Creative Priming: Decoupling Process from Product to Escape the Creative Paradox

Creative Priming. Decoupling Process from Product to Escape the Creative Paradox. By Alon Albalak and four others, from Lila Sciences. Published on the ninth of November.

Many of the largest advances people hope AI will deliver are creative acts at heart: curing diseases, inventing new technologies, formulating new scientific theories. Creativity, though, is unusually hard to learn, because of what the authors call the creative paradox. Train a model to repeat past examples of creative success, and those particular examples become more likely to come back. Learning from creativity's products ends up suppressing the very creativity being sought.

Their way out is to change what counts as the thing being learned. The work argues that creativity is better modelled as an explicit process than as a property of individual artifacts, and proposes learning the generative principles behind that process. Creative priming, or CP, puts this into practice by splitting the work between two models, so that the creative process is explicitly decoupled from generating the response.

A trained priming model looks at a prompt and writes a primer specific to it. That primer is handed to a response model, to help the response model be creative. The priming model is the one that learns, and it learns through GRPO.

What the training rewards is the crux. Primers are rewarded when the responses they lead to are novel, high in quality, and diverse. The abstract describes this design and the reasoning behind it, and does not report experimental figures.

Where the reward lands is what dissolves the paradox, at least as the authors frame it. Credit flows to the primer, the process, and never to a particular response, the product. The procedure that yields creative responses is reinforced, while no single creative answer is ever made more likely to be repeated.