Greedy Decoding — Always Pick the Most Likely Token

Why it matters

Greedy is the baseline. Knowing its behavior explains why sampling and beam search exist and when to use them.

The architecture

At each step: compute logits, pick argmax, generate that token.

Equivalent to temperature=0 or top-k=1.

How it works end to end

Determinism: same prompt → same output. Useful for reproducibility, testing.

Speed: no sampling overhead. Fastest possible generation.

Limitations: can miss globally better sequences; prone to repetition.