Values and Generative AI
If we define values as preferences over the states of the world, and if we can block out the screams of the researchers who devoted their lives to precise and scientifically based value taxonomies, then we will inevitably start wondering how to model and work with these preferences.
This line of research explores how large language models, thanks to their embedded common knowledge and their reasoning abilities, can help tackle this problem.
PUBLICATION
Values and Diffusion Models
How good are diffusion models at capturing human values? Starting from AI-generated images, this study checks whether people can recognise the values represented, and explores how quickly they got to the answer, their confidence, and other aspects, including whether they experienced an “aha!” moment when they discovered the real value represented. Recognition rate among participants was low but the images generated highlight the bias of these models.
PUBLICATION
Values and LLMs
Presented in October 2023, this is the first work on assessing value alignment of generative AI. Through a crowdsourced evaluation, we determined that the LLMs’ responses are as value-aligned as humans’ and they are often preferred. They are, however, biased towards Western values.
PUBLICATION
Values and VLMs
This work extend the evaluation of LLMs alignment by analysing the responses of models with image input. Compared with humans, the models focus on other aspects, so they are not aligned. However, with proper prompting their answers can be more nuanced than the humans and thus they could be used to reason on values.
PUBLICATION
Controlling LLMs Values
This paper introduces a method to measure how well a candidate prompt can steer large language models’ responses toward specific human values, without fine-tuning the model. This is extremely important in dynamic conversational contexts where values, preferences, and interlocutors can change frequently. We apply this method on a dialogue dataset using Schwartz’s theory, and we show that a simple prompt asking to maximise for a specific value can effectively increase its presence in the generated text.
PUBLICATION
Reasoning on Values
In this work, we show that Large Language Models can go beyond just detecting human values in text and images: they can use these to take action, while at the same time reasoning on the preferences, comfort and safety of the people and pets involved… all of this in real-time! As an example, we observed how an LLM would control a simple vacuum cleaning robot.