In Recurrent Neural Networks, the primary issue that limits their efficiency and parallelization during training is the sequential generation of ____, which impacts how information flows through the network.

The bottleneck in training RNNs is the sequential generation of hidden states, because each hidden state depends on the previous one. This dependency forces the network to process one time step after another, preventing simultaneous computation of many steps. As a result, the training loop must wait for the previous state before computing the next, which slows down the process and limits GPU parallelization. For example, if a sequence has ten steps, the network must compute step 1, then step 2, and so on, instead of computing all steps in parallel. This sequential nature is why RNNs are less efficient than feed‑forward networks.