Beyond Parallel Sampling: Diverse Query Initialization for Agentic Search

2026-06-15

AI Summary

Standard parallel sampling in agent-based search often suffers from redundant initial queries, leading to overlapping retrieved evidence and diminishing returns from breadth expansion. This work identifies and addresses this issue for the first time by introducing DivInit, a training-free diversity-aware initialization method. DivInit generates multiple candidate queries in the first step and selects diverse seeds to initiate parallel reasoning trajectories, thereby effectively mitigating redundant retrieval. Integrating large language model–based query generation, diversity selection algorithms, and retrieval-augmented reasoning, DivInit consistently improves multi-hop question answering performance by 5–7 percentage points on average across five open-source models and eight benchmarks, without incurring additional computational overhead.

Abstract

Test-time scaling for agentic search typically increases depth (i.e., more turns and tokens per trajectory) or breadth (i.e., more parallel rollouts). Here we focus on breadth scaling, showing that standard parallel sampling yields diminishing returns, tracing this to query redundancy at the first turn. When models issue similar first queries across rollouts, the threads retrieve overlapping evidence, and subsequent turns are conditioned on this shared retrieval. We address this limitation with DivInit, a training-free intervention at the first turn. Rather than sampling k independent first queries, DivInit draws n candidates from a single call, picks k < n diverse seeds, and runs them as parallel trajectories. Across five open-weight models and eight benchmarks, DivInit consistently improves over standard parallel sampling, with average gains of five to seven points on multi-hop QA at matched compute. Code available at https://github.com/cxcscmu/diverse-query-initialization

Problem

• agentic search
• parallel sampling
• query redundancy
• test-time scaling
• diverse initialization

Innovation

• agentic search
• diverse query initialization
• parallel sampling
• test-time scaling
• multi-hop QA