity within a single unified model.
• Core components:
• MoE architecture: sparse computation with multiple expert sub-networks; only a small subset of experts activates per input.
• MRL integration: hierarchical, multi-scale representation learning.
• Shared router: consistent expert activation across scales.
• Pipeline: Audio-visual input → multi-scale processing → shared router → expert selection → expert sub-networks → knowledge fusion → AVSR output.
• Application: AVSR transcribes speech using both audio and visual lip movements, improving robustness in noisy environments where audio-only systems fail.
• Advantages:
• Dynamic capacity allocation via sparse expert activation.
• State-of-the-art performance on LRS2 and LRS3 for AVSR, ASR, and VSR tasks with fewer active parameters.
• Elastic inference and cross-scale knowledge transfer reduce computational cost, enabling deployment on hardware with limited resources.
• Important distinction: MoME is a project within the Meta AI ecosystem but is distinct from the LLaMA series (e.g., LLaMA 4), even though both may share Mixture-of-Experts principles.

2. Broader MoME landscape: Mixture of Modality Experts

• Mixture of Modality Experts is a separate medical AI model developed at HKUST for non-invasive breast cancer diagnosis using multiparametric MRI, routing across experts specialized for different imaging modalities.
• Related variants and concepts using the MoME label exist across multimodal and medical AI literature; always check the paper's domain and authorship.

3. Foundational architecture: Mixture-of-Experts (MoE)

• Both meanings build on the classic MoE design: a router selects a small subset of specialized expert sub-networks per input, decoupling model capacity from per-token compute.
• MoE is widely used in large-scale models (including some Meta models) to gain capacity without proportional inference cost.

4. Comparative notes

| Aspect | MoME (Meta AI, AVSR) | Mixture of Modality Experts (HKUST) | |---|---|---| | Domain | Speech recognition (audio-visual) | Medical imaging (breast cancer, MRI) | | Institutions | Imperial College London, Meta AI, NatWest AI Research | HKUST | | Key idea | Matryoshka (multi-granularity) + sparse experts | Modality-specific experts fused for diagnosis | | Basis | Mixture-of-Experts + MRL | Mixture-of-Experts across modalities |

Conclusion

When encountering "MoME," check the context: speech/multimodal research by Imperial College London and Meta AI implies Mixture of Matryoshka Experts; medical imaging from HKUST implies Mixture of Modality Experts. Both share the Mixture-of-Experts foundation but serve very different applications.