SkillGym turns human agent skills into RL training environments
SkillGym: Internalizing Human Skills into LLMs for Real-World Problem Solving
Training on internalized skill workflows lets a 35B open-weights model beat far larger proprietary models on agentic benchmarks, shifting the edge from raw model scale to curated environment data. Builders running Claude Code or Codex harnesses can reuse the released environments and trajectories for their own SFT or RL pipelines. The results give Qwen-based deployments a concrete benchmark point against Claude Sonnet 4.6 and DeepSeek V4 Pro.
SkillGym is a framework that converts human-written agent skills into executable, verifiable training environments for LLM agents.
It ships 2,756 environments across 12 categories and 8,364 successful trajectories, averaging 49 tool calls and over 60k logged text tokens per trajectory.
Fine-tuning Qwen3.5-35B-A3B under Claude Code adds 199 Elo on GDPval-AA v2 and 19.
10 percentage points on Terminal-Bench 2.1.
The resulting 35B SkillGym-Agent scores 51.47% on skill-assisted SkillsBench v1.
1, above reported scores for 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.
This report is categorized under AITOP's "Papers" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include SkillGym, Qwen3.5, agents, reinforcement learning.
Primary source verification: published directly by arXiv: DeepSeek with first-party authority. Cross-source corroboration: 6 independent sources have covered this development (including elvis, 向阳乔木, shao__meng), confirming industry-wide signal strength.
