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

Researchers have introduced TextQuests, a new benchmark testing large language models (LLMs) on 25 classic text-based video games. This evaluation measures how well AI agents can reason over long contexts and learn through exploration without external tools. Results show current models struggle with spatial reasoning, context management, and efficient planning when navigating these complex environments that require hundreds of precise actions to complete.