A new arXiv paper studies why parallel sampling in agentic search can hit diminishing returns. The authors find that rollouts often begin with similar first queries, causing agents to retrieve overlapping evidence.
Their proposed method, DivInit, samples multiple candidate first queries in one call, selects diverse seeds, and runs those as parallel trajectories. The goal is to make breadth scaling less redundant.
The work is relevant for research agents and search-heavy workflows where simply running more parallel attempts can waste tokens without improving evidence coverage.