SkillGym: Internalizing Human Skills into LLMs for Real-World Problem Solving

The short version

The problem

AI agents can answer questions and use tools, yet real work often requires following a long sequence of steps, handling setbacks, and confirming that the finished job is actually correct. Reusable skill guides can help, but merely giving an agent a guide does not ensure it learns to carry out the workflow reliably on its own.

The idea

SkillGym turns those guides into practice tasks: for example, a deployment guide becomes a working simulated deployment that an agent must repair and test. Automatic checks judge whether the task was completed, and runs with and without the guide help identify cases where the guide makes a practical difference. The resulting successful work traces can then be used to train the model on whole workflows, not just isolated tool commands.

What it got

The authors created 2,756 practice environments and collected 8,364 successful work traces. After training, their 35-billion-parameter model scored 51.47% on a benchmark with provided skill guides, and improved by 19.10 percentage points on a command-line task benchmark.