Beyond Parallel Sampling: Diverse Query Initialization for Agentic Search

A new method called DivInit improves the efficiency of agentic search systems by generating more diverse initial queries, addressing a key bottleneck in parallel sampling approaches, according to research submitted to arXiv on June 15. Standard techniques for scaling agentic search often increase the number of parallel rollouts, but this yields diminishing returns because models tend to issue similar first queries across different threads, leading to overlapping evidence retrieval [1]. Sidhaarth Sredharan and colleagues propose DivInit, a training-free intervention that operates at the first turn of a search trajectory. Instead of sampling k independent first queries, DivInit draws n candidates from a single call, selects k < n diverse seeds, and runs them as parallel trajectories [1]. Across five open-weight models and eight benchmarks, DivInit consistently outperformed standard parallel sampling, with average gains of five to seven points on multi-hop question-answering tasks at matched compute [1]. The approach does not require additional training, making it a lightweight modification to existing search pipelines. The submission file was 652 KB [1]. The work targets a specific failure mode in test-time scaling. When initial queries lack diversity, subsequent reasoning steps are conditioned on a narrow set of retrieved documents, limiting the system's ability to synthesize information from disparate sources [1]. This challenge echoes broader issues in data analysis where the veracity, or reliability, of information is a critical dimension alongside volume and variety [3]. In big data contexts, ensuring data quality is essential to avoid costs that exceed the value extracted [3]. Agentic search systems are evaluated using standardized benchmarks that measure performance on tasks such as question answering and reasoning [6]. These benchmarks provide datasets and metrics that allow researchers to compare model capabilities in a controlled manner [6]. The gains reported for DivInit were measured on such multi-hop QA benchmarks, where a system must connect information from multiple documents to arrive at an answer [1]. The method's focus on query diversity at the first turn addresses a structural limitation in how language models interact with retrieval systems. By explicitly selecting for diverse seeds from a larger candidate pool, DivInit reduces redundancy without increasing the total number of model calls [1]. The code for the method has been made publicly available [1].