verbalized-sampling

Why we included this project

Anyone who has watched an LLM give back the same safe answer no matter how the prompt is phrased will recognize the problem this library targets. Verbalized Sampling asks the model to produce several candidate responses with a numeric probability attached to each, then samples from that verbalized distribution instead of leaning on temperature. Because it is a prompting technique, it needs no fine-tuning and works with GPT, Claude, Gemini, Llama, and other models, so it slots into existing pipelines without much friction. The Python package wraps the flow in a single call, includes LangChain integration, and provides scripts to reproduce the paper's creative-writing, dialogue-simulation, and synthetic-data experiments, which makes it a practical way to get more variety out of models a team already runs.