arXiv:2601.02320cs.CL2026-01

用大模型反推文本生成时的温度参数,发现人类写作文本温度多接近1。

Estimating Text Temperature with Language Models

  • 基于最大似然法,用语言模型反推任意文本的生成温度
  • 多数文本温度接近1,但笑话、代码等有明显偏差
  • 适合研究生成风格或检测机器生成内容的研究者

自回归语言模型在推理时通常使用温度参数调节生成文本的随机性。生成文本后,可通过最大似然方法估计该参数。本文提出一种方法,可对任意文本(包括人类撰写)相对于给定语言模型估算其温度。我们在一系列中小规模大模型上评估了温度估计能力,最终选用表现最佳的 Qwen3 14B 模型对多个流行语料库进行温度估计,结果发现:大多数文本温度接近1,但部分例外如笑话、GSM8K 和 AG News 的温度为1.1,而 Python 代码的温度为0.9。

原文摘要 · Abstract (English)

Autoregressive language models typically use temperature parameter at inference to shape the probability distribution and control the randomness of the text generated. After the text was generated, this parameter can be estimated using maximum likelihood approach. Following it, we propose a procedure to estimate the temperature of any text, including ones written by humans, with respect to a given language model. We evaluate the temperature estimation capability of a wide selection of small-to-medium Large Language Models (LLMs). We then use the best-performing Qwen3 14B to estimate temperatures of popular corpora, finding that while most measured temperatures are close to 1, notable exceptions include Jokes, GSM8K, and AG News (1.1), and Python code (0.9).

温度估计语言模型文本分析

Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品;中文卡片由大模型生成,请以原文为准。