DivInit Improves Agentic Search by Diversifying Initial Queries

Agentic search systems typically scale by increasing the depth or breadth of their search. When scaling breadth through parallel rollouts, a common issue is that models often generate similar initial queries, leading to redundant information retrieval and diminishing returns. Researchers have introduced DivInit, a training-free intervention designed to overcome this limitation. DivInit works by generating a larger set of candidate first queries from a single model call, then selecting a diverse subset of these to initiate parallel trajectories. Evaluations across multiple open-weight models and benchmarks, particularly in multi-hop Question Answering, demonstrate that DivInit consistently outperforms standard parallel sampling. It achieves average gains of five to seven points at comparable computational cost, by ensuring more varied evidence collection from the outset.