Revisiting Text Ranking in Deep Research

A team of researchers from the University of Glasgow and University of Illinois Urbana-Champaign has published a comprehensive analysis titled 'Revisiting Text Ranking in Deep Research' that systematically evaluates how different information retrieval methods perform when used by AI agents conducting deep research. The study addresses a critical gap: while most LLM-based research agents rely on opaque web search APIs, this research opens the black box to analyze specific retrieval components. Using the BrowseComp-Plus dataset with a fixed corpus, the researchers examined how agent-generated queries interact with various text ranking methods, providing the first systematic comparison in this emerging field.

The research tested 2 open-source agents, 5 different retrievers (including lexical, learned sparse, and multi-vector approaches), and 3 re-rankers across diverse configurations. Key technical findings reveal that agent-issued queries typically follow web-search syntax (like quoted exact matches), which favors certain retrieval methods. Passage-level retrieval units proved more efficient under limited context windows and avoided document length normalization issues. Most significantly, the study found that translating agent-issued queries into natural-language questions dramatically improves performance by bridging the query mismatch problem. These findings provide concrete, actionable guidelines for developers building the next generation of AI research assistants and autonomous agents.