Beyond Parallel Sampling: Diverse Query Initialization for Agentic Search

The paper introduces DivInit, a novel approach for diverse query initialization in agentic search that improves breadth scaling by addressing redundancy in standard parallel sampling. By generating multiple candidate queries from a single model call and selecting diverse seeds for parallel trajectories, DivInit demonstrates significant performance enhancements, achieving average gains of five to seven points on multi-hop question answering benchmarks across five open-weight models. This method provides practitioners with a more efficient way to optimize query diversity, enhancing the effectiveness of AI systems in complex retrieval tasks.