Year: 2025
Type: dissertation
Abstract: Creativity is widely regarded as a hallmark of human intelligence and has historically propelled cultural and technological progress. The rapid development of artificial intelligence AI has prompted the question: Can machines exhibit comparable creativity? Recent advances in large language models (LLMs) offer promising tools to simulate creative activities through language. However, modern LLMs that primarily rely on in-domain fine-tuning still falter in low-resource, long-tail distributions, lack robust mechanisms for unconventional reasoning, and creative outputs remain challenging to evaluate due to their subjective nature and high-level attention required for verification.This dissertation presents a set of modeling, benchmarking, and evaluation techniques to bridge the gap in machine creativity expressed through language. The first part focuses on creative language generation, introducing techniques that disentangle training from inference to bypass the limitations of domain-specific data. Models are augmented at inference-time with domain knowledge and symbolic structures, enhancing their ability to generate humorous or artistically original outputs.The second part centers on unconventional reasoning, introducing a large-scale benchmark of cognitive puzzles that demand non-standard planning and tool use. These tasks probe LLMs’ capacity to reason beyond templates, compare human versus machine problem-solving, and propose techniques that allow models to refine their reasoning paths iteratively.The third part rethinks evaluation, advocating for idea-level analysis to move beyond surface metrics. By projecting outputs into latent event spaces or analyzing narrative arcs and trajectories at the discourse-level, this part proposes idea-level methods that better differentiate human and machine creativity.Together, these contributions build a foundation for creative language modeling—models that can generate, reason, and reflect more like humans. By addressing both generation and evaluation, this thesis aims to advance the field toward more imaginative, controllable, and user-engaging AI systems.
Author: Tian, Yufei
Institution: University of California, Los Angeles
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7065656959