Creative Priming: Decoupling Process from Product to Escape the Creative Paradox
A separate model learns to write primers that help another model produce creative responses.
The authors argue that training on successful creative examples encourages repetition of those examples. Their proposed method, creative priming, instead rewards a model for generating prompt-specific guidance that leads to novel, high-quality and diverse responses. It trains that priming model using GRPO, reinforcing the guidance process without reinforcing any particular answer. The abstract describes the approach but reports no benchmark results.
HF Daily Papers' note
The authors argue that training on successful creative examples encourages repetition of those examples. Their proposed method, creative priming, instead rewards a model for generating prompt-specific guidance that leads to novel, high-quality and diverse responses. It trains that priming model using GRPO, reinforcing the guidance process without reinforcing any particular answer. The abstract describes the approach but reports no benchmark results.
HF Daily Papers' note
score 5