CHATS-lab/verbalized-sampling
A training-free prompting strategy that mitigates mode collapse in LLMs by having models generate responses with probabilities, achieving 2-3x diversity improvements.
Not currently ranked — collecting fresh signals.
star history
Verbalized Sampling is a prompting strategy that asks LLMs to generate multiple candidate responses along with their probabilities, then samples from this distribution to improve output diversity. It works by instructing the model to verbalize its confidence and enumerate alternatives, reducing the tendency of LLMs to produce repetitive outputs. The approach is training-free and model-agnostic, functioning across GPT, Claude, Gemini, Llama, and other LLMs via prompting alone.
Frequently asked
- What is CHATS-lab/verbalized-sampling?
- A training-free prompting strategy that mitigates mode collapse in LLMs by having models generate responses with probabilities, achieving 2-3x diversity improvements.
- Is verbalized-sampling open source?
- Yes — CHATS-lab/verbalized-sampling is an open-source project tracked on heatdrop.
- What language is verbalized-sampling written in?
- CHATS-lab/verbalized-sampling is primarily written in Python.
- How popular is verbalized-sampling?
- CHATS-lab/verbalized-sampling has 785 stars on GitHub.
- Where can I find verbalized-sampling?
- CHATS-lab/verbalized-sampling is on GitHub at https://github.com/CHATS-lab/verbalized-sampling.