The study on how generative language models, like those behind ChatGPT, react when presented with text containing false information has allowed researcher Francesco Ortu to be awarded the Artificial Intelligence Prize from the University of Trieste.
This work was developed at the Institute for Research and Technological Innovation (RIT) of Area Science Park were published in the Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics and presented at one of the most important conferences on Computational Linguistics and Artificial Intelligence for Natural Language which was held last August in Bangkok.
The author explained that “Research on interpretability aims to bridge the gap between empirical approaches and our scientific understanding of the inner workings of generative language models (LLMs). So far, most existing research in this area has focused on how models copy or recall factual knowledge. In our study, we analyzed how information propagates within the neural network, identifying the ‘neurons’ that choose whether to promote or suppress false information proposed by the user.”
