Sanjeev Arora

We introduce Ineq-Comp, a benchmark for testing compositional reasoning in formal inequality proving. Simple human-intuitive transformations cause major accuracy drops, showing that current LLM provers lack robust compositional generalization.