Verbalised Sampling to maximise AI diversity
Ever get that feeling that, when you ask an AI for something new and creative, it just gives you the same dull, repetitive answer over and over again?
What is ‘Mode Collapse’?
Well apparently, it is a well-known phenomenon where AI model has a tendency to produce repetitive, less diverse output. On technical terms this is called ‘Mode Collapse’.
Humans are somewhat responsible for this. We have a ‘subconscious preference’ for things that are familiar, predictable and easy to digest. Now, considering that AI models are partly trained on human data, that ‘subconscious preference’ has a huge unintended consequence.
Issue explained step by step
-
AI model starts out with a number of possible answers.
-
During training stages (fine-tuning and post-training), human raters, who are subconsciously bias, keep rewarding the same set of answers because those feel that most, typical/safe/predictable.
-
This teaches AI model to respond with a very similar set of answers, drastically reducing creativity, without even realising it.
Diverse Base AI (multiple good answers) -> Human Bias (preference for answer A) -> Mode Collapse (AI responds with answer A)
But how can we improve our prompt to increase the diversity of the output?
A typical AI user would normally ask for an answer.
✅ But instead of just asking for an answer, you tell AI model to come up with multiple responses and their probabilities.
With this change in how we ask can create a massive difference in creativity.
This technique can be applied for text, image, video generation.
👉 For full credit and to read research paper: arxiv.org/html/2510.01171v3