Two signals, one long path

Two different signals travel through the RNN's step-by-step chain.

During inference, information about animal travels forward inside the hidden state until the RNN reaches it. Each update may weaken or overwrite that information.

During training, a correction travels backward through those same update steps. Repeated multiplication can shrink it before it reaches the earlier weights.

One signal carries information forward. The other carries a correction backward. They are not the same problem.

What they share is the long path. Every extra word adds another update that the signal must cross.