LLM Discussion: Enhancing the Creativity of Large Language Models via Discussion Framework and Role-Play

TL;DR

This work addresses the creativity gap in large language models by introducing LLM Discussion, a three-phase framework (initiation, discussion, convergence) that fosters collective creativity through multi-turn, diverging-converging exchanges. To counteract homogeneity across models, it employs role-playing with diverse personas via Six Thinking Hats, enabling richer perspectives during discussion rounds. Evaluations across four creativity benchmarks (AUT, Instances, Similarities, Scientific Creativity Test) show that LLM Discussion surpasses single-agent and prior multi-agent baselines in Originality and Elaboration, with human assessments corroborating the gains. The framework, along with prompts, role sets, and evaluation protocols, demonstrates practical potential for enhancing LLM-generated creativity and lays groundwork for future explorations in human-LLM collaborative creativity; code and data are available at the referenced GitHub repository.

Abstract

Large language models (LLMs) have shown exceptional proficiency in natural language processing but often fall short of generating creative and original responses to open-ended questions. To enhance LLM creativity, our key insight is to emulate the human process of inducing collective creativity through engaging discussions with participants from diverse backgrounds and perspectives. To this end, we propose LLM Discussion, a three-phase discussion framework that facilitates vigorous and diverging idea exchanges and ensures convergence to creative answers. Moreover, we adopt a role-playing technique by assigning distinct roles to LLMs to combat the homogeneity of LLMs. We evaluate the efficacy of the proposed framework with the Alternative Uses Test, Similarities Test, Instances Test, and Scientific Creativity Test through both LLM evaluation and human study. The results show that our proposed framework outperforms single-LLM approaches and existing multi-LLM frameworks across various creativity metrics. The code is available at https://github.com/lawraa/LLM-Discussion.

Figures (9)

• Figure 1: Role-Play Enhanced LLM Discussion.
• Figure 2: Discussion Framework. We propose an LLM discussion framework that induces collective creativity by bringing divergent and convergent thinking together. The initiation phase informs LLMs of the discussion setup and the task. The discussion phase allows LLMs to build on the ideas of others as well as diverge and generate their own answers. The convergence phase summarizes the discussed ideas and draws a collective conclusion.
• Figure 3: Role Assignment and Example. At the beginning of a discussion, each LLM is assigned a role with specialties and a detailed description of the role, i.e., role prompt.
• Figure 4: Qualitative Results. We present responses generated by LLM Discussion and other baselines on two benchmarks, AUT and Scientific, along with the Originality scores from LLM evaluation. It demonstrates that ideas generated through LLM Discussion are more innovative and provide greater detail.
• Figure 5: Rounds of Discussion. The performance does not consistently improve with more than 5 rounds.