How to achieve the accuracy stated in the original autokeras paper?

[fwu11]

Hello.
I am using autokeras 0.4.0 on Cifar10 dataset with the 3-layer CNN (CnnGenerator only, no ResnetGenerator or DensenetGenerator) as the base model. I tried to run the experiment for 24 hrs on a single Titan V and the best model I got reached the accuracy around 0.82 while in the autokeras paper, the accuracy should be above 0.88. I am wondering if there are any differences between the implementation in the original paper and autokeras version 0.4.0? Would the gap in results between my experiment and the one from the paper caused by different experiment settings? Any suggestions on this would be very helpful. Thank you for the help!