TowerMind: A Tower Defence Game Learning Environment and Benchmark for LLM as Agents

TL;DR

TowerMind presents a lightweight, multimodal tower-defense RTS environment to evaluate large language models on long-term planning and decision-making while enabling hallucination analysis. Built on Unity with Gym-style interfaces, it provides pixel, textual, and structured observations, a hybrid action space, and five progressively difficult levels, all with a GUI level editor for easy customization. Across language-only and vision-language modalities, commercial LLMs outperform open-source models but all lag behind human experts, highlighting persistent gaps in planning and action effectiveness and revealing the benefits and limits of visual input. The work also includes an RL benchmark using Ape-X DQN and PPO, demonstrating TowerMind’s versatility as a benchmark for both LLM-driven and RL-based agents, and it releases the source code for broad adoption.

Abstract

Recent breakthroughs in Large Language Models (LLMs) have positioned them as a promising paradigm for agents, with long-term planning and decision-making emerging as core general-purpose capabilities for adapting to diverse scenarios and tasks. Real-time strategy (RTS) games serve as an ideal testbed for evaluating these two capabilities, as their inherent gameplay requires both macro-level strategic planning and micro-level tactical adaptation and action execution. Existing RTS game-based environments either suffer from relatively high computational demands or lack support for textual observations, which has constrained the use of RTS games for LLM evaluation. Motivated by this, we present TowerMind, a novel environment grounded in the tower defense (TD) subgenre of RTS games. TowerMind preserves the key evaluation strengths of RTS games for assessing LLMs, while featuring low computational demands and a multimodal observation space, including pixel-based, textual, and structured game-state representations. In addition, TowerMind supports the evaluation of model hallucination and provides a high degree of customizability. We design five benchmark levels to evaluate several widely used LLMs under different multimodal input settings. The results reveal a clear performance gap between LLMs and human experts across both capability and hallucination dimensions. The experiments further highlight key limitations in LLM behavior, such as inadequate planning validation, a lack of multifinality in decision-making, and inefficient action use. We also evaluate two classic reinforcement learning algorithms: Ape-X DQN and PPO. By offering a lightweight and multimodal design, TowerMind complements the existing RTS game-based environment landscape and introduces a new benchmark for the AI agent field. The source code is publicly available on GitHub(https://github.com/tb6147877/TowerMind).

Figures (12)

• Figure 1: These are screenshots from four different TowerMind levels. The icons in the four corners of each image display key gameplay information, including the number of player’s current gold coins, player’s base health, and remaining enemy waves. The maps feature irregular, intersecting roads along which enemies advance toward the player's base in successive waves. Players must strategically build different types of towers at designated locations along these roads to repel the incoming enemies. The cloud-shaped white areas represent fog of war, introducing partial observability to the environment.
• Figure 2: Left: A screenshot of the TowerMind environment with key game elements annotated. The coordinate axes illustrate the alignment between the game map and the 2D coordinate system. The red and blue arrowed curves represent the two roads used by enemies to attack. Labels A–G indicate: (A) player's current gold coins; (B) remaining enemy waves; (C) hero unit's current health; (D) player's base health; (E) fog of war; (F) an unbuilt tower point; and (G) gold coins dropped on the map awaiting collection; (H) knights units; (I) an enemy. Right: Illustrations and brief descriptions of the 12 action types.
• Figure 3: The evaluation results on the benchmark levels, with scores normalized relative to the human expert. Error bars represent the standard error.
• Figure 4: Training curves, the horizontal axis shows the number of training samples, measured in millions. Error bars represent 95% confidence intervals.
• Figure 5: Screenshots of the 15 enemy types, with IDs from left to right ranging from 0 to 14.