AgentIR: Reasoning-Aware Retrieval for Deep Research Agents

Zijian Chen, Xueguang Ma, Shengyao Zhuang, Jimmy Lin, Akari Asai, Victor Zhong, [1ex] University of Waterloo · 2026

Read AgentIR: Reasoning-Aware Retrieval for Deep Research Agents by Zijian Chen, Xueguang Ma, Shengyao Zhuang, Jimmy Lin, Akari Asai, Victor Zhong, [1ex] University of Waterloo (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-04384.txt.

Description

Chen and colleagues exploit agent reasoning traces to improve retrieval for deep research agents. Unlike human users, such agents write explicit reasoning before each search, a signal retrievers ignore. Reasoning-Aware Retrieval embeds that trace jointly with the query, and DR-Synth builds training data from standard question answering sets. Each part helps alone, and together they train AgentIR-4B. On BrowseComp-Plus, AgentIR-4B reaches 68 percent accuracy with Tongyi-DeepResearch, against 50 percent for conventional models twice its size and 37 percent for BM25.

Catalog record

CREATORZijian Chen, Xueguang Ma, Shengyao Zhuang, Jimmy Lin, Akari Asai, Victor Zhong, [1ex] University of Waterloo
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE55581
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.04384

Browse the full library · About the Tower of Babel