Papers with MoEs
NeKo: Cross-Modality Post-Recognition Error Correction with Tasks-Guided Mixture-of-Experts Language Model (2025.acl-industry)
Copied to clipboard
Yen-Ting Lin, Zhehuai Chen, Piotr Zelasko, Zhen Wan, Xuesong Yang, Zih-Ching Chen, Krishna C Puvvada, Ke Hu, Szu-Wei Fu, Jun Wei Chiu, Jagadeesh Balam, Boris Ginsburg, Yu-Chiang Frank Wang, Chao-Han Huck Yang
| Challenge: | Existing methods to train a model on a mixture of domain datasets require separate correction language models. |
| Approach: | They propose a multi-task correction MoE that trains experts to become an "expert" of speech-to-text, language-totext and vision-to text datasets by learning to route each dataset’s tokens to its mapped expert. |
| Outcome: | The proposed model outperforms GPT-3.5 and Claude-3.5-Sonnet on the Open ASR Leaderboard and reaches an average relative 5.0% WER reduction and substantial improvements in BLEU scores. |
Approximating Two-Layer Feedforward Networks for Efficient Transformers (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Recent work uses sparse Mixtures of Experts (MoEs) to build resource-efficient large language models. |
| Approach: | They propose a general framework that unifies various methods to build two-layer NNs . they propose methods to improve both MoEs and PKMs based on their results . |
| Outcome: | The proposed framework improves both MoEs and product-key memories (PKMs) it shows that MoE's are competitive with dense Transformer-XL on two different scales while being much more resource efficient. |
Mixture of Diverse Size Experts (2024.emnlp-industry)
Copied to clipboard
| Challenge: | Recent large language models (LLMs) have shown superior performance in a variety of tasks due to the sub-linearly increasing computational costs. |
| Approach: | They propose a new MoE architecture with designed layers where experts have different sizes to mitigate this defect. |
| Outcome: | The proposed architecture surpasses existing MoEs by adaptively assigning the parameter budget to experts while maintaining the same total parameter size and number of experts. |
ModularMoE: Fast LLM Customization with Parameter-Sharing Mixture-of-Experts for Low-Resource Settings (2026.findings-acl)
Copied to clipboard
Jiaxing Liu, Qi Qi, Haifeng Sun, Dunjun Li, Zirui Zhuang, Bo He, Xiang Yang, Cong Liu, Jianxin Liao, Jingyu Wang
| Challenge: | Large Language Models impose significant computational and storage burdens on personal devices . existing customization approaches incur excessive computational costs or lead to suboptimal performance . |
| Approach: | They propose a training framework that converts pre-trained LLMs into parameter-sharing MoE models for lightweight deployment. |
| Outcome: | The proposed training framework outperforms state-of-the-art training frameworks at the same sparsity level while delivering up to 2.71 inference speedup. |
Demons in the Detail: On Implementing Load Balancing Loss for Training Specialized Mixture-of-Expert Models (2025.acl-long)
Copied to clipboard
Zihan Qiu, Zeyu Huang, Bo Zheng, Kaiyue Wen, Zekun Wang, Rui Men, Ivan Titov, Dayiheng Liu, Jingren Zhou, Junyang Lin
| Challenge: | Existing Mixture-of-Experts training frameworks use a micro-batch to calculate LBL . micro-batches are restricted to a single sequence, preventing expert specialization . |
| Approach: | They propose to use a global-batch to loosen the load balance constraint for MoEs models . they propose to synchronize fi across micro-batches and then use it to calculate the LBL . |
| Outcome: | The proposed global-batch LBL improves the domain specialization of experts . the micro-battery LBL is almost at the sequence level, and the router is pushed to distribute the token evenly . |
Breaking ReLU Barrier: Generalized MoEfication for Dense Pretrained Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to convert pretrained dense models to MoEs are limited to ReLU-based models with natural sparsity. |
| Approach: | They propose a G-MoEfication approach for arbitrary dense models where activation sparsity assumptions no longer hold. |
| Outcome: | The proposed method reduces the inference cost associated with dense models by sparsely activating experts. |
STUN: Structured-Then-Unstructured Pruning for Scalable MoE Pruning (2025.acl-long)
Copied to clipboard
| Challenge: | Mixture-of-experts (MoEs) have been adopted for reducing inference costs by sparsely activating experts in large language models (LLMs). |
| Approach: | They propose a structured-then-unstructured approach outperforming both of structured and unstructured pruning for MoEs. |
| Outcome: | The proposed approach outperforms both of structured and unstructured pruning, especially for MoEs with hundreds of experts. |
Efficient Large Scale Language Modeling with Mixtures of Experts (2022.emnlp-main)
Copied to clipboard
Mikel Artetxe, Shruti Bhosale, Naman Goyal, Todor Mihaylov, Myle Ott, Sam Shleifer, Xi Victoria Lin, Jingfei Du, Srinivasan Iyer, Ramakanth Pasunuru, Giridharan Anantharaman, Xian Li, Shuohui Chen, Halil Akin, Mandeep Baines, Louis Martin, Xing Zhou, Punit Singh Koura, Brian O’Horo, Jeffrey Wang, Luke Zettlemoyer, Mona Diab, Zornitsa Kozareva, Veselin Stoyanov
| Challenge: | Mixture of Experts layers (MoEs) enable efficient scaling of language models . large autoregressive language models such as GPT-3 can be adapted to a wide range of tasks . |
| Approach: | They propose to use Mixture of Experts layers to enable efficient scaling of language models . they find that MoEs are substantially more compute efficient than dense models compared to MoE models - but only when they are more modestly trained . |
| Outcome: | The proposed model outperforms dense models in a wide range of tasks and domains. |
Union-of-Experts: Neurons in Mixture-of-Experts are Secretly Routers (2026.acl-long)
Copied to clipboard
| Challenge: | Mixture-of-Experts (MoE) models rely on an external router to assign tokens to experts, resulting in suboptimal performance. |
| Approach: | They propose an MoE variant that performs "expert-autonomous routing" by pre-designating a fraction of neurons within each expert as "routing neurons" they pre-train UoE models with up to 3B parameters and show they outperform traditional MoEs with matched efficiency. |
| Outcome: | The proposed model outperforms existing models with 3B parameters and provides valuable insights into expert-autonomous selection and the broader routing mechanisms of MoE models. |
Knowledge Injection Exists in MoE? Exploring Expert-Aware Contrast Decoding in MoE for Mitigating LLMs’ Hallucinations (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods to mitigate hallucinations include prompt engineering and model optimization, but lack domain generalization and potential errors in fine-tuning data may exacerbate the hallucism. |
| Approach: | They propose an expert-aware adaptive contrast decoding that uses expert differences in MoE’s higher layers to mitigate hallucinations on QA tasks. |
| Outcome: | The proposed method outperforms baseline models on four datasets Large language models (LLMs) show strong performance but suffer from hallucinations, limiting their application. |