Benchmark results in publications show it being confused in chat—like settings and answering questions incorrectly