Hydra: Unifying Document Retrieval and Generation in a Single Vision-Language Model

Athos Georgiou · 2026

Read Hydra: Unifying Document Retrieval and Generation in a Single Vision-Language Model by Athos Georgiou (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-2603-28554.txt.

Description

Georgiou presents Hydra, a single vision-language model for document search and text generation. Using two models for those tasks doubles memory and system complexity. Hydra toggles one LoRA adapter that was trained only for retrieval. With the adapter on, it makes multi-vector embeddings for search. With it off, generation matches the base Qwen model. The design cuts peak GPU memory by about 63 percent at the 4B scale. Georgiou also releases a smaller 0.8B version and tests the idea on another backbone.

Catalog record

CREATORAthos Georgiou
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42033
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.28554

Browse the full library · About the Tower of Babel