arXiv AI recent: Beyond Parallel Sampling: Diverse Query Initialization for Agentic Search
Researchers introduced DivInit, a training-free intervention for agentic search, to address the limitation of standard parallel sampling.,Standard parallel sampling yields diminishing ret...
DivInit is a training-free intervention that addresses the limitation of standard parallel sampling in agentic search.,The method involves drawing n candidates from a single call, picking k < n diverse seeds, and running them as parallel trajectories, resulting in improved performance over standa...