Knowledge-Graph Paths Improve Self-Evolving Search Agents

Researchers propose using knowledge-graph paths as intermediate supervision to enhance self-evolving search agents. The approach addresses two bottlenecks in the Search Self-Play (SSP) framework: invalid question generation by the Proposer and sparse reward signals for the Solver. By grounding question construction in LLM-guided knowledge-graph subgraphs, the method provides relational context, improving question validity. Additionally, knowledge-graph paths offer richer feedback than binary rewards, helping the Solver learn from partially correct trajectories. The work is published on arXiv under ID 2605.05702.