Research map / Multi-agent systems and games / LLM agents

SkillOS: Learning Skill Curation for Self-Evolving Agents

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SkillOS, an RL recipe (GRPO) that trains a Qwen3-8B skill curator to insert/update/delete Markdown skills in a repository for a frozen executor, using grouped related-task streams and composite rewards.

Problem

Self-evolving LLM agents need high-quality skill curation from streaming experience, but manual curation does not scale, heuristic memory operations lack executor feedback, and prior RL approaches give sparse signals for complex operations like update and delete.

Approach

Decouple a trainable skill curator from a frozen agent executor; train the curator with GRPO on grouped sequences of related tasks so that curation decisions are rewarded by downstream success on later tasks, combined with function-call validity, LLM-judged content quality, and repository compactness rewards.

Key results

• On ALFWorld with Qwen3-8B executor, SkillOS reaches 61.2% average success rate vs 55.7% for the strongest baseline ReasoningBank.
• With Gemini-2.5-Pro executor, SkillOS reaches 80.2% average ALFWorld success vs 66.4% for no memory.
• The 8B trained curator outperforms the Gemini-2.5-Pro zero-shot curator (SkillOS-gemini) across ALFWorld, WebShop and reasoning benchmarks.

Surprising findings

• A trained 8B curator outperforms using Gemini-2.5-Pro itself as the curator, suggesting executor-grounded training matters more than curator scale.
• MemP underperforms the No Memory baseline with the Gemini-3.1-Flash-Lite executor (58.6% vs 61.2%).

Limitations

• Skills are restricted to single Markdown files, a simplification of the folder-based Anthropic skill design.
• Gains are smaller on single-turn reasoning tasks than on agentic tasks.
• Task outcome signal relies on LLM-as-a-judge with the frozen executor rather than ground truth for some settings.

Experimental setup

Datasets: ALFWorld WebShop DeepMath-103K

Benchmarks: ALFWorld WebShop AIME24 AIME25 GPQA-Diamond

Metrics: success rate average score steps accuracy

Baselines: No Memory ReasoningBank MemP SkillOS-base SkillOS-gemini

Builds on: GRPO ReAct Chain-of-thought BM25 ReasoningBank MemP

Closest papers by meaning

• SkillMaster: Toward Autonomous Skill Mastery in LLM Agents (NeurIPS 2026 · similarity 0.91)
• Skill-Pro: Learning Reusable Skills from Experience via Non-Parametric PPO for LLM Agents (ICML 2026 · similarity 0.90)
• SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks (NeurIPS 2026 · similarity 0.89)
• MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents (NeurIPS 2026 · similarity 0.89)
• Group of Skills: Group-Structured Skill Retrieval for Agent Skill Libraries (NeurIPS 2026 · similarity 0.89)
• Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents (NeurIPS 2026 · similarity 0.89)