The big picture

Most benchmarks score literal correctness. XYBench targets a gap: whether models understand user intent in ambiguous or multi-step scenarios.