Society of Mind Meets Real-Time Strategy: A Hierarchical Multi-Agent Framework for Strategic Reasoning

Response to Reviewer W96F (1/2)

We thank Reviewer W96F for the constructive feedback and positive comments on our HIMA framework, experimental results, and clear presentation. For additional experiments, we focus on the Protoss (player) vs. Zerg (opponent) matchup, the most commonly studied scenario in prior work [1,2,3]. We hope our response addresses your concerns.

The number of trials for each experiment is not mentioned. Please specify the trial number.

$\to$ As noted, the number of trials is already specified in the manuscript (L207, L228, L233): 50 randomly sampled matches for the main evaluation (Section 4.1, L207), 10 consecutive games for HIMA vs. baselines (L228), and 20 randomly sampled matches for detailed analysis (Section 4.4, L233). For clarity, we will also state the number of trials in all relevant table and figure captions in the revised manuscript. Thank you for your helpful feedback!

Including human evaluation in this paper to highlight the upper capability boundary of this environment will increase the contribution and the significance of this proposed framework.

$\to$ Good suggestion! For an upper boundary, we first conduct a human evaluation for human vs. built-in AI in the Protoss (human) vs. Zerg (AI) matchup, and their win rates are summarized in the table below. Experimental results with various human players (novice, gold, and platinum) show that platinum and gold-level players consistently outperform the built-in AI, setting an upper bound for agent performance. Here, MMR refers to the matchmaking rating used in StarCraft II to measure a human player's skill level.

We also evaluate HIMA against human players in the Protoss (HIMA) vs. Zerg (human) matchup. HIMA consistently outperforms novice and gold-level players, and performs on par with platinum-level players, demonstrating the practical effectiveness and significance of our framework.

Response to Reviewer crGH

We thank Reviewer crGH for the constructive feedback and positive comments on our use of LLM agents for embodied AI in StarCraft II, the effectiveness of our pipeline, and the experimental validation of our approach. For additional experiments, we focus on the Protoss (player) vs. Zerg (opponent) matchup at difficulty level 7, the most commonly studied scenario in previous work [1,2,3]. We hope our response addresses your concerns.

It is unclear how well this architecture would scale outside the simulation environment or handle more complex physical tasks.

$\to$ HIMA is designed to scale beyond simulation and handle complex real-world tasks, thanks to its hierarchical and modular architecture. Although our current experiments focus on StarCraft II, the core principles underlying HIMA—such as specialized agent collaboration, imitation-based learning, and adaptive planning—directly address challenges common in physical environments, including partial observability and dynamic changes. These strengths make HIMA well-suited for applications like multi-robot coordination and physical system control. We discuss scalability and real-world applicability further in the revised manuscript.

The paper is not well-formulated and very hard to follow. Many terms are used without well definition.

$\to$ We appreciate your feedback and see the review process as a chance to clarify any unclear aspects of the paper. If you can specify which terms or sections are hard to follow, we will gladly address them in the review process. Your detailed input will help us make the paper clearer and more accessible.

The idea basically is not very novel. Actually, the idea is not an exact imitation, actually is role-wise fine-tuning (SFT), which is not new in the literature.

$\to$ The novelty of our work lies not in the basic use of imitation learning or role-wise fine-tuning (SFT), but in our hierarchical, multi-agent structure and meta-level orchestration using a Strategic Planner (SP). While standard imitation learning and SFT struggle to adapt in novel or dynamic scenarios, our framework addresses this limitation by introducing a SP that coordinates multiple specialized imitation agents (L44 and L59, Sec. 3). This planner integrates agent outputs, incorporates environment-aware reasoning, and leverages temporal Chain-of-Thought planning for effective long-term planning. As shown in Tables 2 and 4 and Figure 4, this design enables robust adaptability and strategic coherence in complex environments, clearly distinguishing our method from prior work.

The work only uses one base model (Qwen-2 1.5B) in the evaluation. It is suggested to include more base models.

$\to$ Thank you for your suggestion! We show results w.r.t various small agent models in the table below. Results show that all tested models achieve similar performance in our setting, even though their performance differences may be more noticeable on other benchmarks. We believe this is because all models are trained on the same imitation learning data, which minimizes performance gaps between them.

