SkillGym Framework Internalizes Human Workflows into LLM Training
September 24, 2026
SkillGym converts human-written skills into 2,756 verifiable training environments for agentic reinforcement learning. Supervised fine-tuning on this data improved Qwen3.5-35B-A3B performance by 199 Elo on the GDPval-AA v2 benchmark.
HOW THIS AFFECTS YOU
●
builderThis provides a path to move beyond instruction-following toward internalized agentic capabilities.
●
researcherYou can use this framework to transform unstructured workflows into verifiable RL environments.