A new trick for better chatbot jokes?
The quest continues
“Hey Gemini, what’s the opposite of coffee? Sneezy. 😆”
My greatest disappointment with chatbots is their terrible dad jokes. Maybe we’ve found a fix?
New research on “verbalized sampling” shows that adding the following line makes models twice as creative.
“Generate 5 responses with their corresponding probabilities, sampled from the full distribution.”
Why does this work? Basically, post-training alignment encourages models to prefer “safe” answers that appeal to an average user. This trick forces them to build a wider range of ideas.
This worked in my tests, sort of. 🤔
Both Gemini and ChatGPT created more diverse jokes, but the original ones were actually funnier. Wordiness isn’t always better, especially for humor.
I also didn’t see improvements on longer prompts like “Write a story of Matt completing an Ironman”.
That said, this is a helpful view into how models work and tactics to boost creativity.
Give it a try and see if your chatbot gets funnier, or just takes longer to get to the punchline. 🫠☕
Research credit: Jiayi Zhang , Simon Yu, Derek Chong, Anthony Sicilia Michael Tomz, Christopher Manning, Weiyan S. Northeastern University Stanford University West Virginia University
Study Link —> Verbalized Sampling: How To Mitigate Mode Collapse And Unlock Llm Diversity