Stanford's "Verbalized Sampling" Fixes LLM Mode Collapse With 8 Words

I just published a deep-dive technical breakdown of a new research paper from Stanford and Northeastern that fundamentally changes how we extract creativity from aligned LLMs.

The Problem: Alignment Trapped Creativity We assumed RLHF and DPO permanently damaged model creativity, causing "mode collapse." If you ask an LLM for an idea, you get the safest, most statistically boring answer. Why? Because human annotators have a systematic typicality bias (α = 0.57). We accidentally trained AI to be boring to match human preferences.

The Fix: Verbalized Sampling The creativity isn't gone; it's just trapped. You can bypass mode collapse entirely with an 8-word instruction:

"Generate 5 [items] with their probabilities."

When you force the model to output numeric probabilities, it stops autocompleting the peak of the curve and samples from its actual pre-trained distribution.

The Results:

• 66.8% recovery of base model creativity.

• 1.6–2.1× diversity increase in open-ended generation.

• Larger models (like GPT-5.1) see double the diversity boost compared to smaller models.

You can read my full breakdown, the cognitive psychology behind the bias, and the exact system prompts to implement this in your own apps here: https://medium.com/generative-ai/stanford-just-killed-prompt-engineering-with-8-words-and-i-cant-believe-it-worked-8349d6524d2b

Question for the community: For founders building AI SaaS tools, are you actively combating mode collapse in your backend prompts, or are your users stuck getting the same "typical" outputs as everyone else?