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 network cannot exploit parallel hardware, which slows training and limits scalability. For example, in a language model that processes a sentence, the hidden state for the third word must wait until the second word’s state is computed, so the three computations cannot run in parallel. This sequential nature is what makes RNNs less efficient than feed‑forward nets that can compute all layers at once.