The simplest decoder: at each step, take the single MOST likely next token — the argmax — and never look back. It’s deterministic and fast, and it ignores the whole rest of the snowstorm of probability. That makes it prone to bland, looping text, because the locally-best token isn’t the globally-best sentence. Slide to reshape the distribution; greedy always grabs the tallest bar.
Greedy decoding selects argmaxᵢ pᵢ at every step — the mode of the next-token distribution — with zero randomness, so the same prompt always yields the same output. It is temperature-invariant (scaling logits can’t change which is largest) and throws away all the probability mass it doesn’t pick, which is why it tends toward generic, repetitive continuations and can miss a higher-probability SEQUENCE that starts with a lower-probability token. A fail-loud self-check throws unless the pick is the true maximum and is invariant to temperature. ◆ real decoding, node-verified.
The distribution here is illustrative; the argmax + its temperature-invariance are exact. ‘Bland/looping’ is the well-documented tendency, not a guarantee for every prompt.