SoG lets LLMs navigate knowledge graphs without fine-tuning, beats SOTA

Search-on-Graph (SoG) introduces a novel paradigm for integrating large language models (LLMs) with knowledge graphs (KGs) during knowledge-intensive reasoning tasks. Existing approaches often rely on separate modules to select reasoning paths, using criteria weakly connected to the actual reasoning needed. This leads to incorrect relations or premature pruning. SoG eliminates this disconnect by having the LLM itself—rather than an external component—perform context-aware navigation through the KG. The process follows an observe-think-navigate loop: at each entity, the LLM observes available relational connections, reasons about which path best advances toward answering the question, and then navigates accordingly. This exploits the LLM's inherent reasoning abilities while keeping the full KG structure and history in view.

Evaluated on six knowledge graph question answering (KGQA) benchmarks, SoG outperforms state-of-the-art methods while requiring no task-specific fine-tuning and generalizing effectively across different KG schemas. Accepted to KDD '26 (32nd ACM SIGKDD Conference), the work demonstrates that giving the LLM direct control over path selection yields superior accuracy and robustness. For professionals building knowledge-intensive AI systems, SoG offers a practical way to enhance reasoning quality without incurring the cost and complexity of additional training.