DPrivBench: Benchmarking LLMs' Reasoning for Differential Privacy

In three linesDPrivBench is a benchmark evaluating LLMs' ability to reason about differential privacy (DP). It tests whether functions satisfy stated DP guarantees under specified assumptions. Strongest models handle textbook mechanisms well but all struggle with advanced algorithms, revealing substantial gaps in DP reasoning capabilities.

Summary generated by Claude — human-verified