Quantified Diversity Gains: The Numbers Behind Verbalized Sampling

~10 min read

Verbalized sampling isn't just a clever trick — it's backed by measured results: 1.6-2.1x diversity improvement over direct prompting, larger gains on more capable models, and ~66.8% diversity retention across post-training stages.

Quantified Diversity Gains: The Numbers Behind Verbalized Sampling is a Pro topic

Sign in, then upgrade to Pro or Power to unlock this topic and the full AI Engineering curriculum.

Key points

  • VS improves diversity 1.6-2.1x over direct prompting, while maintaining or improving quality — not a trade-off
  • VS-based CoT and VS-based Multi-turn variants push diversity gains even further
  • Larger, more capable models (GPT-4.1, Gemini-2.5-Pro) benefit MORE from VS, not less — up to 2x the gain of smaller models
  • VS retains ~66.8% of the base model's original diversity across post-training stages (SFT, DPO, RLVR), versus a much lower rate for direct prompting
  • VS's gains are independent of temperature/top-p sampling — it composes with them rather than replacing them