LLMs may be overthinking Word-in-Context benchmarks

Explicit sense inventories improve results and reveal annotation boundaries hiding inside apparent model errors.