In tests, all baselines achieved no more than 20% accuracy, while table-augmented generation did far better with 40% or better accuracy.