Imagine an AI trying to navigate a virtual house, not with computer vision, but purely by thinking in words. That’s the challenge posed by LangSuit·E, a new testing ground for large language models (LLMs). Researchers wanted to see how well LLMs could act as embodied agents – AI that can perceive and interact with an environment – using only text descriptions. LangSuit·E presents six everyday tasks, like rearranging furniture or answering questions about the location of objects, within a simulated text world. Traditional LLMs often struggle with these embodied tasks because they lack the spatial reasoning abilities of humans. To tackle this, the researchers developed a clever trick called EmMem, or Embodied Memory. EmMem essentially gives the LLM an internal “map” by prompting it to continuously summarize its location and the surrounding environment in words. This helps the AI keep track of its movements and actions within the virtual space. The results were promising. When using EmMem, LLMs like GPT-3.5 performed significantly better in complex, multi-step tasks, demonstrating an improved ability to ‘think’ spatially. The LLM could better keep track of its location and remember past actions, leading to more effective problem-solving. While there's still room for improvement, LangSuit·E and EmMem offer exciting progress in the quest to create more adaptable and intelligent AI agents. This research opens doors to exciting possibilities. Imagine AI assistants that can truly understand and interact with the real world, from navigating complex environments to carrying out intricate tasks, all through the power of language.