SkillGym training reportedly improves AI agent scores even without skill files

A post describing the paper says SkillGym turns human-written skill files into sandboxed tasks and trains a model on runs that pass a checker. In Claude Code, Qwen3.5-35B-A3B reportedly rose from 39.33% to 58.43% on Terminal-Bench 2.1. On SkillsBench, the trained model scored 26.81% without skill files, versus 23.34% for the base model with skills. Loading skills onto the trained model raised its score to 51.47%.