Temperature is a sampling hyperparameter that controls how "confident" or "random" an LLM's next-token choices are — a single number with an outsized effect on generated text's character.
Formula
This is exactly the softmax formula from Softmax Function, with every logit \(z_i\) divided by a temperature \(T\) before exponentiating. \(T=1\) reproduces standard softmax exactly.
What Different Temperature Values Do
| Temperature | Effect on the Distribution | Generated Text Character |
|---|---|---|
| \(T \to 0\) | Sharpens toward the single highest-probability token (approaches argmax) | Highly deterministic, repetitive, "safe" |
| \(T = 1\) | Unchanged — the model's raw learned distribution | Balanced |
| \(T > 1\) | Flattens the distribution, making less-likely tokens relatively more probable | More random, varied, sometimes less coherent |
Numerical Example
Logits \([2.0, 1.0, 0.1]\) at \(T=1\) (standard softmax, from Softmax Function): \([0.659, 0.242, 0.099]\). At \(T=0.5\) (dividing logits by 0.5, i.e. doubling them to \([4.0,2.0,0.2]\) before softmax): the distribution sharpens considerably, e.g. approximately \([0.843, 0.140, 0.017]\) — the top token becomes much more dominant. At \(T=2\) (halving the logits to \([1.0,0.5,0.05]\)): the distribution flattens toward more uniform, e.g. approximately \([0.475, 0.288, 0.237]\).
Code
import torch
import torch.nn.functional as F
logits = torch.tensor([2.0, 1.0, 0.1])
for T in [0.5, 1.0, 2.0]:
scaled_probs = F.softmax(logits / T, dim=0)
print(f"T={T}: {scaled_probs}")
# T=0.5: sharper, more confident distribution
# T=1.0: the model's original, unmodified distribution
# T=2.0: flatter, more uniform, more "random" distributionWhen to Use Which Temperature
| Task Type | Typical Temperature |
|---|---|
| Factual Q&A, code generation, math | Low (0–0.3) — favors reliability and determinism |
| Creative writing, brainstorming | Higher (0.7–1.2) — favors variety and novelty |
Common Mistakes
- Setting temperature very high expecting purely "more creative" output without downside — excessive temperature can degrade coherence and factual reliability, not just increase variety.
- Confusing temperature with top-k or top-p sampling — temperature reshapes the entire probability distribution's sharpness; top-k/top-p (next two notes) instead restrict which tokens are even eligible to be sampled from, a distinct and complementary mechanism.
Interview Relevance
Q: "Why would you use a low temperature for a code-generation task but a higher one for creative writing?" Code generation typically benefits from reliability and correctness — a low temperature sharpens the distribution toward the model's most confident (and typically most likely to be syntactically/logically correct) predictions. Creative writing benefits from variety and novelty — a higher temperature flattens the distribution, giving less-likely but potentially more interesting or original word choices a real chance of being sampled.
Practice Question
As \(T \to 0\), what does the sampling process approach — a random draw, or a deterministic choice? Explain using the softmax formula.