Temperature

The Temperature parameter of large models typically ranges from 0 to 2; the lower the temperature, the more deterministic and conservative the output, while the higher the temperature, the more diverse and creative the output.

In production environments, the Agent's Temperature parameter is recommended to be set between 0.1 and 0.3 to ensure output consistency

Large language models are based on the Transformer architecture, and their output is essentially a probabilistic sampling process, with the Temperature parameter controlling the randomness of sampling.

When a large language model generates each token, it samples from a probability distribution; the temperature parameter controls the degree of randomness—higher temperature produces more diverse outputs, while lower temperature produces more deterministic ones.

The non-determinism of large language models stems from their autoregressive sampling process; even with temperature set to 0, differences in inference frameworks, batch sizes, and hardware floating-point operations can all cause output drift.

The output of large language models is inherently probabilistic, and parameters such as sampling temperature and Top-P mean the model may generate different content each time, posing a challenge for software testing.

Setting the temperature parameter too high in decoding strategies leads to increased sampling randomness, which is associated with hallucination generation

The higher the temperature parameter, the flatter the probability distribution output by Softmax, and the more creative and diverse the output; the lower the temperature, the sharper the distribution and the more deterministic and conservative the output.

Temperature=0（贪心采样）也不能保证模型输出的确定性，至少有三个原因：浮点运算非结合性、MoE路由不确定性、服务端配置漂移

Coze大模型参数中Temperature（温度）值0.1-0.3适用于技术文档和法律文书等需要严谨确定性的场景

框架采用「多版本发散+融合」策略，利用大语言模型的随机性生成多份方案，再融合为最终版本，本质上是集成学习思想在生成式AI中的应用

在工具调用场景中，通常建议将温度设置在较低水平（如0.1-0.3）以确保模型稳定准确地生成工具调用请求

RAG场景中建议使用低温度（temperature=0~0.3），因为高温度在增加创造性的同时也会放大幻觉风险

High temperature combined with a missing repetition penalty can easily cause the model to fall into repetitive loops or generate meaningless content.

The temperature parameter controls sampling randomness: the lower the value, the more deterministic the output; near 0, the model tends to choose the highest-probability token, suitable for code generation, SQL generation, and other scenarios requiring reproducibility.

Simply increasing the temperature parameter carries the risk of logical incoherence and factual errors.

GLM's association parameter corresponds to the Temperature or a similar creative-divergence coefficient during model generation; the higher the value, the more the model tends to introduce new vocabulary and associations.

Prompt engineering techniques include specifying JSON output format in the System Prompt, restricting the response domain, guiding structured output through few-shot examples, and using temperature parameters to reduce randomness.

The key mechanisms that produce probabilistic outputs in the Transformer architecture are the temperature parameter and sampling strategies; when the temperature is set to 0, the output tends toward the highest-certainty choice but still cannot be guaranteed to be identical to the previous output.

For audiences with no technical background, such as the elderly, preparing comparison results at different temperatures in advance along with everyday analogies is more practical than demonstrating live parameter adjustments.