HF Paper 'Optimizing the Optimizer' Uses Claude Fable 5.1 and GPT-6 Astra to Cut Molecular-Relaxation Force Calls ~60% at DFT

Researchers from ITMO and Zelinsky Institute report that Claude Fable 5.1 and GPT-6 Astra agents, run in a 'gated autoresearch loop,' evolved Sella 2.5.0 into AutoSella-F — needing only 40.2–77.2% of baseline force calls at the r2SCAN-3c DFT level while matching energy reductions. The agent never evaluated DFT gradients itself; improvements transferred across four held-out molecule datasets and multiple potentials from GFN2-xTB to full DFT.