CAMBRIDGE, MASSACHUSETTS — Researchers from MIT, Worcester Polytechnic Institute, and Google introduced a debiasing technique for vision language models called Weighted Rotational DebiasING (WRING) in a paper accepted to the 2026 International Conference on Learning Representations. The method is a post-processing approach that can be applied on the fly to a pre-trained vision language model without requiring retraining.

Vision language models are multi-modal systems that can understand and interpret video, image, and text simultaneously. WRING can be applied to vision language models such as OpenAI's OpenCLIP, though it is currently limited to Contrastive Language-Image Pre-training (CLIP) models, a type of vision language model that connects images to language for search or classification.

Model architecture can contain and amplify bias, negatively influencing AI model performance in real-world settings. In hospitals and clinics, dermatologists use AI models to classify skin lesions as at risk of developing into cancer or as benign, and bias in skin lesion classification models toward certain skin tones can lead to failure to identify a high-risk patient.

The most commonly used debiasing approach for vision language models is projection debiasing, which removes biased information from model embeddings by projecting the subspace out of the representation space of relationships. That approach can amplify and create other biases, leading to what has been called the "Whac-A-Mole dilemma," a concept introduced to AI research in 2023. Projection debiasing, by contrast, is a post-processing technique that operates differently from WRING's rotational method.

"When you do that, you inadvertently squish everything around," said Walter Gerych, an assistant professor of computer science at Worcester Polytechnic Institute and a co-author of the paper. WRING moves coordinates in the model's high-dimensional space responsible for bias to a different angle so the model can no longer distinguish between different groups within a concept.

The technique alters representations in targeted subspaces while preserving the model's other relationships. Researchers found that WRING reduced bias for a target concept without increasing bias in other areas.

"People already spent a lot of resources, a lot of money, training these huge models, and we don't really want to go in and modify something during training because then you have to start from scratch. [WRING is] very efficient. It doesn't require more training of the model and it's minimally invasive," Gerych said.

"Extending this for ChatGPT-style, generative language models, is the reasonable next step for us," he said. The work was supported in part by a National Science Foundation CAREER Award; an AI2050 Award Early Career Fellowship; a Sloan Research Fellow Award; the Gordon and Betty Moore Foundation Award; and the MIT-Google Computing Innovation Award.