SkillGym: Fine-Tuning on Verified Skill Trajectories Boosts Agent Performance

Shanghai AI Lab's SkillGym fine-tunes models on verified human-written agent skill trajectories, internalizing skills into model capabilities. It boosts Claude Code success rates by 19 points and lets a 35B model outperform several flagship models without needing skill files at inference.

2026-10-06 ~ 2026-10-06 · 2 related posts