Verbalized Sampling is shaking up how we coax LLMs to be creative. By asking for probability distributions instead of single outputs, it fights mode collapse without retraining. [1]
What is Verbalized Sampling? Instead of prompting for one output (e.g., “Tell me a joke”), you prompt for distributions (for example, “Generate 5 jokes with their probabilities”). This reframes the model’s baked-in preferences, reducing repeats and broadening the set of plausible responses. [1]
Why it matters - Mode collapse comes from biases baked into human annotation; RLHF sharpening can amplify this bias. [1] - Diversity boosts performance on creative tasks—about 2.1x—with no drop in quality. [1] - No retraining required; gains come from prompting rather than model weights. [1] - The approach has been demonstrated as a general method to counteract the bias and improve variety across models. [1]
Getting started & resources The paper’s authors are Jiayi Zhang, Simon Yu, Derek Chong, Anthony Sicilia, Michael Tomz, Christopher Manning, and Weiyan Shi. [1] Resources include the Paper, Blog, X Thread, Video, and Quickstart & Colab for hands-on use. [1] Quickstart & Colab makes it easy to try Verbalized Sampling without retraining. [1]
In short, Verbalized Sampling offers a retraining-free path to broader, more reliable LLM creativity. [1]
References
[R] Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity
Proposes Verbalized Sampling: prompt distributions to counter mode collapse, restore output diversity in LLMs with no retraining; shows broad gains.
View source