CreativeBench: Benchmarking and Enhancing Machine Creativity via Self-Evolving Challenges
Researchers have introduced CreativeBench, a new benchmark designed to evaluate and enhance machine creativity, particularly in code generation. The benchmark utilizes a cognitive framework and includes two subsets, CreativeBench-Combo and CreativeBench-Explore, to assess combinatorial and exploratory creativity. A key finding is that while scaling models improves combinatorial creativity, it can lead to diminished returns in exploration and a tendency towards "convergence-by-scaling," making models more correct but less divergent. To address this, the paper proposes EvoRePE, a strategy to improve machine creativity by incorporating evolutionary search patterns. AI
IMPACT Introduces a new benchmark for evaluating machine creativity, potentially guiding future model development towards more divergent and exploratory capabilities.