SkillGym trains large language models to master reusable real-world skills
SkillGym: Internalizing Human Skills into LLMs for Real-World Problem Solving
Computation and Language
Summary
Many programs have step-by-step instructions written by people to solve everyday problems, but they usually only follow these steps once instead of really learning them. The authors created SkillGym, a system that turns these instructions into training games where large language models practice and verify their abilities repeatedly. They made thousands of these training games and collected many example solutions, which helped improve the models when fine-tuned. This approach made the models better at using skills to solve tasks and also improved their general problem-solving abilities without relying on skills.
What this means in practice
- •For ai model developers: Create training environments that allow language models to internalize and reuse complex skill workflows for improved problem solving.
- •For software automation engineers: Build language model agents that can reliably execute and verify multi-step procedural tasks based on human-written instructions.
- •For business process architects: Use skill-internalized AI agents to automate and optimize real-world workflows with verifiable outcomes.$Commercial implications: This paper enables AI-powered workflow automation tools that can be sold to businesses for reliable and scalable task execution.
Authors
Zhilong Ge, Yuting Shao, Yutao Yang, Yuxuan Cai, Jie Zhou, Kai Chen, Bo Zhang, Qin Chen, Liang He
Abstract
Human-written agent skills encode rich workflows for real-world problem solving, but are typically used as external inference-time instructions rather than internalized as reusable model capabilities. We introduce \texttt{SkillGym}, a framework that transforms these skills into executable, verifiable training environments for large language model agents. Its skill-to-task pipeline instantiates concrete tasks, verifies outcomes with code-based checkers, and assesses empirical skill dependence through contrastive executions. We construct and release 2,756 environments across 12 categories and collect 8,364 successful trajectories from multiple models and harnesses, averaging 49 tool calls and over 60k logged text tokens. These resources support supervised fine-tuning on verified workflows and reinforcement learning with outcome-based rewards. Under Claude Code, supervised fine-tuning improves Qwen3.5-35B-A3B by 199 Elo on GDPval-AA v2, 19.10 percentage points on Terminal-Bench 2.1, and 28.13 and 12.38 points on SkillsBench v1.1 with and without skills, respectively. Our 35B \texttt{SkillGym-Agent} reaches 51.47\% on skill-assisted SkillsBench, exceeding reported scores for Claude Sonnet 4.6, GPT-5.4 Mini, and DeepSeek V4 Pro. Without skills, it also surpasses skill-assisted bases under Codex and Claude Code, suggesting reusable procedural competence.