Greedy decoding
Always taking the single most likely next word — the safe, repeatable, sometimes boring way to generate text.
Think of it like
Always ordering the bestseller — reliable, never surprising.
Example
Set temperature to 0 and you get greedy decoding: the same answer every time.
How it actually works
Greedy decoding skips the dice roll and grabs the top token at every step. That makes it reproducible and good for tasks where you want one right answer — extraction, classification, structured output. The cost is blandness and the occasional dead end, since the locally-best word isn’t always the globally-best sentence.
For product teams
Turn it on (temperature 0) when you need consistent, repeatable outputs.
For engineers
Argmax decoding at each step; deterministic, equivalent to temperature→0.
Related
- Sampling — The deterministic opposite of sampling.
- Temperature — What temperature 0 effectively turns on.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome