AgentIR: Reasoning-Aware Retrieval Model Outperforms 2x Larger Models in Deep Research

hllo_wrld · x · 2026-08-01

AgentIR is a new retrieval paradigm designed specifically for Deep Research agents. Unlike traditional methods that only process the standalone query, it jointly embeds the agent's explicit reasoning trace alongside the query to capture richer intent and context.

To address the lack of existing training data, the researchers introduced DR-Synth, a method to synthesize retriever training data from standard QA datasets. Experiments show that combining both components yields substantial gains: the 4B parameter model AgentIR-4B achieved 68% accuracy on the BrowseComp-Plus benchmark when paired with an open-weight agent, outperforming conventional embedding models twice its size (50%) and BM25 (37%).

Interestingly, the study found that feeding more reasoning history doesn't help. The latest trace already filters out dead ends, making "forgetting" a beneficial feature in this context.

Original post →

AgentIR: Reasoning-Aware Retrieval Model Outperforms 2x Larger Models in Deep Research

More from coding & agent