Source URL: https://simonwillison.net/2025/May/4/llm-sampling/#atom-everything
Source: Simon Willison’s Weblog
Title: Dummy’s Guide to Modern LLM Sampling
Feedly Summary: Dummy’s Guide to Modern LLM Sampling
This is an extremely useful, detailed set of explanations by @AlpinDale covering the various different sampling strategies used by modern LLMs. LLMs return a set of next-token probabilities for every token in their corpus – a layer above the LLM can then use sampling strategies to decide which one to use.
I finally feel like I understand the difference between Top-K and Top-P! Top-K is when you narrow down to e.g. the 20 most likely candidates for next token and then pick one of those. Top-P instead “the smallest set of words whose combined probability exceeds threshold P" – so if you set it to 0.5 you’ll filter out tokens in the lower half of the probability distribution.
There are a bunch more sampling strategies in here that I’d never heard of before – Top-A, Top-N-Sigma, Epsilon-Cutoff and more.
Reading the descriptions here of Repetition Penalty and Don’t Repeat Yourself made me realize that I need to be a little careful with those for some of my own uses of LLMs.
I frequently feed larger volumes of text (or code) into an LLM and ask it to output subsets of that text as direct quotes, to answer questions like "which bit of this code handles authentication tokens" or "show me direct quotes that illustrate the main themes in this conversation".
Careless use of frequency penalty strategies might go against what I’m trying to achieve with those prompts.
Via Hacker News
Tags: prompt-engineering, llms, ai, generative-ai
AI Summary and Description: Yes
Summary: The text discusses various sampling strategies utilized by modern large language models (LLMs), providing clarity on concepts like Top-K and Top-P sampling. It emphasizes the importance of understanding these strategies for enhancing the effectiveness of LLM interactions, particularly in practical applications related to security and coding.
Detailed Description: The text provides insights into the methodologies employed by LLMs to choose the next token during text generation. This is particularly relevant for professionals in AI and AI Security, as understanding these strategies can impact the effectiveness and security of applications leveraging LLMs.
Key points include:
– **Sampling Strategies Overview**:
– **Top-K Sampling**: This method involves selecting the top K most probable tokens and randomly choosing one from this set. It allows for some diversity while maintaining relevance.
– **Top-P Sampling**: Also known as nucleus sampling, it selects tokens based on their cumulative probability, filtering out those that fall below a certain threshold P. For example, a threshold of 0.5 means only the smallest set of tokens whose probabilities sum up to over 0.5 is considered.
– **Additional Strategies**:
– Introduction of other less common strategies like Top-A, Top-N-Sigma, and Epsilon-Cutoff, which may offer nuanced approaches to sampling.
– **Practical Implications**:
– The text highlights the significance of understanding these strategies for optimal use cases, especially when performing tasks like extracting specific code snippets or thematic quotes from larger texts.
– A cautionary note regarding the use of repetition penalties and frequency penalties is provided, indicating that improper application of these strategies could hinder the intended results when querying LLMs, especially in security and coding contexts.
Overall, for AI, AI Security, and generative AI stakeholders, mastering these sampling strategies can enhance the performance and reliability of models in real-world applications, especially where precise language generation is crucial.