Preprint reports small model beats larger rivals at generating VASP settings

Researchers report in a preprint that a fine-tuned Qwen3-4B model paired with automated checks outperformed every general-purpose language model they evaluated at generating VASP INCAR files, which specify settings for materials calculations. Their combined system, INCAR-SLM, exceeded GPT-5.4 by 15.55 points on the 100-point INCAR Score in INCARBench. The model learns from reference VASP calculations, while a deterministic post-processor called VASPGuard checks syntax, workflow requirements and material-dependent constraints. Most of the improvement came from fine-tuning. Performance saturated at a few billion parameters after training and post-processing; the reported advantage concerns this specialized settings-generation benchmark.