A new preprint from arXiv reveals that running multiple AI agent rollouts with diverse query initialization outperforms the standard approach of simply launching many parallel copies of identical prompts. The research, titled "Beyond Parallel Sampling: Diverse Query Initialization for Agentic Search," demonstrates that breadth scaling—increasing the number of simultaneous agent trajectories—yields diminishing returns when all rollouts follow the same initial instructions. Instead, the study shows that varying how queries are framed, the reasoning frameworks applied, and the prompts presented to agents produces significantly better search outcomes. This finding challenges the prevailing industry practice where teams attempting to improve agentic search performance typically resort to increasing computational resources by running more identical parallel processes, effectively solving the problem through brute-force scaling rather than architectural innovation.
The implications extend beyond theoretical interest into practical system design. Current large language model agents used in everything from code generation to information retrieval systems rely heavily on test-time scaling—allocating extra compute budget during inference to improve output quality. The research indicates this compute could be deployed far more efficiently. Rather than allocating resources to redundant parallel rollouts, teams should invest in generating diverse prompt variations, alternative reasoning pathways, and different query formulations that guide agents toward more comprehensive solution spaces. This approach mirrors human problem-solving, where approaching a challenge from multiple angles typically yields better results than simply spending more time on a single approach. The research quantifies specific performance gaps, showing that diverse initialization strategies consistently outperform standard parallel sampling across multiple benchmark tasks, though exact percentages vary by domain.
The findings carry particular significance for resource-constrained environments where compute budgets remain limited. Organizations deploying agentic systems in legal case retrieval, supply chain optimization, or other specialized domains can now redirect engineering effort from infrastructure scaling toward prompt engineering and query diversification strategies. However, the research also identifies important limitations: the study focuses specifically on breadth scaling and doesn't address interaction effects with depth scaling (increasing tokens per trajectory). Teams implementing these insights will need to determine optimal ratios between diverse query initialization and other scaling dimensions. Future work should explore how these principles apply across different model sizes and whether gains compound when combined with other efficiency techniques, potentially unlocking substantially more capable agent systems without proportional increases in computational overhead.