Mapping AI Benchmark Data to Quantitative Risk Estimates Through Expert Elicitation

The literature and multiple experts point to many potential risks from large language models (LLMs), but there are still very few direct measurements of the actual harms posed. AI risk assessment has so far focused on measuring the models' capabilities, but the capabilities of models are only indicators of risk, not measures of risk. Better modeling and quantification of AI risk scenarios can help bridge this disconnect and link the capabilities of LLMs to tangible real-world harm. This paper makes an early contribution to this field by demonstrating how existing AI benchmarks can be used to facilitate the creation of risk estimates. We describe the results of a pilot study in which experts use information from Cybench, an AI benchmark, to generate probability estimates. We show that the methodology seems promising for this purpose, while noting improvements that can be made to further strengthen its application in quantitative AI risk assessment.

The complexity of unsupervised learning of lexicographic preferences

This paper considers the task of learning users' preferences on a combinatorial set of alternatives, as generally used by online configurators, for example. In many settings, only a set of selected alternatives during past interactions is available to the learner. Fargier et al. [2018] propose an approach to learn, in such a setting, a model of the users' preferences that ranks previously chosen alternatives as high as possible; and an algorithm to learn, in this setting, a particular model of preferences: lexicographic preferences trees (LP-trees). In this paper, we study complexity-theoretical problems related to this approach. We give an upper bound on the sample complexity of learning an LP-tree, which is logarithmic in the number of attributes. We also prove that computing the LP tree that minimises the empirical risk can be done in polynomial time when restricted to the class of linear LP-trees.