1,858 reusable skills mined from 22,511 real chats — zero weight updates
Notes on AutoSkill: Experience-Driven Lifelong Learning via Skill Self-Evolution (arXiv:2603.01145) — Yutao Yang, Junsong Li, Qianjun Pan, Bihao Zhan, Yuxuan Cai, Linge Du, Jie Zhou, Kai Chen, Qin Chen, Xin Li, Bo Zhang, Liang He · March 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 evolutionAutoSkill is a training-free plug-in layer that turns repeated user requests into versioned skill artifacts instead of treating them as disposable chat history. Two loops run in tandem. A serving loop rewrites the incoming query, ranks stored skills with a hybrid BM25+dense retrieval score, and injects only the ones that clear a similarity threshold. A background evolution loop extracts a candidate skill from the user's own queries — never the model's replies — and hands it to an LLM judge that compares it against its single nearest neighbor and decides add, merge, or discard. A merge doesn't concatenate text: it bumps the existing skill's version number and folds in only the new, non-conflicting constraints, so a skill like a writing-style preference accumulates as one evolving artifact rather than a pile of near-duplicate prompts.
Running the extraction pipeline across four WildChat-1M subsets — English and Chinese, GPT-3.5 and GPT-4, each filtered to conversations with more than 8 turns — turns up 1,858 extracted skills from 22,511 conversations. The resulting SkillBank skews toward programming and technical writing: Programming & Software Dev. is the largest category at 482 skills, roughly a quarter of the corpus, and "python" is the single most common tag. It also catches softer behavioral preferences — roleplay personas, translation conventions. Case studies confirm the versioning mechanism works as designed: one English text-rewriting skill reached version 0.1.34 after 34 rounds of incremental refinement from recurring feedback, while a rarely-revisited Chinese counseling-persona skill sat untouched at 0.1.0.
Key numbers
| Skills extracted (4 WildChat subsets) | 1,858 |
| Conversations analyzed (>8 turns each) | 22,511 |
| Largest category share | 482 skills, Programming & Software Dev. (~26%) |
| Deepest single-skill refinement depth | v0.1.34 (34 rounds) |
| Most common skill tag | python — 98 skills |
Skills related to this research
Related research
Related notes
- MUSE-Autoskill's Self-Written Skills Beat Human-Authored Ones, 85.24% vs. 81.17% →
- 215 Skills, 165 Contributors, No Fidelity Test →
- A skill library built by GLM-4.6 boosts GPT-4.1's task success by 10 points →
- Self-tested skill edits beat teacher-curated skill banks by up to 9.3 points →
- Splitting agent memory into skills and experience gains up to 11 points of task success →
- A self-writing skill library lifts Humanity's Last Exam accuracy 116% — with zero weight updates →
- Flat Skill Invocation Craters From 48 to 17 as the Pool Hits 200,000 →
- Skill folders, not weights: +12.1 points on noisy-retrieval QA →
- 200k+ skills wired into one graph, +40% reward across three backbones →
References
- Yang, Li, Pan et al., "AutoSkill: Experience-Driven Lifelong Learning via Skill Self-Evolution," arXiv:2603.01145 (2026)
- Zhao et al., "WildChat: 1M ChatGPT Interaction Logs in the Wild" (2024)
- Packer et al., "MemGPT: Towards LLMs as Operating Systems" (2023)
- Zhong et al., "MemoryBank: Enhancing Large Language Models with Long-Term Memory," AAAI (2024)
- Wang et al., "Voyager: An Open-Ended Embodied Agent with Large Language Models" (2023)