In adversarial environments, especially within highly dynamic real-time strategy (RTS) game scenarios, AI agents powered by large language models (LLMs) still face significant challenges in executing fine-grained tactical decision-making. Although LLMs exhibit strong language-based reasoning capabilities, they often lack real-time responsiveness and contextual adaptability when dealing with unit-level operations such as movement planning and attack prioritization. This study focuses on enhancing the strategic generation capabilities of LLM-driven agents in such complex tasks by integrating expert policies with multi-chain-of-thought (MCoT) reasoning. The aim is to strengthen their situational awareness and strategic coherence, thereby improving the accuracy, interpretability, and flexibility of decision-making.

This study proposes a method called AE-MCoT (adversarial strategy generation integrating expert policies and multi-chain-of-thought reasoning), which integrates multimodal inputs, expert policies, and MCoT reasoning to enhance the fine-grained decision-making capabilities of large language models (LLMs) in highly dynamic adversarial environments. The core of AE-MCoT lies in combining temporally segmented expert strategies with a parallel reasoning framework that explores multiple tactical paths. The expert policy module injects tactical priors into the LLM via natural language prompts embedded in the system prompt, enabling phase-focused reasoning throughout the game. Based on prior experiments and typical game rhythms, the gameplay is divided into three phases: early (0-6s), mid (6-20s), and late (>20s), corresponding respectively to goals such as evading early enemy contact, balancing offense and defense, and preserving unit survival at low health. These phase-specific strategies are marked with EARLY, MID, and LATE labels and dynamically invoked based on the in-game timestamp. Guided by these prompts, the model generates three distinct reasoning paths: aggressive (attack-first), conservative (survival-first), and balanced (offense-defense tradeoff), based on current environmental observations. A self-assessment mechanism is then employed to evaluate the effectiveness of each chain and select the optimal strategy for generating specific move-and-attack actions. Without the need for additional training, AE-MCoT enables high-precision, interpretable, and temporally coherent strategy generation, making it particularly suitable for fast-paced and complex adversarial scenarios.

In the high-difficulty custom StarCraft Ⅱ scenario “1 Colossus vs. 32 Zerglings,” the AE-MCoT method demonstrated strong effectiveness. The full approach integrating expert policies and MCoT reasoning achieved a 95% win rate across 20 matches, surpassing both the single-chain and expert-ablated variants (each at 55%) and far outperforming the no-chain setting (5%). It reached a high kill-to-loss ratio (629), showing clear advantages in unit survivability and enemy suppression. Average kills per game reached 31.45, with 0.32 health ratio remaining upon victory, highlighting the model’s balance of offense and caution. In contrast, the chain-ablated version performed worst, with only 17.4 kills on average, the shortest match duration, and near-zero health retention. Under different enemy scales, AE-MCoT maintained robustness, achieving 65% wins against 40 enemies and 100% against 24, while the baseline agent failed in all cases. In an optimal match, the agent adjusted dynamically through early, mid, and late phases—initially evading encirclement, maintaining steady offense in midgame, and executing high-ground kiting for full elimination. The composite performance score (0.782) further confirms the method’s superior efficiency, survivability, and pacing in high-intensity fine-grained adversarial scenarios.

The results demonstrate the effectiveness of combining expert policies with MCoT reasoning for adversarial strategy generation. The proposed method improves the agent's ability to generate adaptive, fine-grained control strategies without additional training, making it a valuable tool for complex decision-making tasks. The integration of expert policies with MCoT reasoning ensures high adaptability and precision, enabling the model to adjust its strategy based on dynamic game conditions. This approach shows significant promise not only in RTS games but also in other complex, adversarial environments such as autonomous systems and military applications. Moreover, it highlights the potential for improving LLM-based agents' performance in highly dynamic adversarial environments without the need for retraining or extensive datasets.

