SkillGraph: Graph Foundation Priors for LLM Agent Tool Sequence Recommendation

2026-04-24 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

SkillGraph is a novel framework designed to improve tool selection and ordering for LLM agents interacting with large API libraries. It addresses the limitations of semantic similarity methods, which often fail to account for inter-tool data dependencies, leading to negative Kendall-τ scores in structured workflows. SkillGraph introduces a directed weighted execution-transition graph, a "graph foundation prior," derived from 49,831 successful LLM agent trajectories. This prior encodes workflow-precedence regularities. The framework employs a two-stage decoupled approach: GS-Hybrid retrieval for candidate tool selection and a learned pairwise reranker for optimal ordering. On the ToolBench dataset (9,965 test instances, ~16,000 tools), SkillGraph achieved a Set-F1 of 0.271 and a Kendall-τ of 0.096. On API-Bank, it significantly improved Kendall-τ from -0.433 to +0.613, outperforming LLaMA-3.1-8B Stage-2 rerankers under identical Stage-1 inputs.

Key takeaway

For NLP engineers developing LLM agents that interact with extensive API libraries, SkillGraph offers a robust method to enhance tool selection and sequencing accuracy. Your current semantic-only approaches may be insufficient for complex workflows, leading to suboptimal agent performance. Consider integrating graph-based priors derived from successful agent trajectories to capture critical inter-tool dependencies, significantly improving tool ordering and overall agent efficacy.

Key insights

SkillGraph uses a graph-based prior from agent trajectories to improve LLM tool sequence recommendation beyond semantic similarity.

Principles

• Tool ordering requires inter-tool data dependency awareness.
• Workflow regularities can be encoded as a graph prior.

Method

SkillGraph uses a two-stage decoupled framework: GS-Hybrid retrieval for candidate selection and a learned pairwise reranker for ordering, built on a graph foundation prior.

In practice

• Mine agent trajectories to build workflow priors.
• Decouple tool retrieval from ordering for better performance.

Topics

• LLM Agent Tooling
• Tool Sequence Recommendation
• SkillGraph
• Graph Foundation Priors
• Workflow Precedence

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

Related on AIssential

• Compositional Skill Routing for LLM Agents: Decompose, Retrieve, and Compose — Compositional skill routing for LLM agents requires decomposition, retrieval, and plan composition.
• GraphFlow: A Graph-Based Workflow Management for Efficient LLM-Agent Serving — GraphFlow uses a unified graph (wGraph) to dynamically manage LLM-agent workflows, improving performance and memory efficiency.
• How AI Agents Navigate Massive Skill Graphs — Skill-DAG enables LLMs to dynamically interpret and evolve a typed skill graph for scalable, context-aware skill selection, moving beyond semantic similarity.
• SING: Synthetic Intention Graph for Scalable Active Tool Discovery in LLM Agents — Intention-aware graph structures enable accurate, context-efficient tool discovery for LLM agents in large ecosystems.
• Integrating Graphs, Large Language Models, and Agents: Reasoning and Retrieval — Integrating graphs with LLMs enhances reasoning, retrieval, and decision-making by combining structured knowledge with semantic understanding.
• A Framework for Evaluating Agentic Skills at Scale — A new framework rigorously evaluates LLM agent skills at scale, revealing varied model adherence and significant behavioral changes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.