I have read parts of the Med-Gemini paper and I was intrigued by the uncertainty-guided search technique used by Med-Gemini models to improve the accuracy of their answers. In essence, It involves an iterative process of generating multiple reasoning paths (chains of thought) and filtering them based on uncertainty. This is something that I would love to have in the base Gemini family models. When Gemini returns an answer it would be great to have an uncertainty score, so that we know how much we can rely on the answer.
Furthermore, it would be great to have with each token a confidence/uncertainty score. If i ask the model to generate a JSON, I am less interested in the uncertainty score of the keys and more on the uncertainty score of each individual value.
It would be great to have any thoughts by the teams at google on this topic.
Thank you
LinkedIn
Twitter
