Generate diverse, high-quality responses by producing multiple candidates with probability distributions instead of a single answer. Use when the user asks for brainstorming, creative options, diverse alternatives, original ideas, or when you detect a task that benefits from exploring multiple approaches before committing to one. Also use when explicitly asked for 'verbalized sampling', 'diverse candidates', 'tail sampling', or 'show me options'. Covers creative writing, synthetic data, dialogue simulation, open-ended QA, and any task where mode collapse (repetitive/predictable outputs) is a risk.
