DPrivBench Benchmark Tests LLMs' Ability to Reason About Differential Privacy Algorithms

A new benchmark called DPrivBench evaluates whether large language models can automate reasoning about differential privacy, a technique for protecting data privacy. Differential privacy requires expert-level knowledge to design and verify algorithms, creating barriers for non-specialists. Previous approaches have depended on specialized verification languages needing substantial domain expertise or remained semi-automated with human guidance. The benchmark presents instances asking whether a function or algorithm satisfies a stated differential privacy guarantee under specific assumptions. It covers a broad range of differential privacy topics and spans diverse difficulty levels while resisting shortcut reasoning through trivial pattern matching. Experiments reveal that while the strongest models handle textbook mechanisms adequately, all models struggle with advanced algorithms. The work investigates the potential for LLMs to lower the high barrier faced by practitioners lacking expertise in this complex field. The research was announced on arXiv with the identifier 2604.15851v1.