[1] Li et al., Hierarchical Expert Prompt for Large-Language-Model …, arXiv, 2025.
[2] Ma et al., Large language models play starcraft ii …, NerIPS, 2024.
[3] Wu et a., LLMs Are Not Good Strategists …, ICLR workshop, 2025.\

Response to Reviewer RWbE (1/2)

We thank Reviewer RWbE for the constructive feedback and positive comments on our method’s strong performance, clear ablation studies, and comprehensive benchmark expansion. For additional experiments, we focus on the Protoss (player) vs. Zerg (opponent) matchup at difficulty Level 7, which is the most commonly studied scenario in prior work [3,4,5]. We hope our response addresses your concerns.

The method parts are difficult to understand, particularly the Nominal Group Technique, Strategic Objective, and SC2EGSet.

$\to$ The Nominal Group Technique is a structured decision-making principle we apply to aggregate proposals from multiple agents, as described in [1] (L155) and shown in Figures 3-(b), 14, and 15. This process has five stages: generating agreed and conflicted viewpoints, resolving conflicts, generating isolated viewpoints, and formulating the final strategy (Figures 14 and 15). The Strategic Objective is the high-level goal assigned to each specialized agent based on expert gameplay analysis (L158), with examples in Appendix K and Figure 10. SC2EGSet, a dataset of professional StarCraft II replays with game states and actions [2], is used for training and evaluation, as detailed in Figure 3 (left) and Appendix E.1. We will clarify these in the revised version.

The problem is not well motivated and the design decisions seem arbitrary at points.

$\to$ Our work is motivated by the challenges of real-time strategy games like StarCraft II, where existing LLM-based methods [3,4,5,6] struggle with strategic coherence and adaptability in dynamic, partially observable environments (see L37 and L44). To address these limitations, HIMA incorporates an imitation multi-agent component to support diverse and robust long-horizon planning, combined with a Strategic Planner that enables adaptive, environment-aware decision-making within a hierarchical structure.

No error bars or significance tests.

$\to$ We conduct statistical tests to compare HIMA’s win rates with previous methods in Table 2. In Protoss (player) vs. Zerg (opponent) at difficulty 7, HIMA achieves an 82% win rate (50 trials, L207), which is statistically significant compared to HEP’s 25% (12 trials in [3], p ≈ 0.00015). At difficulty 6, HIMA’s 84% win rate (50 trials) is also statistically significant compared to TextStarCraft’s 8% (12 trials in [4], p-value ≈ 0.00000117) and EpicStar’s 30% (40 trials in [5], p-value ≈ 0.00000000344).

In the Zerg (player) vs. Terran (opponent) matchup, the difference between SwarmBrain [6] and HIMA is not statistically significant, likely due to limited samples. Increasing HIMA’s trials to 150 yields an 84% win rate (±1.5 standard deviation) at difficulty 5; although not statistically significant, HIMA’s average remains higher than SwarmBrain’s 76%, given the observed variance.

Across all detailed analyses (Tables 4 and 5, each with 20 repetitions), HIMA consistently shows statistically significant improvements (all p < 0.05).

Response to Reviewer bnMe

We thank Reviewer bnMe for the constructive feedback and positive comments on our method’s strong empirical results, the introduction of the TEXTSCII-ALL benchmark, and our novel use of LLMs to enhance datasets with tactical rationale. We also appreciate your recognition of our framework’s ability to output action sequences, leverage clustering for agent specialization, and improve efficiency through reduced LLM calls. We hope our response addresses your comments and concerns.

This paper is limited to a single benchmark (StarCraft II), raising questions about the generalizability of the findings. Future work should explore the application of HIMA in other real-time strategy games to validate its effectiveness across different contexts.

$\to$ Thank you for your feedback. We recognize that our experiments are limited to StarCraft II and agree that testing HIMA in other environments is important for demonstrating its generalizability. As future work, we plan to apply HIMA to additional RTS environments such as Civilization [1], poker games [2], Minecraft [3] to evaluate its generalizability. We believe that HIMA’s hierarchical coordination, imitation-based learning, and environment-aware planning are broadly applicable to complex, sequential decision-making tasks, not only in games but also in domains like multi-agent robotics and industrial automation.

[1] Allen et al., "Artificial Intelligence and Civilization: Using Deep Reinforcement Learning to Play Civilization VI," AAAI Workshop on Artificial Intelligence for Strategy Games, 2021.
[2] Zhuang et a., "PokerBench: Training Large Language Models to Become Professional Poker Players," AAAI, 2025.
[3] Guss et al., "The MineRL Competition on Sample Efficient Reinforcement Learning using Human Priors," NeurIPS Competition Track, 2019.