TowerMind: A New Benchmark for LLM Agents Using Tower Defense Games

Researchers have introduced TowerMind, a novel environment and benchmark for evaluating Large Language Models (LLMs) as agents, grounded in the tower defense (TD) subgenre of real-time strategy (RTS) games. TowerMind addresses limitations of existing RTS game environments, which either have high computational demands or lack textual observations, by offering low computational requirements and a multimodal observation space. This allows for assessing LLMs' long-term planning and decision-making capabilities, which are crucial for adapting to diverse scenarios. The environment preserves key evaluation strengths of RTS games while being more accessible for LLM testing.