Machine Learning3 min reading time
Understanding Alignment in Multimodal LLMs: A Comprehensive Study
Apple Research Blog
Read full postResearchers conducted a detailed study on preference alignment in Multimodal Large Language Models (MLLMs), analyzing alignment algorithms and datasets. They introduced Bias-Driven Hallucination Sampling (BDHS), a new method to create multimodal preference data without extra annotation, achieving competitive results.



