In agentic search, test-time scaling typically increases depth (i.e., more turns and tokens per trajectory) or breadth (i.e., more parallel rollouts). This study focuses on breadth scaling, showing that standard parallel sampling yields diminishing returns due to query redundancy at the first turn.
When models issue similar first queries across rollouts, the threads retrieve overlapping evidence, conditioning subsequent turns on this shared retrieval. We address this limitation with DivInit, a training-free intervention at the first turn.
Instead of sampling k independent first queries, DivInit draws n candidates from a single call and picks k. This approach avoids query redundancy, enhancing the diversity and efficiency of the search.
Blogger's Review: The DivInit method proposed in this paper effectively addresses the limitations of traditional parallel sampling, enhancing the performance of agentic search through diverse query initialization. It opens up new avenues for future research.