Apple Unveils BDHS, an Alignment Technique to Reduce Multimodal AI Hallucination

Why does AI say things that don't match the image it's looking at?

A paper released by Apple Research on August 3 addresses why multimodal large language models (MLLMs — AI that understands images and text together) sometimes say things that don't match what's in a photo, and how this can be reduced. Text-only language models already suffer from "hallucination," the problem of stating things that aren't factually true, but models that also handle images face an added layer of complexity. On top of getting facts wrong, they can also fabricate content that isn't in the image at all. Apple directly examines "preference alignment," a training method used to reduce this problem.

Alignment, an old challenge still unresolved for multimodal models

Preference alignment is a training method in which, among multiple answers an AI produces, the one humans prefer is selected and the model is gradually nudged toward it. This technique is already widely used in text-only models like ChatGPT, but the paper's starting point is that it has been relatively less studied in multimodal models that also handle images. Various research teams have reported performance improvements using different datasets, base models, and alignment methods (DPO, PPO, etc.), but what actually drove those improvements has remained unclear and conflated. Apple's researchers separated alignment algorithms into two categories — "offline" methods that train on pre-built pairs of answers, and "online" methods that generate answers in real time during training and adjust accordingly — and analyzed the effect of each independently. The result showed that combining both approaches improves performance in certain situations. The paper goes a step further, also examining how the composition of various publicly available multimodal preference datasets actually affects performance.

Zhipu AI Unveils GLM-OCR, an Image-to-Text Model

Training data made without human hands: BDHS

The most notable proposal in this paper is a new data-generation method called "Bias-Driven Hallucination Sampling (BDHS)." Existing approaches often relied on humans manually marking which answer is more accurate, or on deploying a separate external AI model to generate preference data. Both approaches cost time and money. BDHS creates preference data by leveraging the model's own bias, without such additional work, and Apple stated that it achieved performance competitive with previously published alignment techniques across multiple benchmarks.

What does this change?

This research is likely to serve as a practical reference for teams building multimodal AI. Until now, even when improved alignment performance was reported, it was difficult to tell whether the gains came from the dataset or the algorithm. Since Apple has separated these factors and presented comparative results, follow-up researchers now have a basis for judging which combinations offer the best cost-effectiveness. In particular, the BDHS approach, which generates preference data without manual human work, may also offer a useful reference for the open-source community seeking to lower the cost of alignment training.