SIRA: Single-Query Retrieval Agent Outperforms Multi-Round Agentic Search Across 10 BEIR Benchmarks

arXiv·medium signalXBlueskyLinkedInCopy link

SIRA consolidates multi-round exploratory searches into a single optimized BM25 call by using LLM cognition to predict missing search vocabulary offline and validating terms via corpus statistics. Training-free, interpretable, and efficient, it outperforms both dense retrievers and multi-round agentic baselines on ten BEIR benchmarks and downstream QA tasks.