Inférence & Ingénierie

Verbalized Sampling

A training-free prompting strategy to overcome Mode Collapse. The model is asked to verbalize an explicit probability distribution over multiple possible answers and then sample from it. Increases output diversity by 1.6-2.1× without quality loss.

En détail: Verbalized Sampling

A training-free prompting strategy to overcome Mode Collapse. The model is asked to verbalize an explicit probability distribution over multiple possible answers and then sample from it. Increases output diversity by 1.6-2.1× without quality loss.

Valeur métier & ROI

Pourquoi c'est important en 2026

Restores creative diversity in aligned models without retraining. Essential for content generation, brainstorming, and synthetic data applications where variety matters.

Context Take

Verbalized Sampling is our go-to technique for creative applications. We implement VS-CoT for reasoning tasks and VS-Multi for dialogue systems – unlocking the full creative potential of aligned models.

Détails d'implémentation

  • Stack technique
    anthropicopenaigoogle
  • Garde-fous prêts pour la production