This forum post introduces and explains the survey paper "Creativity in LLM-based Multi-Agent Systems: A Survey" (arXiv:2505.21116v1) from National Taiwan University, describing how multiple collaborating AI agents can outperform a single model in creative tasks.
Team vs. Solo: The Creativity Gap
A single AI can get stuck in fixed thinking patterns, while a multi-agent system (MAS) works like a professional team:
- Role division: each agent has a specialty (e.g., writer, designer, evaluator)
- Dynamic collaboration: agents debate, give feedback, and iteratively refine solutions
- Emergent creativity: outputs that even human designers did not anticipate
- Objective metrics (fast screening): Distinct-n and Self-BLEU for text; FID and TIE for images
- Subjective evaluation: TTCT (fluency, flexibility, originality, elaboration) and task-specific criteria
- User studies: questionnaires like the Creativity Support Index (CSI) and behavioral analysis. Notably, CoQuest data shows user satisfaction drops ~30% when AI suggestions become too proactive.
- Open-source repository: https://github.com/MiuLab/MultiAgent-Survey
- Evaluation tools: TTCT, Creativity Support Index
- Case studies: HoLLMwood screenplay system, CollabStory collaborative writing
> Example: In the HoLLMwood system, three AI roles (Writer, Editor, Actor) jointly create screenplays whose character development and narrative coherence significantly exceed what a single AI produces.
Three Core Techniques for AI Team Creativity
1. Divergent Exploration — brainstorm from many angles before converging. In *Group-AI Brainwriting*, students ideate first, then GPT-3 expands ideas, AI agents refine collaboratively, and GPT-4 evaluates; 75% of final proposals contained AI-contributed ideas. 2. Iterative Refinement — polish solutions repeatedly. *Baby-AIGS-MLer* splits machine learning research tasks into ideation → coding → testing → evaluation, handled by different agents in relay, improving benchmark accuracy by 20%. 3. Collaborative Synthesis — assemble fragments into complete works. *CollabStory* has agents take turns writing story sections with GPT-4 judging coherence; 75% of transitions were rated highly coherent.
Agent Persona Design
Persona granularity matters:
| Granularity | Description | Example | |---|---|---| | Coarse | role label only (e.g., "designer") | Solo Performance Prompting | | Medium | domain knowledge + toolkits | HoLLMwood's writer role | | Fine | full personality profile + psychological traits | PersonaFlow's research expert team |
Personas are double-edged: too fine may constrain creativity, too simple yields shallow output. Best practice is dynamic adjustment of persona granularity per task.
Evaluating AI Creativity
Five Open Challenges
1. Balancing AI initiative — too proactive erodes user control; too passive reduces AI to a tool. Future: dynamically calibrated systems. 2. Persona bias — gender/cultural biases in training data get amplified; MALIBU benchmark aims to quantify this. 3. Conflict management — agent debates can spiral; controlled mechanisms like the MAD framework supervise debates. 4. Lack of unified evaluation — different fields use incomparable metrics; MultiAgentBench proposes a common benchmark platform. 5. Copyright — legal frontier treats AI as a creator "somewhere between puppet and actor."
Conclusion: Human-AI Creative Symbiosis
The survey's central finding: multi-agent systems are not meant to replace humans but to serve as a "second brain." In human-AI co-creativity studies, 15 creators described the collaboration as "having another thinking partner." As agents learn role-play, dynamic debate, and iterative refinement, humans are freed from mechanical creation to focus on higher-level creative decisions.