Tag
This paper introduces LogiHard, a framework that uses combinatorial hardening to expose compositional failures in frontier LLMs, demonstrating significant accuracy drops in logical reasoning tasks.