finite impulse response

In a finite input response (FIR) system observations or events far enough in the past have no effect on the current behaviour of the system. Typically this is true of systems that use a fixed window over the data of, say, the last N items and base responses only on these. However, this may also be less obvious. Some statistical time series analysis techniques are FIR such as a moving average as well as window-based neural networks.

FIR is in contrast to infinite impulse response (IIR) systems where there is no limit on how far backward data influences persist.

Note that while most LLMs use a fixed window, so are strictly FIR, the window is often larger than a complete data sequence, so they is therefore effecitively IIR.