SciML Agents: Write the Solver, Not the Solution
Saarth Gaonkar, Xiang Zheng, Haocheng Xi, Rishabh Tiwari, Kurt Keutzer, Dmitriy Morozov, Michael W. Mahoney, Amir Gholami · 2025
Read SciML Agents: Write the Solver, Not the Solution by Saarth Gaonkar, Xiang Zheng, Haocheng Xi, Rishabh Tiwari, Kurt Keutzer, Dmitriy Morozov, Michael W. Mahoney, Amir Gholami (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-2509-09936.txt.
Description
Saarth Gaonkar and 7 co-authors test a different use of language models in a 2025 paper: writing the solver instead of the solution. Much scientific machine learning trains networks to predict answers directly, which can be fragile. Their SciML agents instead read a plain-language description of an ordinary differential equation and write code that calls proven numerical methods, choosing fit solvers and checking stability.
To measure this skill they built two test sets, one of tricky problems that look stiff but are not, and a benchmark of a thousand varied equation tasks. Newer instruction-following models, given guided prompts with domain knowledge, wrote code that ran and matched reference answers well. Older or smaller models still gained from fine-tuning.
Catalog record
| CREATOR | Saarth Gaonkar, Xiang Zheng, Haocheng Xi, Rishabh Tiwari, Kurt Keutzer, Dmitriy Morozov, Michael W. Mahoney, Amir Gholami |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-09936