Greedy decoding
Also known as: greedy
Greedy decoding picks the highest-probability token at every step. No sampling, no diversity — just the model’s top guess, repeated.
- Fast and deterministic — same input, same output.
- Repetitive — the model tends to loop on its favorite high-probability phrases.
It’s the baseline that sampling strategies improve on.
Use it for: tasks where you want maximum determinism (e.g., a classification label). Avoid it for open-ended text — it loops.