Model

Top-p

Why You Need to Understand Top-p

When a model generates each word, it's actually "sampling" from all possible words by probability.

The problem: the probability distribution is long — the "most likely word" may account for 80%, "second most likely" for 10%, and everything else adds up to only 10%. If you sample from all words every time, occasionally you'll pick low-probability tail words, causing output to deviate.

Top-p solves this: controlling the range of the "candidate pool" rather than adjusting the shape of the probability distribution.


What Is Top-p

One-line definition: Top-p limits the sampling candidate range, selecting only from the smallest set of words whose cumulative probability reaches the threshold.

When top-p = 0.9:

  1. The model first lists all candidate words, sorted by probability
  2. Starting from the highest, accumulate until cumulative probability reaches 90%
  3. Only select the next word from the words covered by that 90% probability
  4. The remaining 10% of low-probability tail words are directly excluded

Analogy: Like a teacher grading an exam, only grading the top 90% of the most carefully written parts of the answer sheet — directly ignoring the most illegible parts.


Top-p vs Temperature

Both control randomness but in different directions:

ParameterControls WhatAnalogy
TemperatureAdjusts probability distribution smoothnessAdjusting "risk tolerance"
Top-pLimits the candidate word rangeAdjusting "candidate pool width"

Simply put:

  • Temperature = "dare to pick low-probability words"
  • Top-p = "how wide is the pool of words allowed to be selected"

About Token consumption: these two parameters mainly affect output diversity and stability, not directly affecting Token count, but can indirectly affect Token consumption through output length changes.


How to Do It: When to Adjust Top-p

Suitable for lowering Top-p (0.7-0.9):

  • Wanting some variation but not too far off
  • Being sensitive to low-probability absurd words
  • Needing to control divergence boundaries

Suitable for keeping default Top-p (usually 1.0 or unset):

  • Already adjusting Temperature
  • Wanting the model to have more autonomous space

Practical advice:

  • If already adjusting temperature, usually keep top-p at default first
  • Generally not recommended to significantly increase both simultaneously, otherwise output will noticeably destabilize
  • In many products, temperature is sufficient, Top-p is more for fine-tuning

Remember this: Top-p is the control valve for "candidate pool width" — it controls "which words are eligible to be selected," not "how conservative the selected words are."

Related terms: Temperature