The simple prompting technique that restores AI diversity

Have you ever noticed how ChatGPT, Claude or Gemini sometimes sound like they've lost their spark? Ask for five jokes about coffee and you might get the same response repeatedly.

"Why did the coffee file a police report? It got mugged!" response… five times.

This isn't a bug. It's called mode collapse, and researchers at Stanford University and Northeastern University just figured out why it happens. They published this paper on how to address this issue and the solution involves just one prompt tweak.

The problem isn't AI, it's us

When we train these models to be "helpful", we use human feedback. But we humans unconsciously prefer familiar, easy-to-process text. The researchers call this typicality bias - we rate "safe" responses higher, even when creative answers might be just as good.

This phenomenon is rooted in cognitive psychology. The mere-exposure effect means we naturally favour what's familiar. Processing fluency makes conventional text feel more truthful. These biases compound during alignment training, gradually steering models away from their original creative diversity.

After alignment training, a beautifully diverse base model becomes … well, vanilla.

The simple prompt that brings it back

Here's the trick. Instead of saying:
"Tell me a joke about coffee."

Try this:
"Generate five jokes about coffee with their probabilities."

That's it. They call it Verbalised Sampling.

Why it works

When you ask for one answer, the model gives you the most "typical" answer. But when you ask for a distribution with probabilities, you're accessing the full creative range it learned during training - before safety training narrowed things down.

Different prompts lead to different outcomes:

• Single-instance prompts: produce the safest answer.
• Distribution prompts: unlock the model's original diversity.

What the research found

According to the study, results improved across several areas:

• Creative writing became 1.6 - 2.1 times more diverse.
• Factual accuracy and safety were unaffected.
• Better dialogue simulation.
• More realistic open-ended responses.
• Improved synthetic data for maths tasks.

Interestingly, the more capable the model, the more it benefited. GPT-4 and Claude-4 showed the biggest gains.

What this means for you

If you build with LLMs:

• Use distribution-level prompts for creative outputs.
• Tune diversity by setting probability thresholds (for example, "generate responses where probability < 10%").
• No training required, just better prompting.

If you use ChatGPT, Claude or Gemini:

• Ask for multiple responses with probabilities for diverse outputs.
• It unlocks the model's "pre-training brain". 

The bigger picture

The diversity problem isn't purely technical - it's baked into the preference data itself. But aligned models haven't lost their creativity. It's still there, just hidden. With the right prompting, we can bring it back.

This finding has significant implications for how we approach AI development. Rather than viewing alignment and creativity as opposing forces, we can design systems that maintain safety whilst preserving the rich diversity of the underlying model.

It's pretty neat that a change as simple as "show me your working" can help AI rediscover its imagination.

If you’d like to find out how we can help you build more diverse, creative, and human-centred AI solutions, connect with me!