Researchers unveil DivInit to boost agentic AI search
DivInit improves AI agent search accuracy by 5-7 points without extra compute...
A team of researchers from institutions including Carnegie Mellon and the University of Lisbon has introduced DivInit, a novel approach to improve agentic search systems that use LLM agents to answer complex queries. Traditional parallel sampling methods often suffer from diminishing returns because the initial queries generated are too similar, leading to redundant evidence retrieval and overlapping search paths.
DivInit addresses this by generating n candidate queries from a single model call and selecting k diverse ones for parallel execution. This training-free intervention consistently improved performance across five open-weight models and eight benchmarks, delivering average accuracy gains of 5-7 points on multi-hop question answering tasks at matched compute levels. The method is available open-source, with code released on GitHub.
- DivInit improves agentic search performance by 5-7 points on multi-hop QA benchmarks without additional compute
- Method generates n candidate queries from one LLM call and selects k diverse ones for parallel execution
- Works across five open-weight models and eight benchmarks, outperforming standard parallel sampling
Why It Matters
This research delivers measurable accuracy gains for AI agents without increasing computational costs, improving real-world search applications.