10/26/2026
By Vijeta Deshpande

The Kennedy College of Sciences, Miner School of Computer and Information Sciences, invites you to attend a doctoral dissertation proposal defense by Vijeta Deshpande titled "Navigating Homogeneity: Understanding, Evaluating, and Enhancing Diversity and Creativity in Language Models."

Details

  • Date: Nov. 6, 2026
  • Time: 1 – 2 p.m. EST
  • Location:Zoom meeting

Committee members

  • Prof. Anna Rumshisky, advisor: Associate Professor, Miner School of Computer & Information Sciences
  • Prof. Benyuan Liu: Professor, Miner School of Computer & Information Sciences
  • Prof. Ming Shao: Associate Professor, Miner School of Computer & Information Sciences
  • Prof. Jessy Li: Associate Professor, Computer Science Department, The University of Texas at Austin

Abstract

Human creativity is fundamentally exploratory: it drives us to re-examine existing knowledge, discover unexpected associations between concepts, and synthesize new ideas from familiar elements. In written discourse, this creative exploration is partially embodied in text diversity. Today, large language models (LLMs) possess unprecedented capabilities and mediate a substantial proportion of our written communication. Consequently, the diversity of LLM outputs now plays an active role in reflecting and shaping the collective human creativity landscape. However, empirical studies consistently reveal that current LLMs suffer from constrained output diversity, risking homogenized human ideas, discourse, and expressions. In response, this thesis pursues the overarching goal of understanding, evaluating, and enhancing diversity and creativity in LLM generation.

We begin with a targeted investigation into pre-training, the foundational stage of language model (LM) development. We show that the language understanding capabilities of LMs improve when trained on a less diverse pre-training corpus with fewer unique words. Moving our attention to pre-trained and aligned models in the second study, we investigate how response diversity, both lexical and syntactic, can be enhanced. We present a length-controlled preference-pair curation strategy that effectively isolates the diversity signal, demonstrating that LLMs fine-tuned on this data exhibit significant diversity gains across a variety of downstream tasks. Third, we identify a critical failure mode in data-efficient alignment, specifically activation steering: while activation steering effectively controls target behaviors, it consistently contracts collective diversity, precipitating severe homogenization when model outputs are reused at scale.

Because text diversity serves as the essential foundation for creative thought, we extend our inquiry to the broader construct of creativity by introducing the Artificial General Creativity Benchmark (AGC-Bench), a large-scale meta-benchmark for creativity evaluation. Evaluation of 83 LLMs on AGC-Bench reveals a dominant, domain-general creative factor (the c-factor) governing the cross-task creative performance of LLMs. Finally, we formulate verifiable reinforcement environments that explicitly operationalize divergent exploration and convergent validation through semantic association games.

Together, these contributions establish diversity and creativity as twin objectives across the model lifecycle, offering scalable pathways to foster genuinely creative machine intelligence.