TextQuests: How Good are LLMs at Text-Based Video Games?

The study evaluates the performance of various large language models (LLMs) in text-based video games, specifically analyzing their ability to understand and navigate complex narratives and decision-making scenarios. Key findings include benchmark results indicating that models like GPT-3 and GPT-4 achieve higher success rates in game completion compared to earlier models, with notable improvements in contextual understanding and action selection. This research is significant for practitioners as it highlights the potential of LLMs to enhance interactive storytelling and decision-making systems in game development.