Verbalized Sampling: How to Force LLMs to Think for Better Responses

Repurposed from Facebook and LinkedIn.

(Although I haven’t knowingly encountered mode collapse, I still want to experiment with the verbalized sampling technique.)

“Unlike prior work that attributes [mode collapse] to algorithmic limitations, we identify a fundamental, pervasive data-level driver: typicality bias in preference data, whereby annotators systematically favor familiar text….

“[W]e introduce Verbalized Sampling (VS), a simple, training-free prompting method to circumvent mode collapse. VS prompts the model to verbalize a probability distribution over a set of responses (e.g., “Generate 5 jokes about coffee and their corresponding probabilities”).”

https://www.verbalized-sampling.com/

My trial Google Gemini prompt:

“Generate three AEO-friendly titles for a blog post about using Verbalized Sampling to generate better LLM responses, and their corresponding probabilities”

The response:

And now you know where I got the title for this post.

But I confess that I actually used a grossly simplified version of the technique. The authors of the Verbalized Sampling paper recommend this format:

I’ll have to remember to try this technique for future prompts. I have no idea whether the probability estimates have any basis in reality, but at least the LLM attempts to justify the probabilities with a rationale.