How Temperature Controls Creativity in Large Language Models

Temperature measures the degree of randomness in a large language model’s generation, similar to how molecular motion reflects heat; a low temperature (e.g., 0.1) makes the model follow the “recipe” closely, producing almost identical, safe outputs, whereas a high temperature (e.g., 1.5) lets the model “play” with the ingredients, yielding creative but sometimes ill‑structured results.

Effect of Different Temperature Settings

When temperature is near 0, the model prefers the most probable token at each step, resulting in precise, conservative answers. As temperature rises toward 1 or higher, the model samples less probable tokens more often, increasing variability, imagination, and occasional logical drift.

Illustrative Examples

Temperature = 0.1 – "春天来了，花儿开了，鸟儿在歌唱。" (Consistent but bland)

Temperature = 0.7 – "细雨润新绿，微风抚柳枝，蝴蝶吻过桃花的唇。" (Moderately varied and poetic)

Temperature = 1.5 – "春天是一块融化的彩虹糖，蚂蚁在云朵上野餐！" (Highly imaginative, may be incoherent)

Task‑Specific Recommended Temperatures

Precise Q&A / technical dialogue – 0.0 ~ 0.3 – prioritize accuracy and avoid hallucination.

Code generation / structured writing – 0.2 ~ 0.5 – allow slight variation while staying reliable.

Copywriting / creative content – 0.7 ~ 1.0 – encourage diverse expression and inspiration.

Novel or poetry writing – 1.0 ~ 1.8 – seek rich imagination and expressive freedom.

Code Demonstration

from openai import OpenAI

client = OpenAI(api_key="your-api-key")
# Low temperature: suitable for precise answers
response = client.ChatCompletion.create(
    model="gpt-4",
    messages=[{"role": "user", "content": "解释相对论"}],
    temperature=0.2  # stable and accurate output
)
print(response.choices[0].message.content)

# High temperature: suitable for storytelling
response = client.ChatCompletion.create(
    model="gpt-4",
    messages=[{"role": "user", "content": "写一个外星人童话"}],
    temperature=0.9  # more creative output
)
print(response.choices[0].message.content)

Important Considerations

Setting temperature to 0 can cause repetitive, overly deterministic output, while values above 1.5 may generate garbled or nonsensical text. Choose a temperature that matches the desired balance between reliability and creativity for the specific use case.