Accuracy drops to 50% when lstm stacked layers are added

Hello. I’m using the function nn.LSTM for my analysis. My LSTM has 2 layers and the accuracy reaches 90%. When i change the number of layers to 3 the accuracy goes down to 50%. Why does this happen?

if anyone is interested in this, I figured it out! To stack more LSTM I decreased the learning rate by a factor of 10. Training is slower but the accuracy doesn’t decrease.