Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Creativity in LLM-based Multi-Agent Systems: A Survey

Yi-Cheng Lin, Kang-Chieh Chen, Zhe-Yan Li, Tzu-Heng Wu, Tzu-Hsuan Wu, Kuan-Yu Chen, Hung-yi Lee, Yun-Nung Chen

TL;DR

This work addresses the problem of evaluating and enhancing creativity in LLM-driven multi-agent systems. It proposes a unified framework that decomposes MAS creativity into three core techniques—Divergent Exploration, Iterative Refinement, and Collaborative Synthesis—and analyzes how agent proactivity and persona design shape ideation, evaluation, and collaboration. The paper surveys datasets, evaluation metrics (both objective and subjective), and real-world studies, while outlining challenges such as biased personas, unclear authorship, and the need for standardized benchmarks. By providing a roadmap and a structured taxonomy, the survey aims to standardize evaluation, foster transparent collaboration between humans and AI agents, and accelerate the development of creative MAS across text and image domains.

Abstract

Large language model (LLM)-driven multi-agent systems (MAS) are transforming how humans and AIs collaboratively generate ideas and artifacts. While existing surveys provide comprehensive overviews of MAS infrastructures, they largely overlook the dimension of \emph{creativity}, including how novel outputs are generated and evaluated, how creativity informs agent personas, and how creative workflows are coordinated. This is the first survey dedicated to creativity in MAS. We focus on text and image generation tasks, and present: (1) a taxonomy of agent proactivity and persona design; (2) an overview of generation techniques, including divergent exploration, iterative refinement, and collaborative synthesis, as well as relevant datasets and evaluation metrics; and (3) a discussion of key challenges, such as inconsistent evaluation standards, insufficient bias mitigation, coordination conflicts, and the lack of unified benchmarks. This survey offers a structured framework and roadmap for advancing the development, evaluation, and standardization of creative MAS.

Creativity in LLM-based Multi-Agent Systems: A Survey

TL;DR

This work addresses the problem of evaluating and enhancing creativity in LLM-driven multi-agent systems. It proposes a unified framework that decomposes MAS creativity into three core techniques—Divergent Exploration, Iterative Refinement, and Collaborative Synthesis—and analyzes how agent proactivity and persona design shape ideation, evaluation, and collaboration. The paper surveys datasets, evaluation metrics (both objective and subjective), and real-world studies, while outlining challenges such as biased personas, unclear authorship, and the need for standardized benchmarks. By providing a roadmap and a structured taxonomy, the survey aims to standardize evaluation, foster transparent collaboration between humans and AI agents, and accelerate the development of creative MAS across text and image domains.

Abstract

Large language model (LLM)-driven multi-agent systems (MAS) are transforming how humans and AIs collaboratively generate ideas and artifacts. While existing surveys provide comprehensive overviews of MAS infrastructures, they largely overlook the dimension of \emph{creativity}, including how novel outputs are generated and evaluated, how creativity informs agent personas, and how creative workflows are coordinated. This is the first survey dedicated to creativity in MAS. We focus on text and image generation tasks, and present: (1) a taxonomy of agent proactivity and persona design; (2) an overview of generation techniques, including divergent exploration, iterative refinement, and collaborative synthesis, as well as relevant datasets and evaluation metrics; and (3) a discussion of key challenges, such as inconsistent evaluation standards, insufficient bias mitigation, coordination conflicts, and the lack of unified benchmarks. This survey offers a structured framework and roadmap for advancing the development, evaluation, and standardization of creative MAS.

Paper Structure

This paper contains 57 sections, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. MAS Workflow and Proactivity
  3. MAS Workflow
  4. Spectrum of Agent Proactivity
  5. Planning
  6. Process
  7. Decision Making
  8. Creativity Analysis on Proactivity
  9. MAS Techniques for Creativity
  10. Divergent Exploration
  11. Co-GPT Ideation
  12. Group-AI Brainwriting
  13. ICCRI
  14. Long-Term Impact Study
  15. Iterative Refinement
  16. ...and 42 more sections

Figures (3)

  • Figure 1: Overview of multi‑agent creativity systems. Given user inputs in text or image form, agents engage in a three‑stage process: Planning, Process, and Decision Making, using a variety of techniques (Sec. \ref{['sec:mas-techniques-for-creativity']}) and persona (Sec. \ref{['sec:Persona and Agent Profile']}), with outputs evaluated both subjectively by humans and objectively by automated metrics (Sec. \ref{['sec:Evaluation']}).
  • Figure 2: MAS frameworks positioned along a two-dimensional spectrum reflecting levels of proactivity in Process and Decision-Making. The Planning phase is omitted here due to consistently low agent proactivity in existing literature. Details of proactivity categorization criteria are shown in Appendix \ref{['spectrum details']}.
  • Figure 3: Categories of Persona Granularity: A conceptual framework illustrated with selected attributes, accompanied by a concise example representing each defined persona.