Utility-Oriented Visual Evidence Selection for Multimodal Retrieval-Augmented Generation

Weiqing Luo, Zongye Hu, Xiao Wang, Zhiyuan Yu, Haofeng Zhang, Ziyi Huang · 2026

Read Utility-Oriented Visual Evidence Selection for Multimodal Retrieval-Augmented Generation by Weiqing Luo, Zongye Hu, Xiao Wang, Zhiyuan Yu, Haofeng Zhang, Ziyi Huang (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-2605-13277.txt.

Introduction

Luo and colleagues rethink visual evidence choice for multimodal retrieval augmented generation. Current methods rank images by surface similarity, which often misses what actually helps reasoning. They define evidence value as information gain on model output, then estimate it through a latent helpfulness variable with a light surrogate model. Tests on two benchmarks show steady gains over top retrieval baselines at much lower compute cost.

Catalog record

CREATORWeiqing Luo, Zongye Hu, Xiao Wang, Zhiyuan Yu, Haofeng Zhang, Ziyi Huang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE134599
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.13277

Browse the full library · About the Tower of Babel