Imagine an AI assistant describing a photo. It confidently points out details that aren’t there, like a non-existent cat or a red traffic light when it's clearly green. This "hallucination" problem plagues today's Multimodal Large Language Models (MLLMs), which are designed to both understand images and generate text. New research from Apple delves into the intricacies of why these models struggle with visual accuracy, particularly focusing on *alignment*—how well the models’ responses match the actual content of images. One of the root causes, researchers suggest, is that the language model part of an MLLM is often pre-trained separately on massive text datasets, creating inherent biases. When faced with an image, these biases can outweigh the visual evidence, leading to fabricated details. The Apple researchers categorized different alignment techniques into two main groups: offline and online. Offline methods, like Direct Preference Optimization (DPO), use pre-collected data where a 'preferred' (correct) response is paired with a 'rejected' (incorrect) response. Online methods, on the other hand, generate responses dynamically during training, getting feedback from human annotators or other AI models. Interestingly, the researchers found that combining both offline and online approaches can lead to more balanced results. They also discovered that the *quality* of training data matters more than the sheer quantity. In fact, they developed a new, highly efficient technique called Bias-Driven Hallucination Sampling (BDHS). BDHS cleverly triggers the model's biases by selectively masking parts of the image's representation. Then, it guides the model to generate 'rejected' responses that are subtly but meaningfully different from the correct answer. The core innovation lies in the selective attention masking to identify specific weaknesses in the MLLM's understanding. This approach avoids expensive human annotations or external models, and still achieves similar or even better performance compared to other methods, highlighting its potential for more efficient and cost-effective alignment processes. The study sheds light on the challenges of integrating vision and language in AI and introduces new avenues for aligning MLLMs, pushing them closer to our expectations of seeing and reasoning like humans. While we are still in the early days, this type of research paves the way for more trustworthy, visually grounded AI assistants.