CreativeInstruct: A New Method to Balance Quality and Creativity in LLMs

A recent paper on arXiv presents CreativeInstruct, an adaptable method for instruction tuning that enables large language models (LLMs) to achieve a balance between creative outputs and the quality of models that have undergone post-training. This approach incorporates unique [StartCreativity] spans to encourage a more creative generation process. Additionally, the paper introduces a new metric for structural diversity, utilizing graph edit distance to assess narrative variation that traditional lexical and semantic metrics may overlook. In terms of narrative generation, CreativeInstruct demonstrates comparable or superior diversity to both multi-model benchmarks and distilled output variants, all while maintaining quality and eliminating the need for multiple models. The paper can be found on arXiv with the identifier 2608.07460.