A new Stanford paper just dropped a bombshell on how we prompt AI models.
It’s called Verbalized Sampling, and it proves that aligned models like ChatGPT or Claude aren’t “stale” — we’ve just been prompting them wrong.
Here’s the core problem: Post-training alignment causes mode collapse.
Ask an AI for a joke 5 times, and it’ll give you the same one every time.
The real cause? Typicality bias — human annotators in preference training pick the most familiar, safest answer. The model learns to play it safe.
But the researchers found the diversity isn’t gone — it’s trapped.
Their fix is hilariously simple:
Instead of “Tell me a joke,” ask “Generate 5 jokes with their probabilities.”
That’s it. No retraining, no fine-tuning — just one extra phrase.
The results:
✅ 1.6–2.1× more diversity in creative tasks
✅ 66.8% recovery of pre-alignment variety
✅ Zero drop in factual accuracy or safety
Why it works: When asked for one response, the model gives its safest guess.
When asked for a distribution, it reveals the full range of what it actually knows.
This method boosts everything — from creative writing to open-ended reasoning — and larger models benefit the most.
The diversity wasn’t lost.
We just forgot how to ask for it.
📄 Paper: “Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity”
👉 arxiv.org/abs/2510.01171
