CoEvoKG: Co-Evolving Knowledge Graphs with Self-Evolving Search Agents

The introduction of CoEvoKG marks a significant advancement in enhancing large language models (LLMs) for search agent applications by utilizing knowledge graphs as a reliable source for training tasks and a lasting memory for evidence. This framework, outlined in an arXiv paper (2608.01904), simultaneously trains a search agent and a task generator. The task generator formulates multihop questions derived from entity chains in the knowledge graph, while the search agent is rewarded based on the accuracy of answers and search paths validated by graph evidence. Upon a successful search, CoEvoKG verifies and eliminates duplicate evidence, reintegrating it into the relevant graph nodes and edges. This enriched graph is subsequently employed in future task generation and reward assessment, facilitating a continuous cycle of model self-evolution and knowledge enhancement. CoEvoKG addresses the shortcomings of current self-play agents that often overlook knowledge acquired from successful searches. By treating the knowledge graph as an evolving memory, it seeks to boost the performance and efficiency of LLM-driven search agents over time.