O1 Embedder: Let Retrievers Think Before Action

Ruiran Yan, Zheng Liu, Defu Lian · 2025

Read O1 Embedder: Let Retrievers Think Before Action by Ruiran Yan, Zheng Liu, Defu Lian (2025) 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-2502-07555.txt.

Description

Search engines built on language models fetch first and think later, and one design flips that order. Ruiran Yan, Zheng Liu, and Defu Lian present O1 Embedder in a 2025 paper. Their retriever writes out useful thoughts about the query before hunting documents. Two hurdles had to fall: making training data for such thoughts, and teaching a model to think and fetch together. An expert model drafted the thoughts, and a committee of retrievers refined them.

Joint fine-tuning then taught both skills in one model. Targets include multi-task search, unseen-task search, and queries needing heavy reasoning. The name nods to well-known reasoning models that think before answering. The authors claim the pause improves the fetch.

Catalog record

CREATORRuiran Yan, Zheng Liu, Defu Lian
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE54 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-07555

Browse the full library · About the Tower of Babel