MM-instruct: Generated visual instructions for large multimodal model alignment

doi:10.1016/j.neucom.2025.131164

MM-instruct: Generated visual instructions for large multimodal model alignment

2025-08-05

0

PRE

AI

DOI：10.1016/j.neucom.2025.131164

Original

Original request for help

Save

Abstract

En 中文

• Introduces MM-Instruct: An automated pipeline generating diverse visual instruction data to improve LMM performance on real-world tasks beyond simple QA. • Demonstrates Improved Instruction Following: Models trained with MM-Instruct data significantly outperform prior models (e.g., LLaVA-1.5) in executing diverse visual instructions. • Releases Large-Scale Dataset: Contributes a new dataset with 293 diverse instructions and 234 k high-quality visual instruction pairs to advance LMM instruction-following research.

Keywords:

MM-Instruct

visual instruction data

large language models

instruction following

multimodal learning

MM-instruct: Generated visual instructions for large multimodal model alignment

Abstract

Journal

Organization