Local Mixtures of Experts: Essentially Free Test-Time Training via Model Merging
Ryo Bertolissi, Jonas Hübotter, Ido Hakimi, Andreas Krause · 2025
Read Local Mixtures of Experts: Essentially Free Test-Time Training via Model Merging by Ryo Bertolissi, Jonas Hübotter, Ido Hakimi, Andreas Krause (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-14136.txt.
Description
Ryo Bertolissi and co-authors scale up mixture-of-experts models by merging experts at test time, dodging the cost of training each one per prompt. Their 2025 paper starts from a trend: models that route each input to a few specialist sub-models gain skill without a matching rise in run cost. Few experts are used today, since more would cost too much to train and run.
Their method builds far more experts and fuses them on the fly for each task. They show this fusion acts like a cheap stand-in for test-time training, where a model is fine-tuned for each prompt at great expense. With more experts the scores near that pricey method, and on a small base model the fusion ran a hundred times faster.
Catalog record
| CREATOR | Ryo Bertolissi, Jonas Hübotter, Ido Hakimi, Andreas Krause |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 55 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-14136