Knowledge Graphs as Files: An Agent Harness Where Tool Creation and Tool Use Co-Evolve
- Published
- Source
- arXiv
- Paper number
- 1350
- Field
- Agents
- arXiv ID
- 2610.18616
Key points
- The researchers exported the nodes and edges of knowledge graphs wholesale into files and built the DKGI harness, in which agents follow relations with Python code to gather evidence.
- Training-free DKGI beat the strongest published baseline by at least 6.7 Hit@1 points on every graph of the STaRK benchmark.
- Writing Python code over several turns raised overall Hit@1 by 6.9 points compared with using shell commands only.
- The agent wrote 4.0 programs per query on average and reused earlier outputs to revise its code.
- As a limitation, the evaluation covered only the three STaRK knowledge graphs, and scaling to larger graphs is left to future work.
Paper links
External research summaries. These are not HDATF publications or measured product results.