← All articles

AI & LLMs

Temperature and sampling — how an LLM picks its words

2026-07-30 · 5 min read

Ask a model the same question twice and you often get two different answers. That's not a bug — it's a dice roll, and there's a dial controlling how loaded the dice are. A model never simply "picks a word." It produces a probability for every possible next token, and then samples one. Temperature is the setting that decides how much it favours the safe, likely choice versus a surprising one.

From probabilities to a choice

For "the sky is ___", the model might assign 80% to "blue," 8% to "grey," 5% to "clear," and a long tail of everything else. Sampling then draws from that distribution. Temperature reshapes it before the draw: low temperature sharpens the peaks (the likeliest token dominates → predictable), high temperature flattens them (rarer tokens get a real chance → creative, or unhinged).

Temperature reshapes the token distribution prompt: "the sky is ___" low temp · sharp bluegreycleardark high temp · flat bluegreycleardark
Same underlying odds. Low temperature makes "blue" almost certain; high temperature gives the other words a real shot.
The gist

The model rolls weighted dice for each word. Temperature 0 = always take the single most likely word (predictable). Crank it up and the dice flatten out, so it reaches for rarer, riskier words.

Choosing a temperature

  • Low (0–0.3): factual answers, code, extraction, classification — anywhere you want the safest, most consistent token and reproducibility. This is also your friend against hallucination: it favours the well-supported continuation.
  • Medium (0.7-ish): the default for general chat — coherent but not robotic.
  • High (1+): brainstorming, creative writing, variety — at the cost of coherence and correctness.

Its cousins: top-k and top-p

Temperature is usually paired with a cutoff so the flattened tail can't produce nonsense. Top-k keeps only the k most likely tokens before sampling; top-p (nucleus) keeps the smallest set of tokens whose probabilities add up to p (say 0.9). Together they let you turn up creativity without letting the model reach for genuinely absurd words.

The practical takeaway: if you need the same answer every time, turn temperature down. If a task keeps coming out flat and samey, turn it up. It's not a quality knob — it's a boldness knob, and the right setting is entirely about the job.


LLMTemperatureSamplingAI
← Back to the blog