Researchers propose DivInit, a training-free method to improve agentic search by selecting diverse initial queries before running parallel search trajectories. The paper reports that standard parallel sampling suffers from redundant first queries and overlapping retrieved evidence, while DivInit improves performance across five open-weight models and eight benchmarks, including average gains of five to seven points on multi-hop question answering at matched compute.

Researchers propose DivInit, a method to improve agentic search by diversifying initial queries in parallel sampling, avoiding redundant evidence retrieval. The approach boosts performance by five to seven points on multi-hop question-answering benchmarks without additional training. The technique is tested across five open-weight models and eight datasets, offering a compute-efficient alternative to standard parallel sampling.