Study: Iterative Generation Improves LLM Creativity in Recipe Design

A recent pilot study published on arXiv (2608.07243) examines the potential of iterative generation and evaluation to boost the creativity of large language models (LLMs) in crafting recipes. Researchers modified the FunSearch technique to create recipes for the 2024 Pillsbury Bake-Off and assessed the results against human standards using a TTCT-based evaluation method. Two experiments explored the impact of iteration counts, generator temperature, and the size of the in-loop selection scorer model. Findings reveal that while iterative generation-selection can yield creativity scores similar to human benchmarks, merely increasing iterations does not enhance creativity. The in-loop evaluator plays a crucial role; a smaller selection scorer significantly improves scores across most TTCT dimensions, whereas temperature affects originality minimally. This suggests that the design of the evaluator is a critical factor in LLM-driven creative tasks.