Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2603.11535 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866915856675504128 |
|---|---|
| author | Sun, Hanchi Liu, Yixin Wu, Yonghui Sun, Lichao |
| author_facet | Sun, Hanchi Liu, Yixin Wu, Yonghui Sun, Lichao |
| contents | Token-choice Mixture-of-Experts (TC-MoE) routes each token to a fixed number of experts, limiting dynamic computation allocation and requiring auxiliary losses to maintain load balance. We propose Expert Threshold (ET) routing, where each expert maintains an exponential moving average (EMA) threshold estimated from the global token distribution. At both training and inference, each token is independently routed to an expert if its score exceeds the expert's threshold, enabling dynamic computation allocation while achieving load balance without auxiliary losses. This fully causal mechanism eliminates dependence on other tokens in the batch, making it well-suited for autoregressive language modeling. In pretraining experiments scaling to 2.4B parameters on FineWeb-Edu, ET achieves 0.067 lower cross-entropy loss than TC-MoE, equivalent to reaching the same performance with 1.6$\times$ fewer tokens. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_11535 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Expert Threshold Routing for Autoregressive Language Modeling with Dynamic Computation Allocation and Load Balancing Sun, Hanchi Liu, Yixin Wu, Yonghui Sun, Lichao Artificial Intelligence Computation and Language Token-choice Mixture-of-Experts (TC-MoE) routes each token to a fixed number of experts, limiting dynamic computation allocation and requiring auxiliary losses to maintain load balance. We propose Expert Threshold (ET) routing, where each expert maintains an exponential moving average (EMA) threshold estimated from the global token distribution. At both training and inference, each token is independently routed to an expert if its score exceeds the expert's threshold, enabling dynamic computation allocation while achieving load balance without auxiliary losses. This fully causal mechanism eliminates dependence on other tokens in the batch, making it well-suited for autoregressive language modeling. In pretraining experiments scaling to 2.4B parameters on FineWeb-Edu, ET achieves 0.067 lower cross-entropy loss than TC-MoE, equivalent to reaching the same performance with 1.6$\times$ fewer tokens. |
| title | Expert Threshold Routing for Autoregressive Language Modeling with Dynamic Computation Allocation and Load Balancing |
| topic | Artificial Intelligence Computation and Language |
| url | https://arxiv.org/abs/2603.11535 |