Researchers from MIT and the MIT-IBM Watson AI Lab have developed a technique to estimate the reliability of foundation models, which are massive deep-learning models pretrained on a large amount of unlabeled data. These models, such as ChatGPT and DALL-E, can sometimes provide incorrect or misleading information, leading to potentially serious consequences. The researchers trained a set of slightly different foundation models and used an algorithm to assess the consistency of the representations each model learned about the same test data point.

By comparing their technique with baseline methods, they found that it better captured the reliability of foundation models across various classification tasks. This technique could be useful in determining whether a model should be applied in a specific setting, even without testing it on a real-world dataset, which can be particularly beneficial in privacy-sensitive environments like healthcare. The researchers also mentioned the possibility of ranking models based on reliability scores, allowing users to choose the most suitable one for their task.