SkillGym Turns Human Agent Skills into Training Environments
SkillGym: Internalizing Human Skills into LLMs for Real-World Problem Solving
The key finding is that skills can be internalized as model weights rather than injected at inference time, since the fine-tuned agent beats skill-assisted base models even without the skills provided. Agent engineers and fine-tuning practitioners can reuse the released environments and trajectories for SFT or RL with outcome-based rewards. The result also shows a 35B open Qwen-based model outscoring much larger closed models on SkillsBench, a useful cost comparison for teams building coding and tool-use agents.
SkillGym is a framework that converts human-written agent skills into executable, verifiable training environments for LLM agents. Its pipeline creates tasks, verifies outcomes with code-based checkers, and measures skill dependence through contrastive executions. The authors release 2,756 environments across 12 categories plus 8,364 successful trajectories averaging 49 tool calls and over 60k logged text tokens. Under Claude Code, supervised fine-tuning lifts Qwen3.5-35B-A3B by 199 Elo on GDPval-AA v2 and 19.10 percentage points on Terminal-Bench 2.1. The resulting 35B SkillGym-Agent reaches 51.47% on skill-assisted SkillsBench v1.1, exceeding reported scores of Claude Sonnet 4.6, GPT-5.4 Mini and DeepSeek V4 Pro.
SkillGym: Internalizing Human Skills into LLMs for Real-World Problem Solving
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.
