Learning interactions to boost human creativity with bandits and GPT-4
Project Overview
The document explores the application of generative AI, particularly utilizing a multi-armed bandit algorithm and GPT-4, to enhance creativity in educational settings by aiding users in overcoming cognitive roadblocks during idea generation tasks. The research reveals that AI-generated hints can significantly increase both the diversity and quantity of responses in situations where individuals typically face difficulties in producing ideas. By simulating human interactions with GPT-4, the study assesses the effectiveness of various hinting strategies, concluding that generative AI can effectively augment human creative processes. These findings highlight the potential for AI to serve as a valuable tool in education, fostering creativity and enhancing learning outcomes by supporting students in generating innovative ideas.
Key Applications
Generative AI for Feature Generation and Hinting
Context: Engagement in cognitive tasks involving feature generation, aimed at undergraduate students at the University of Wisconsin-Madison. The tasks explored AI's ability to simulate human-like creative processes, leveraging both multi-armed bandit algorithms and GPT-4 for hint generation and feature production.
Implementation: Participants engaged in a feature fluency task where they could request hints generated by a multi-armed bandit based on prior responses. Simultaneously, GPT-4 was prompted to generate features and request hints, enabling the bandit to learn from its output and improve hint quality.
Outcomes: Both approaches led to increased production of features and greater diversity in the features generated. The AI systems exhibited behaviors similar to human participants, producing comparable outcomes and benefiting from hints.
Challenges: Determining effective hinting strategies and training bandits on human behavior remain complex issues. Additionally, the lack of cognitive constraints in GPT-4 complicates direct comparisons with human performance.
Implementation Barriers
Technical barrier
The need for extensive training data to effectively train AI models on human behavior.
Proposed Solutions: Utilizing simulations with models like GPT-4 to capture patterns of human behavior and refine hinting strategies.
Cognitive barrier
Humans can experience cognitive roadblocks that limit their creativity.
Proposed Solutions: Implementing AI-generated hints to help users overcome these roadblocks and stimulate idea generation.
Project Team
Ara Vartanian
Researcher
Xiaoxi Sun
Researcher
Yun-Shiuan Chuang
Researcher
Siddharth Suresh
Researcher
Xiaojin Zhu
Researcher
Timothy T. Rogers
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Ara Vartanian, Xiaoxi Sun, Yun-Shiuan Chuang, Siddharth Suresh, Xiaojin Zhu, Timothy T. Rogers
Source Publication: View Original PaperLink opens in a new window
Project Contact: Dr. Jianhua Yang
LLM Model Version: gpt-4o-mini-2024-07-18
Analysis Provider: Openai