Paper Examines Temperature Sensitivity in LLM Truncation Sampling

A new arXiv paper analyzes how the temperature parameter in large language model decoding interacts with truncation samplers such as top-p and min-p, which remove the least likely tokens from the candidate pool. The authors describe "temperature fragility" and argue that truncation can offer conditional benefits depending on how sampling is configured. The work is a theoretical and empirical study of decoding strategies rather than a released model or product.