Ice-kun: A Virtual Pet on a Low Energy Display Controlled by a Large Language Model

This paper presents “Ice-kun,” a virtual pet controlled by a large language model (LLM) and displayed on addressable light-emitting diode (LED) boards. Unlike virtual pets with limited interactions, Ice-kun leverages an LLM to interpret natural language input and generate countless playful animations. Ice-kun is configured into several parts to generate the animations. The LLM is tasked with generating animations for these parts. Ideally, as a real-time control agent for virtual pets, an LLM is expected to perform quickly and respond in the correct format. To achieve this, two prompting strategies and six prompt engineering techniques are compared. Fast response time is achieved through a prompting strategy that uses multiple shorter prompts, each corresponding to a different part and executed asynchronously in parallel. To ensure the LLM responds with high accuracy in the correct format, few-shot prompting and zero-shot CoT prompting are incorporated into the prompts. Code, experiment results, and demo videos are available in our repository https://github.com/anonymouspenguin27/icekun