Sampling at intermediate temperatures is optimal for training large language models in protein structure prediction
L. Ghiringhelli, A. Zambon, G. Tiana · 2026
Read Sampling at intermediate temperatures is optimal for training large language models in protein structure prediction by L. Ghiringhelli, A. Zambon, G. Tiana (2026) 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-2603-29529.txt.
Description
Ghiringhelli and co-authors probe how transformer models learn protein structure from sequences. They sample the loss landscape at varied temperatures with Langevin dynamics. Unlike non-attention networks, transformers show no sharp first-order-like transition in loss. That leaves a band of mid-range temperatures with good learning traits. At those temperatures, most layer weights stay highly conserved. The team gives a practical way to pick the embedding size. Models sampled there also predict mutation energy shifts better than standard training.
Catalog record
| CREATOR | L. Ghiringhelli, A. Zambon, G. Tiana |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 62338 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.29529