Matryoshka Model Learning for Improved Elastic Student Models Conference: Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2; August 3–7, 2025; Toronto, ON, Canada Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD ’25)
Chetan Verma, Aditya Srinivas Timmaraju, Cho-Jui Hsieh, Suyash Damle, Ngot Bui, Yang Zhang, Wen Chen, Xin Liu, Prateek Jain, Inderjit Dhillon · 2025
Read Matryoshka Model Learning for Improved Elastic Student Models Conference: Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2; August 3–7, 2025; Toronto, ON, Canada Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD ’25) by Chetan Verma, Aditya Srinivas Timmaraju, Cho-Jui Hsieh, Suyash Damle, Ngot Bui, Yang Zhang, Wen Chen, Xin Liu, Prateek Jain, Inderjit Dhillon (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-2505-23337.txt.
Description
Chetan Verma and nine co-authors train families of models in twenty twenty-five. Industry models must meet serving limits that change fast, and rebuilding for each limit is costly. Their framework, MatTA, trains several accurate student models in one run. It uses a Teacher-TA-Student recipe, where TA models sit between teacher and student in size.
The TA models have more capacity, so students can relate better to the teacher. They also bring more domain expertise into the chain. Multiple students can then be drawn from one TA model. That gives servable options that trade accuracy against cost. Tests on proprietary data and live A/B tests in a production system backed the method. Readers get one training run yielding many deployable sizes.
Catalog record
| CREATOR | Chetan Verma, Aditya Srinivas Timmaraju, Cho-Jui Hsieh, Suyash Damle, Ngot Bui, Yang Zhang, Wen Chen, Xin Liu, Prateek Jain, Inderjit Dhillon |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 40 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-23337