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

SkillMaster: Toward Autonomous Skill Mastery in LLM Agents

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 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.

Trains LLM agents with RL to propose, update, or keep their own skills via tool calls, credited by counterfactual probe-based utility and optimized with a dual-advantage GRPO variant.

Problem

Skill banks for LLM agents are typically curated by external teachers, fixed rules, or auxiliary modules, so skill quality is not directly optimized and skill edits receive only sparse task-outcome rewards.

Approach

Integrate skill-editing tool calls into the agent's own policy after each episode, score each candidate edit by its counterfactual utility on same-family probe tasks (success and step efficiency), and normalize action and skill-edit advantages separately (DualAdv-GRPO).

Key results

• On ALFWorld, SkillMaster reaches 98.7% overall success versus 89.9% for SkillRL, the strongest baseline.
• On WebShop, SkillMaster achieves 95.0 score and 82.0% success, versus 85.2 and 72.7 for SkillRL.
• On seven search-augmented QA datasets, SkillMaster achieves 48.0% overall versus 47.1% for SkillRL.

Surprising findings

• Appending a post-episode review turn without skill management improves GRPO by only 0.8 points on each benchmark.
• SkillMaster with a fresh or empty initial skill bank remains strong, with restoring the original bank changing ALFWorld success by only 0.3 points.

Limitations

• Relies on predefined task groupings (task families or product categories) for probe selection.
• Probe evaluation adds O(K) extra rollouts per skill mutation, growing with episode length and mutation frequency.
• Evaluated only on two structured benchmarks plus SearchQA; open-ended environments untested.

Experimental setup

Datasets: ALFWorld WebShop SearchQA training questions

Benchmarks: ALFWorld WebShop NQ TriviaQA PopQA HotpotQA 2Wiki MuSiQue Bamboogle

Metrics: success rate score exact match accuracy

Baselines: GPT-4o Gemini-2.5-Pro ReAct Reflexion Mem0 ExpeL MemP RLOO GRPO EvolveR MemRL SkillRL

Builds on: GRPO SkillRL PPO Reflexion

Closest papers by meaning

• SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks (NeurIPS 2026 · similarity 0.91)
• SkillOS: Learning Skill Curation for Self-Evolving 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)
• Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents (NeurIPS 2026 · similarity 0.90)
• SkillOpt: Executive Strategy for Self-Evolving Agent Skills (NeurIPS 2026 · similarity 0.89)
• From Raw Experience to Skill Consumption: A Systematic Study of Model-Generated Agent Skills (NeurIPS 2026 · similarity 0.89)