SkillDAG: Self-Evolving Typed Skill Graphs for LLM Skill Selection at Scale

2026-06-02 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

SkillDAG is a novel system addressing the challenge of selecting the right skills for LLM agents from large libraries, moving beyond simple similarity matching to a structural problem. It models inter-skill relationships as a typed directed graph, providing an inference-time, agent-callable structural retrieval interface. This system allows agents to query and evolve the graph during execution using a propose-then-commit protocol, accumulating structure across episodes. On ALFWorld and SkillsBench with MiniMax-M2.7, SkillDAG achieved 67.1% success and 27.3% reward, outperforming the strongest Graph-of-Skills baseline by +12.8 and +8.6 points, respectively. The advantage extends to gpt-5.2-codex, with intrinsic SkillsBench Ret@K increasing from 65.5 to 78.2 under matched queries. These improvements stem from robust candidate ranking and set-monotone online edits.

Key takeaway

For Machine Learning Engineers building LLM agents with extensive skill libraries, consider implementing dynamic, graph-based skill selection. Your agent's performance can significantly improve by modeling inter-skill relationships structurally and allowing the agent to evolve this graph during execution. This approach, exemplified by SkillDAG's gains on MiniMax-M2.7 and gpt-5.2-codex, offers a path to more robust and adaptable agent behavior, especially as skill pools grow.

Key insights

Dynamic, graph-based structural retrieval significantly enhances LLM agent skill selection and performance over static methods.

Principles

• Inter-skill relationships are structural, not merely semantic
• Dynamic graph evolution improves retrieval recall and adaptability
• Candidate ranking robustness is crucial for large skill pools

Method

SkillDAG models inter-skill relationships as a typed directed graph, exposing an agent-callable structural retrieval interface that queries and evolves during execution via a propose-then-commit protocol.

In practice

• Implement typed directed graphs for LLM skill management
• Enable agents to register execution-backed edges for graph evolution
• Prioritize robust candidate ranking for expanding skill libraries

Topics

• SkillDAG
• LLM Agents
• Skill Selection
• Typed Skill Graphs
• Structural Retrieval
• Graph Evolution

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• SkillGraph: Graph Foundation Priors for LLM Agent Tool Sequence Recommendation — SkillGraph uses a graph-based prior from agent trajectories to improve LLM tool sequence recommendation beyond semantic similarity.
• SkillPyramid: A Hierarchical Skill Consolidation Framework for Self-Evolving Agents — SkillPyramid enables AI agents to dynamically evolve and generalize skills through hierarchical consolidation and self-evolution, improving performance.
• SkillsInjector: Dynamic Skill Context Construction for LLM Agents — Dynamic, adaptive skill context construction significantly enhances LLM agent performance by optimizing skill selection and presentation.
• SkillBrew: Multi-Objective Curation of Skill Banks for LLM Agents — Principled, multi-objective curation of LLM agent skill banks is crucial for efficiency and self-improvement, moving beyond append-only growth.
• SkillMAS: Skill Co-Evolution with LLM-based Multi-Agent System — SkillMAS couples skill evolution and MAS restructuring using verified execution traces for adaptive LLM agent specialization.
• EXG: Self-Evolving Agents with Experience Graphs — EXG uses a structured experience graph to enable LLM agents to systematically learn and improve from past interactions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.