Apple AI Updates: August 4, 2026

1. Apple Study Maps Preference Alignment Techniques for Multimodal LLMs

Apple. Apple ML researchers published a comprehensive study of preference alignment in multimodal large language models, systematically comparing alignment algorithms, preference datasets, and data-construction methods for image understanding. The work introduces Bias-Driven Hallucination Sampling (BDHS), a method for generating multimodal preference data that requires no additional annotation or external models yet remains competitive with prior alignment approaches across benchmarks. The findings target hallucination, where models produce responses inconsistent with image content, and show that combining offline methods like DPO with online approaches can improve performance. Source