Generating VASP INCAR files from natural-language requests is a demanding task because the settings are tightly coupled and highly sensitive to the physics of the system. According to a new preprint on arXiv, only large proprietary cloud models have so far come close to handling this reliably, creating a practical dependence on those services.
The preprint, titled "Fine-Tuning Small Language Models for Reliable VASP INCAR Generation," explores whether targeted fine-tuning can bring smaller models up to a similar level of reliability. The authors suggest that compact models, when fine-tuned on the specific structure of INCAR settings, may be able to manage the coupled constraints well enough to serve as a viable alternative.
Because this is a single preprint, the results have not been independently verified, and the abstract does not provide specific benchmark numbers. Still, the direction is notable: making accurate INCAR generation feasible without large cloud models could lower barriers and improve reproducibility in computational materials science.