verbalized-sampling

To get diverse LLM outputs, prompt the model to generate a probability distribution over multiple responses rather than a single answer. This technique, Verbalized Sampling, counteracts mode collapse and ensures genuine diversity in brainstorming and creative writing.

LLMs return the same cluster of answers due to mode collapse caused by typicality bias in alignment data. Verbalized Sampling addresses this by requesting a probability distribution across responses, encouraging the model to explore tail and distribution-aware outputs.

Verbalized sampling requires no fine-tuning. It is a training-free technique that prompts the model to produce k diverse outputs with explicit probabilities, returning them as a JSON array to increase coverage in open-ended QA and synthetic data generation.

Use probability distribution prompting for open-ended QA when the task needs genuine diversity and multiple valid answers exist. It is also effective for persona simulation, adversarial examples, and synthetic data generation where typicality bias limits output coverage.

To generate synthetic data with diverse outputs, prompt the model to produce k distinct candidate responses, each with an explicit probability. This verbalized sampling method formats results as a JSON array, avoiding typicality bias and expanding data coverage.