Geometry optimization is a major cost in quantum-chemical workflows: each optimization step requires one force evaluation, and at the density-functional level that evaluation dominates wall time. A new research paper, Optimizing the Optimizer: Language Models Discover Faster Molecular Relaxation, asks whether a language model can improve on the best existing optimization methods through autoresearch.
The approach uses an agent that rewrites the optimizer itself to minimize force-call counts. Two admission gates constrain the search: one rejects premature stopping, and another rejects improvements that do not generalize to unseen molecules. Starting from Sella, described as the fastest open-source optimizer available, the search produces AutoSella, a family of two optimizers.
According to the paper, both variants deliver consistent force-call reductions relative to Sella across held-out molecular benchmarks and potentials not used during the search. Most notably, at the r2SCAN-3c DFT level, the best variant requires only 40.2–77.2% of Sella's force calls while achieving the same energy reduction, even though the agent used no DFT gradients during the search.
In a comment, one of the authors explained that the search relies on inexpensive GFN2-xTB calculations, with checks against premature stopping and overfitting. The improvements transfer to DFT: AutoSella uses 23–60% fewer force evaluations than Sella across four held-out molecular benchmarks at the r2SCAN-3c level, while maintaining or improving average energy recovery.
The work sits alongside related recent efforts on self-evolving optimizers, cost-aware transfer for evolutionary optimization, and benchmarks for LLM understanding of chemistry, suggesting growing interest in using language models to improve scientific algorithms.