A decoding strategy decides which token is actually emitted from the next token distribution. The model is unchanged by this choice, only the output is.

Temperature

  • the logits are divided by the temperature before the softmax
  • a temperature below 1 makes the distribution sharper, so generation is more conservative
  • a temperature above 1 makes it flatter, so generation is more random

Truncation

  • greedy decoding always takes the single highest probability token
  • top-k keeps a fixed number of candidates
  • top-p keeps the smallest set of tokens whose cumulative mass is at least p, so the number of candidates varies from step to step
  • beam search keeps a fixed number of whole candidate sequences

Pipeline order

  • divide the logits by the temperature
  • truncate with top-k or top-p
  • renormalise the mass that is left
  • sample from what remains

Note

  • changing the sampling policy changes the realised output only, and never the learned probabilities, the weights or the tokenizer
  • an early sampling error matters because it changes every later conditional distribution and can compound

A worked top-p example:

with and cumulative mass , the third token is the one that first reaches , so the first three tokens are kept.

The softmax step this sits on top of is in Sampling.