greedy decoding

Greedy decoding is the simplest rule of all: at each step, look at the probabilities and take the single highest one, ignoring every other option. It is like a hiker who always steps toward the steepest local uphill — fast, decisive, and never second-guessing. Because there is no randomness, the same prompt gives the same output every time, which makes greedy decoding handy for tasks where you want reproducibility, like extracting a fact or running a test.

The catch is that locally best is not globally best. Always grabbing the top token can walk the model into a corner — a phrase that looked great one word at a time but reads flat, generic, or stuck in a loop overall. It also throws away the model's own uncertainty: a near-tie between two good tokens is decided by a hair, with no memory that the choice was close. For open-ended writing, greedy output tends to feel safe and a little lifeless.