In order to build more useful AI systems, a natural inclination is to try to make them more agentic . But while agents built from language models are touted as the next big advance [Wang et al., 2024],

In general, interventional data, i.e., data collected after a perturbation in

Current P AC-Bayes generalisation bounds are restricted to scalar metrics of performance, such as the loss or error rate. However, one ideally wants more information-rich certificates that control the entire distribution of possible outcomes, such as the distribution of the test loss in regression, or the probabilities of different mis-classifications.