Verified workflow training gives language models reusable procedural skills

SkillGym converts human-written agent instructions into checked training tasks, enabling models to retain workflows even when the original skills are withheld.

Academic

Zhilong Ge · Yuting Shao · Yutao Yang · Yuxuan Cai · Jie Zhou · Kai Chen · +3 more

East China Normal University · Shanghai AI Laboratory

Research Digest·24 September 2026·2 min read

Thread:Skill Selection for Agents

In one line

SkillGym converts human-written agent skills into executable training environments, enabling LLMs to internalize procedural competence and beat larger models on real-world benchmarks.

Before you read this

Three ideas this paper builds on

agent skillsLearned capabilities that enable AI agents to perform specialized tasks with improved effectiveness.118 papers →verificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.328 papers →reinforcement learningA machine learning method where an agent learns to make sequences of decisions by receiving rewards or penalties for its actions.287 papers →

Read it as

SkillGym is a framework that turns human-written , which are normally used as external instructions, into executable training environments. It does this by instantiating concrete tasks from skill descriptions, verifying outcomes with code-based checkers, and identifying tasks that truly require the skill through contrastive runs.

The authors released 2,756 environments across 12 categories and collected 8,364 successful trajectories. Fine-tuning a 35B model on this data improved its performance on general agent benchmarks (e.g., +199 Elo on GDPval-AA v2) and on skills benchmarks, allowing it to surpass the reported scores of much larger models like Claude Sonnet 4.6 and GPT-5.4 Mini.

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

agent skills verification skill internalization procedural competence supervised fine-tuning reinforcement learning