Skills Trained Into the Model Beat Runtime-Loaded Skills by Up to 10.1 Points
Notes on SKILL0: In-Context Agentic Reinforcement Learning for Skill Internalization (arXiv:2604.02268) — Zhengxi Lu, Zhiyuan Yao, Jinyang Wu, Chengcheng Han, Qi Gu, Xunliang Cai, Weiming Lu, Jun Xiao, Yueting Zhuang, Yongliang Shen · April 2026
Note published · written by SkillFed’s research pipeline from the paper above · how these notes are made
AI-assisted notes · reviewed by SkillFed Skill evolutionSKILL0 trains an agent to stop needing its own skill files. That's skill internalization — procedural knowledge migrates from context into weights instead of getting reloaded every time the agent runs. The recipe starts with reinforcement learning (GRPO) on Qwen2.5-VL backbones, and it renders skill files and interaction history together as a single compact visual context — an image, not a wall of prepended text — which trims token overhead before any curriculum even kicks in. Then comes the actual mechanism: a training schedule that hands every rollout the full skill file at first and slowly takes it away. An offline pass groups skills by category and pairs each with matched validation sub-tasks; a Dynamic Curriculum then rechecks every few steps how much each file is still worth — accuracy with it minus accuracy without — and keeps only the ones still earning their keep, inside a budget that shrinks on a fixed linear schedule until it hits zero and the policy runs fully zero-shot.
Across ALFWorld, Search-QA, and WebShop, the internalized zero-shot policy beats the standard skill-augmented RL baseline (AgentOCR): +9.7 points, +6.6, and +10.1, all while running under 0.5k tokens of context per step — a fraction of what the baseline spends keeping those same files in every prompt. And the curriculum's shape isn't decorative, it's load-bearing: hold the skill budget constant during training and strip it only at test time, and accuracy collapses by more than 13 points. Decay that same budget smoothly to zero across training instead, and the cliff disappears — the model even picks up a small positive transfer.
Key numbers
| ALFWorld success-rate gain over skill-augmented RL baseline | +9.7 pts (87.9% vs. 78.2%) |
| WebShop task-accuracy gain over baseline | +10.1 pts (66.4% vs. 56.3%) |
| Search-QA score gain over baseline | +6.6 pts (40.8% vs. 34.2%) |
| Per-step context cost at inference | under 0.5k tokens/step |
| Accuracy collapse when skills are stripped at test time without a decay curriculum | -13.3 pts |
Skills related to this research
Related notes
- Route skill treatment by task difficulty, not skill category: +13.2 points OOD success →
- Self-tested skill edits beat teacher-curated skill banks by up to 9.3 points →
- One shared reward signal pushes skill-augmented agents to 97.5% success, beating every tested baseline — skill-augmented or skill-free — by up to 20 points →
- Testing a Skill Before It's Banked Lifts WebShop Success From 72.7% to 78.1% →
- The right skill text for one model is the wrong one for another — model-aware rewriting recovers up to 25.8 points →
- An 8B Model Beats 4 Frontier LLMs by 25%+ — By Mining Its Own Skill Bank →
- 1,184 skills fell in one campaign — because Agent Skills has no line between data and instructions →
- Debug Logging Alone Causes 73.5% of Agent-Skill Credential Leaks →
- A skill's own examples smuggle malicious code past agent defenses up to 33.5% of the time →
References
- Lu, Z. et al. (2026). SKILL0: In-Context Agentic Reinforcement Learning for Skill Internalization. arXiv:2604.02268.
- Yao, S. et al. (2022). ReAct: Synergizing Reasoning and Acting in Language Models. arXiv:2210.03629.
- Shinn, N. et al. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. arXiv:2303.11366.
- Shao, Z. et al. (2024). DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv:2402.03300.