Inspiration
Our inspiration stems from the groundbreaking Stanford generative agents paper, which demonstrated how AI systems can serve as powerful lenses for understanding human behavior and social dynamics. That seminal work revealed how computational agents, when imbued with memory and personality, naturally exhibit emergent social patterns that mirror real human interactions—from forming relationships to coordinating activities to developing shared narratives. We were captivated by the profound implications this holds for the social sciences: if artificial agents can authentically replicate the complexities of human psychology and sociology, they become invaluable tools for exploring questions that would be difficult or impossible to study with human subjects. This realization sparked our vision for CutiesVote. We recognized that political systems, with their intricate web of persuasion, coalition-building, and strategic decision-making, represent one of the most fascinating domains of human social behavior. By extending the generative agents framework to democratic processes, we could create a living laboratory for understanding how beliefs evolve, how consensus emerges, and how individual psychology shapes collective outcomes. The charm and accessibility of our "cutey AI friends" make these deep insights into human nature approachable and engaging, transforming abstract social science concepts into tangible, observable interactions.
What it does
CutiesVote reimagines voting systems through the lens of adorable AI companions, creating an engaging simulation where artificial agents with distinct personalities participate in democratic processes. The system brings political discourse to life through charming characters that debate, persuade, and ultimately vote on important issues.
How we built it
Our technical stack combines cutting-edge AI infrastructure with modern web technologies. We leveraged vLLM and Verl for efficient language model inference, paired with Qwen-8B as our foundation model and Claude for enhanced reasoning capabilities. The frontend visualization was crafted in ReactJS with custom sprite animations, while our generative agents were carefully designed through iterative prompt engineering to exhibit realistic political behaviors and coherent belief systems.
Challenges we ran into
The development journey presented several significant hurdles. Integrating coding agents into our workflow required careful orchestration and debugging. Training our reinforcement learning models demanded extensive experimentation to achieve stable convergence. Perhaps most challenging was striking the delicate balance between computational efficiency and behavioral realism in our political simulations. Additionally, coordinating the diverse technical components—from RL training pipelines to frontend animations—required meticulous planning and communication across different workstreams.
Accomplishments that we're proud of
We're particularly proud of several breakthrough achievements. Our visualization layer features polished animations that bring the town hall meetings to life with charm and clarity. The AI characters we've created possess genuinely complex personalities with internally consistent belief systems that evolve through interactions. Most exciting are the emergent behaviors we've observed: our agents spontaneously employ sophisticated persuasion techniques, appeal to emotional reasoning, and even engage in strategic deception when advancing their positions. Throughout development, we benefited tremendously from collaborating with our partners at Modal, Cognition, and OpenAI, whose platforms and guidance elevated our project significantly.
What we learned
This project deepened our understanding across multiple domains. We gained hands-on experience orchestrating complex multi-agent simulations where numerous AI entities interact with persistent memory systems. Working with Modal and Devin transformed our development workflow, enabling rapid iteration and deployment. The technical challenges of multi-turn reinforcement learning taught us valuable lessons about training stability and reward shaping. Beyond the technical aspects, we developed a richer appreciation for the nuances of political processes and the intricate dynamics of group decision-making.
What's next for CutiesVote
The future roadmap for CutiesVote is ambitious and exciting. We plan to expand the simulation framework to encompass more diverse scenarios and interaction settings, from local town halls to international summits. Our multi-agent reinforcement learning system will evolve to support increasingly sophisticated strategic behaviors and coalition-building dynamics. We're also committed to scaling up our training infrastructure, leveraging larger models and longer training runs to unlock even more nuanced and realistic political behaviors. Ultimately, we envision CutiesVote becoming a powerful tool for understanding democratic processes through the engaging medium of AI-driven simulation.
Log in or sign up for Devpost to join the conversation.