HyperSkill preprint reports reasoning gains with frozen model weights

Researchers presenting HyperSkill report that their method beat the best baseline by 13.0 average F1 points across 11 text, vision-language and omnimodal benchmarks without updating the underlying model's weights. The preprint says the method was best or tied on every dataset. HyperSkill improves reasoning by editing a textual library of skills rather than changing model parameters. It checks candidate skills, records how they combine and retrieves selected skills to guide the model's answers. The researchers say their software and data are publicly available. The reported gains come from benchmark evaluations in a preprint.