LSTM细胞状态更新是否考虑当前输入？兼论文与博客公式差异问询

架构老林

2026-5-22

Answers to Your LSTM Questions

1. Does LSTM cell update consider the current input?

Absolutely! Every part of an LSTM's step-wise update—including the input gate, forget gate, output gate, and core cell state update—directly incorporates the current input $x_t$.

To break it down:

All gate units (input gate $i_t$, forget gate $f_t$, output gate $o_t$) are calculated using both the previous hidden state $h_{t-1}$ and the current input $x_t$.
The new candidate cell state $\tilde{C}_t$ is also generated from these two variables.
The final cell state $C_t$ combines the old state (regulated by the forget gate) and the new candidate state (regulated by the input gate)—neither component would work without the current input's contribution.

2. Discrepancy between blog formula and original LSTM paper for $C_t$

This kind of mismatch is actually pretty common in educational content. Many blog posts simplify LSTM explanations for clarity, which can lead to minor omissions or streamlined formulations that don't perfectly mirror the original paper.

In the original LSTM paper, the cell state update does include the current input: the candidate state $\tilde{C}t$ (which feeds into $C_t$) relies on both $x_t$ and $h{t-1}$, and the input gate $i_t$ (which controls how much of $\tilde{C}_t$ is added) also uses $x_t$.

If the blog's formula for $C_t$ only mentions $h_{t-1}$, it's almost certainly a simplification: the author might have omitted explicit reference to $x_t$ because they assumed readers understand that the gate units and candidate state already include the current input, or they streamlined the explanation to focus on the core state transition logic.

The standard, accurate cell state update formula is: