https://k4i.top/tags/sampling/Recent content in Sampling on k4i's blogHugo -- gohugo.ioensky_io@outlook.com (K4i)sky_io@outlook.com (K4i)All content is subject to the license of <a rel="license noopener" href="https://creativecommons.org/licenses/by-nc-sa/4.0/" target="_blank">CC BY-NC-SA 4.0</a> .Thu, 18 Jun 2026 21:20:00 +0800https://k4i.top/posts/llm-sampling-temperature-top-p-top-k/Thu, 18 Jun 2026 21:20:00 +0800sky_io@outlook.com (K4i)Thu, 18 Jun 2026 21:20:00 +0800https://k4i.top/posts/llm-sampling-temperature-top-p-top-k/<p>Why does lowering <code>temperature</code> make an answer more stable? Why does lowering <code>top_p</code> make it more conservative? Why does a small <code>top_k</code> feel like asking the model to choose only from the obvious candidates? These are not three separate magic styles. They are three concrete operations on the <strong>next-token probability distribution</strong>.</p> <p>This post builds the intuition with a 5-token example, then maps the same idea to the vLLM V1 sampler.</p>K4ifeatured imagellminferencesamplingvllmsource-readingai-infraAIvLLM and SGLang Source Reading