Research2h ago
Apple Studies Multimodal LLM Alignment
Researchers at Apple analyzed how preference alignment reduces hallucinations in multimodal models. These systems often produce responses that contradict visual evidence. The study focuses on forcing models to prioritize image content over internal linguistic biases. This provides a technical framework for developers to improve factual consistency in vision-language tasks.