Training-free method DivInit improves agentic search performance by five to seven points across five open-weight models and eight benchmarks

Agentic search systems often scale breadth by running more parallel rollouts, but standard parallel sampling can suffer from diminishing returns due to query redundancy at the first turn. When models issue similar first queries across rollouts, subsequent turns retrieve overlapping evidence, limiting diversity in downstream reasoning.

The proposed method, DivInit, addresses this by drawing n candidate first queries from a single model call and selecting k < n diverse seeds to initialize parallel trajectories. This approach reduces redundancy at the source while maintaining matched compute compared to standard parallel sampling.

Across five open-weight models and eight benchmarks, DivInit consistently improves performance on multi-hop question answering, with average gains of five to seven points at the same compute budget. The improvements are reported without requiring additional training, making DivInit a practical intervention for existing agentic search systems.

The authors release code and note that the method is under review at EMNLP 2026. The paper includes eight figures and spans 15 pages, positioning DivInit as a lightweight yet effective technique for test-time scaling in agentic search.