Verbalized Sampling: The Prompt Against Mode Collapse

Illustration of a robin and a penguin to illustrate mode collapse in AI models.


On this question and the new promptingtechnique 𝗩𝗲𝗿𝗯𝗮𝗹𝗶𝘀𝗶𝗲𝗿𝘁𝗲𝘀 𝗦𝗮𝗺𝗽𝗹𝗶𝗻𝗴 (𝗩𝗦), we have a fresh, detailed video on the YouTube channel of the SRH Distance Learning University – The Mobile University K. I. Crime Novels uploaded:

𝙆.𝙄. 𝙆𝙧𝙞𝙢𝙞𝙨: 𝙂𝙞𝙗𝙩 𝙚𝙨 𝙙𝙚𝙣 𝙢𝙖𝙜𝙞𝙨𝙘𝙝𝙚𝙣 𝙋𝙧𝙤𝙢𝙥𝙩?

Verbalized sampling was introduced in the study How to Mitigate Mode Collapse and Unlock LLM Diversity as a powerful tool against mode collapse in large language models:

In mode collapse, AI models (and thus also large language models) assign a disproportionately high probability of their typical outputs to a single mode—the default mode: The output is not creative, not original, but merely typical. Just as most people, when asked to name a bird, are more likely to choose a robin than a penguin. Even though a penguin would be just as correct.

I explain the further background of the 𝗩𝗲𝗿𝗯𝗮𝗹𝗶𝘀𝗶𝗲𝗿𝘁𝗲𝗻 𝗦𝗮𝗺𝗽𝗹𝗶𝗻𝗴𝘀 and how you can best use it for creative, non-generic chatbot outputs, I explain in detail in the video. Here is just a rough infographic on the topic:



Source link

Leave a Reply

Your email address will not be published. Required fields are marked *