Large language models (LLMs) have expanded beyond natural language processing into decision-making domains like military strategy and real-time gaming, yet face critical challenges in adversarial environments. Existing methods prove inadequate: free-generation mode lacks structured reasoning for tactical analysis, while chain-of-thought (CoT) generates overly detailed steps causing computational overhead. Additionally, LLMs’ hallucination issues and lack of domain-specific constraints often yield suboptimal decisions in high-stakes scenarios. This study develops the dynamic chain-of-reasoning-and-decision (DCoRD) framework to address these limitations by combining structured reasoning with dynamic output constraints specifically designed for adversarial decision-making. Our objectives are threefold: (1) create an efficient reasoning architecture preserving CoT’s interpretability while eliminating its inefficiencies; (2) develop constraint mechanisms ensuring feasible, effective decisions; and (3) establish performance benchmarks using StarCraft Ⅱ’s complex environment, which provides ideal evaluation metrics for testing LLM-based decision systems under time pressure and competitive conditions. The research aims to bridge the gap between LLMs’ potential and their practical application in professional adversarial scenarios where accuracy, speed, and reliability are paramount.

The DCoRD framework introduces a two-step Markov reasoning architecture comprising state reasoning and strategy reasoning. This architecture is enhanced by a carefully designed dynamic decision options library, which provides standardized semantic units for tactical situation assessment and action selection. Crucially, DCoRD extends the conventional CoT triplet prompt into a quadrilateral structure—integrating input, reasoning framework, dynamic options library, and constrained output—as structured few-shot exemplars. Unlike traditional CoT methods that generate overly detailed intermediate steps, DCoRD strategically constrains outputs to predefined, domain-specific options (e.g., “[Terrain]<High Ground>” or “Attack<Enemy Unit>”), while permitting concise free-form rationales, implementing a soft constraint mechanism adaptable to real-time dynamics. This approach achieves three key improvements: (1) substantial reduction in token consumption, (2) minimized risk of hallucination, and (3) maintained decision interpretability. The framework dynamically adjusts output complexity based on real-time environmental demands, achieving an optimal balance between reasoning depth and operational responsiveness. Comprehensive experiments were conducted on the challenging StarCraft Ⅱ multi-agent challenge (SMAC) platform, testing performance across scenarios of varying complexity (2s_vs_1sc and 2c_vs_64zg). The evaluation compared DCoRD against three established baseline methods: Few-Shot, CoT, and chain-of-draft (CoD), with all methods using identical experimental setups and the same LLM backbone (DeepSeekV3) for fair comparison.

The experimental results demonstrate DCoRD’s significant advantages in both decision quality and operational efficiency. In the particularly demanding 2c_vs_64zg scenario, DCoRD achieved an impressive 50% win rate, representing a 40% absolute improvement over traditional CoT methods. The framework also showed superior tactical effectiveness, increasing average enemy eliminations by 41% (from 36.8 to 51.9 units per match). Detailed analysis revealed that these performance gains stem from DCoRD’s ability to focus computational resources on critical decision factors while filtering out irrelevant reasoning branches. Efficiency metrics were equally strong, with DCoRD reducing token usage by 37.5% (from 333.52 to 208.39 tokens per decision) and decreasing response delay by 14.2% compared to CoT. These improvements were achieved without sacrificing decision transparency, as the structured options framework actually enhanced the interpretability of the model’s reasoning process. Additional analysis of decision patterns showed that DCoRD consistently made better use of terrain advantages and unit positioning compared to baseline methods.

This work establishes that structured output constraints and domain-optimized reasoning frameworks are critical for successful LLM deployment in adversarial environments. The DCoRD methodology represents a significant advance in prompt engineering for decision-making tasks, with several key contributions: (1) a novel two-step reasoning architecture specifically designed for dynamic environments, (2) a flexible options library system that balances expressiveness with computational efficiency by allowing manual design or RAG-based dynamic expansion, and (3) empirical validation confirming that moderate output constraints yield better tactical alignment than lengthy reasoning chains. The framework’s flexible design allows for straightforward adaptation to other high-stakes decision domains including military simulations and emergency response planning. Future research directions will focus on two areas: (1) developing automated methods for expanding the options library, (2) extending the framework to support multi-agent cooperation scenarios.