Research map / Multi-agent systems and games / LLM agents
Group of Skills: Group-Structured Skill Retrieval for Agent Skill Libraries
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 GoSkills, an inference-time method that retrieves anchor-centered skill groups from a typed skill graph, expands support groups, bottlenecks them into bounded skill payloads, and renders a role-labeled execution contract, evaluated on SkillsBench and ALFWorld.
Problem
Flat or dependency-bundled skill retrieval for LLM agents leaves the agent to infer entry points, support roles, visible requirements, and failure-avoidance cues under a limited context budget.
Approach
Change the agent-facing retrieval object from a flat list of atomic skills to anchor-centered, role-labeled skill groups expanded via a group graph and rendered as a fixed execution contract (Start/Support/Check/Avoid) with coverage-debt backfill, without changing the agent, payloads, or environment.
Key results
- On SkillsBench under GPT-5.4, GoSkills reaches 48.9 reward vs 36.4 for Graph of Skills and 18.6 for No Skills, with runtime 352.9s vs 603.7s.
- Retrieval gate: GoSkills passes all 40 visible-requirement items per mode with must-hit 1.00 and fewer than four presented skills on average.
- On ALFWorld, GoSkills achieves the highest reward across the six backbones (e.g., 97.9 with Gemini 3 Pro vs 95.0 for Graph of Skills), with smaller margins than on SkillsBench.
Surprising findings
- Contract-matched control applied to Graph of Skills payloads raises reward from 36.4 to 41.2 and cuts runtime, but remains below full GoSkills, indicating prompt formatting alone does not explain the gain.
- GoSkills is not token-minimal; it often uses more input tokens than Graph of Skills while reducing runtime.
Limitations
- Limited to inference-time retrieval; cannot bind arguments, execute skills, or repair after failures; missing capabilities only reported as debt.
- Depends on deterministic schema extraction and skill metadata; may underperform on long setup chains, sparse metadata, or ambiguous/hidden constraints.
- Reward may trail flat structural rendering in some matched cases; efficiency gains are runtime-specific rather than token-minimal.
Experimental setup
Datasets: SkillsBench ALFWorld
Benchmarks: SkillsBench ALFWorld
Metrics: reward input tokens agent-only runtime must-hit rate
Baselines: No Skills Vanilla Skills Vector Skills Graph of Skills
Builds on: Graph of Skills Vector retrieval
Closest papers by meaning
- SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks (NeurIPS 2026 · similarity 0.89)
- SkillOS: Learning Skill Curation for Self-Evolving Agents (NeurIPS 2026 · similarity 0.89)
- Learning Generalizable Skills from Offline Multi-Task Data for Multi-Agent Cooperation (ICLR 2025 · similarity 0.88)
- Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents (NeurIPS 2026 · similarity 0.87)
- SkillMaster: Toward Autonomous Skill Mastery in LLM Agents (NeurIPS 2026 · similarity 0.87)
- Skill-Pro: Learning Reusable Skills from Experience via Non-Parametric PPO for LLM Agents (ICML 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.