Apple researchers have argued that performance benchmarks may overstate “fragile” AI reasoning capabilities.