SkillGraph: Graph Foundation Prior for LLM Tool Sequence Recommendation
A novel approach known as SkillGraph enhances the selection and sequencing of tools for LLM agents by extracting a directed weighted execution-transition graph from 49,831 successful agent paths. This graph captures the regularities of workflow precedence, serving as a reusable foundational prior and overcoming the shortcomings of semantic-only techniques that overlook inter-tool data dependencies. The framework operates in two stages, employing GS-Hybrid retrieval for selecting candidates and a learned pairwise reranker for arranging them. On ToolBench, which includes 9,965 test cases and approximately 16,000 tools, it records a Set-F1 of 0.271 and a Kendall-τ of 0.096; on API-Bank, Kendall-τ rises from -0.433 to +0.613.
Key facts
- SkillGraph is a directed weighted execution-transition graph mined from 49,831 successful LLM agent trajectories.
- It encodes workflow-precedence regularities as a reusable graph foundation prior.
- The method uses a two-stage decoupled framework: GS-Hybrid retrieval and a learned pairwise reranker.
- On ToolBench (9,965 test instances, ~16,000 tools), it achieves Set-F1 = 0.271 and Kendall-τ = 0.096.
- On API-Bank, Kendall-τ improves from -0.433 to +0.613.
- Existing semantic-only methods can produce negative Kendall-τ in structured workflow domains.
- The graph foundation prior addresses the absence of inter-tool data dependencies in tool descriptions.
- The work is published on arXiv with ID 2604.19793.
Entities
Institutions
- arXiv