Estimating Text Temperature
This work addresses a specific technical challenge in language model analysis, offering a tool for researchers and practitioners to infer generation parameters from text, but it is incremental as it builds on existing maximum likelihood methods.
The paper tackles the problem of estimating the temperature parameter used in autoregressive language models for text generation, proposing a procedure to estimate this temperature for any text, including human-written ones, and evaluates it across various models, with Qwen3 14B achieving the best performance and being used to estimate temperatures of popular corpora.
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 LLMs. We then use the best-performing Qwen3 14B to estimate temperatures of popular corpora.