GRA: Schema-Agnostic Graph Reasoning Agent for Hybrid Knowledge Graphs
A new AI agent, GRA (Graph Reasoning Agent), has been developed to navigate hybrid knowledge graphs using only seven generic tools, analogous to basic file operations. The agent was tested on the UFK-M (Unified Factory Knowledge Model) industrial benchmark, which contains 258 analytical questions with validated SQL-generated answers. GRA outperformed a full-context agent by 5.1 percentage points (88.4% vs. 83.3%) while reading less than a third of the input tokens. The research, available on arXiv (2608.15834), demonstrates that the performance gain is primarily due to selective agentic access rather than graph topology, as shown by a graph-free control. This work highlights the potential of tool-calling LLM agents to generalize across different data structures, offering a more efficient approach to querying complex knowledge graphs.
Key facts
- GRA is a Graph Reasoning Agent for hybrid knowledge graphs.
- It uses seven generic tools for exploration.
- Tested on UFK-M benchmark with 258 analytical questions.
- GRA achieved 88.4% accuracy vs. 83.3% for full-context agent.
- GRA read under a third of the input tokens.
- Performance gain is from selective agentic access, not graph topology.
- Research paper available on arXiv with ID 2608.15834.
- The agent discovers domain-specific information at run time.
Entities
Institutions
- arXiv