Scale-Dependent Input Representation and Confidence Estimation for LLMs in Materials Property Prediction
Shuichiro Ozawa, Izumi Takahara, Teruyasu Mizoguchi · 2026
Read Scale-Dependent Input Representation and Confidence Estimation for LLMs in Materials Property Prediction by Shuichiro Ozawa, Izumi Takahara, Teruyasu Mizoguchi (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-2605-03515.txt.
Description
Shuichiro Ozawa, Izumi Takahara, and Teruyasu Mizoguchi study how input form and model scale interact when language models predict materials properties. Which representation works best, and how confidence can be judged, have both stayed unclear for this growing use of language models in materials science.
Fine-tuning two Llama models of different sizes on crystal data, they test five input forms for formation energy and bandgap. Small models favored compact inputs, while the larger model handled long descriptions and structure files well. After fine-tuning, token likelihood of the predicted numbers tracked error size, offering a cheap confidence signal.
Catalog record
| CREATOR | Shuichiro Ozawa, Izumi Takahara, Teruyasu Mizoguchi |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 37 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.03515