Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

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

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper tries to establish that hierarchical coordination of specialist imitation agents beats flat LLM planners in StarCraft II, at a fraction of the LLM cost.

desk verdict A solid hierarchical multi-agent system for text-based StarCraft II with a genuinely useful nine-matchup benchmark, but the 'outperforms SOTA' claim rests on quoted baselines rather than same-environment re-runs. read the letter →

arxiv 2508.06042 v1 pith:N7CMZH6G submitted 2025-08-08 cs.AI

Daechul Ahn , San Kim , Jonghyun Choi This is my paper
classification cs.AI
keywords hierarchicalmulti-agentimitationlearningStarCraftIILLMagentsstrategicplanningtemporalchain-of-thoughtreal-timestrategyTEXTSCII-ALL
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper sets out to show that a hierarchy of specialized imitation agents can handle long-horizon strategic reasoning in StarCraft II where flat LLM planners fail. Three small agents, each trained on professional replays grouped by unit composition, emit structured multi-step action plans with tactical rationales; a meta-controller called the Strategic Planner merges those plans into one adaptable decision. The reported payoff is a Protoss-vs-Zerg win rate of 82% against level-7 built-in AI versus 25% for the HEP baseline, while cutting LLM calls from 636 to 11 per 20-minute game. If the comparison holds, the value is practical: coherent strategic play from an LLM does not require dense, expensive querying at every game tick.

What carries the argument

The load-bearing mechanism is the two-level hierarchy itself. At the bottom, imitation agents trained by supervised fine-tuning output temporally structured action sequences for a fixed horizon (Δ=3 minutes) together with a tactical rationale; at the top, the Strategic Planner orchestrates their proposals using Nominal Group Technique resolution, reads rationale and strategic objective per agent, and re-expresses the winning strategy as immediate, short-term, and long-term actions (the 'temporal CoT'). The feedback system (failed-command repair, enemy-force re-planning threshold) is what lets a long-horizon plan stay responsive. Together they convert the planner's role from per-tick decision

What would settle it

Re-run TextStarCraft, SwarmBrain, EpicStar, and HEP inside the paper's TEXTSCII-ALL environment on the same map (Ancient Cistern LE), patch, and LLM backends, with at least 50 games per matchup; if the level-7 Protoss-vs-Zerg win rate of any baseline reaches or exceeds HIMA's 82%, the claimed superiority is not established.

Watch

Extended reading notes

Core claim

On its own terms, HIMA's discovery is that 'society of mind' decomposition changes the planning economy of LLM agents in real-time strategy games. Imitation agents trained on SC2EGSet professional replays generate variable-length action sequences over a three-minute window, annotated with a tactical rationale, instead of single actions per timestep. Clustering the replays by supply-weighted unit composition creates specialists (air-dominant, ground-dominant, hybrid) that cover distinct strategic routes. The Strategic Planner then applies the Nominal Group Technique, reads each specialist's rationale and strategic objective, and produces final actions through a temporal chain-of-thought that

Load-bearing premise

The baseline win rates quoted from earlier papers are treated as directly comparable to HIMA's numbers, even though those baselines were not re-run in the paper's own testbed.

Editorial extensions

If this is right

  • If the comparison is right, LLM-based RTS agents can plan at coarse time scales and reserve dense reasoning for rare conflicts, making real-time deployment practical with ordinary API budgets.
  • The three-minute planning window is a real trade-off reported in the paper: shorter horizons waste calls, longer horizons miss threats, so similar horizons should be tuned for any ported environment.
  • Removing the tactical rationales or the strategic planner's NGT+TR-SO aggregation lowers win rates in the paper's own ablations, so both components are load-bearing for the stated result, not decorations.
  • The nine-matchup benchmark reveals the technique is not uniformly strong: Terran play lags badly, which the paper attributes to micro-control demands outside its macro-focused action space.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An extension the paper leaves implicit: the hierarchy is not StarCraft-specific in principle, so any partially observable sequential task with expert demonstrations and an expensive oracle planner could reuse the same split between cheap imitation specialists and a rare meta-controller.
  • Clustering by final supply-weighted unit composition uses late-game information to label early-game behavior; a temporal clustering that respects information availability at decision time might change which specialists emerge and how cleanly they separate.
  • The head-to-head table reports 100% over only 10 games per opponent; with that sample, the true margin over HEP and TextStarCraft could be much smaller than the table suggests, and a larger tournament would sharpen the estimate.
  • The enemy-threshold trigger (τ=10) is a heuristic; an adaptive threshold learned from replays or tuned per matchup is a direct extension the paper does not test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes HIMA, a hierarchical multi-agent framework for text-based StarCraft II. Three imitation-learning agents (trained by SFT on professional replays from SC2EGSet, clustered by unit composition) each propose structured multi-step action sequences with tactical rationales; a Strategic Planner LLM (GPT-4o-mini) orchestrates these proposals using Nominal Group Technique and a temporal Chain-of-Thought, with a feedback mechanism for failed actions and enemy-threshold triggers. The authors also introduce TEXTSCII-ALL, an extension of TextStarCraft II covering all nine race matchups. Empirical sections report win rates against built-in AI across difficulty levels, head-to-head wins against three prior methods, ablations of the multi-agent design, and a large reduction in LLM call overhead (11 calls / 247 s vs. 636 calls / 7,695 s for TextStarCraft). The central claim is that HIMA outperforms state-of-the-art LLM-based SC2 agents in win rate, adaptability, and efficiency.

Significance. If the empirical claims hold, HIMA is a meaningful step for LLM-based RTS agents: it replaces frequent short-horizon LLM queries with longer-horizon structured plans from small specialized imitation models, coordinated by a meta-controller, and it evaluates across all nine race matchups. The paper is also commendable for releasing code/models/datasets, for including ablations (Table 4, Fig. 4, App. G, App. I), and for attempting a more comprehensive testbed. However, the headline comparative claims depend on baseline numbers taken from other papers without re-running those baselines in the new environment, and the head-to-head evidence is thin. The design ideas are plausible, but the current evaluation does not yet substantiate the 'outperforms state of the arts' claim as stated.

major comments (5)
  1. [Section 4.3, Table 2] The central comparison against TextStarCraft, SwarmBrain, EpicStar, and HEP uses win rates quoted from the original papers, which were obtained in the original TextStarCraft II environment, not in TEXTSCII-ALL. The manuscript does not report re-running any baseline in its own harness, and Appendix C.1 specifies a particular game patch and map without showing the baselines used the same settings. Since action spaces, map, patch, and built-in AI behavior differ, the reported gaps (e.g., HIMA Protoss vs. Zerg 82% at Lv.7 vs. HEP 25%) may reflect environment differences rather than strategic superiority. Please re-run the available baselines in TEXTSCII-ALL or clearly restrict the claim to 'HIMA in TEXTSCII-ALL versus literature-reported numbers.'
  2. [Table 3] Head-to-head results are based on only 10 consecutive games per opponent, use only HIMA's best race (Protoss), and report 100% win rates against each baseline. No confidence intervals or significance tests are given, and the opponent's race is chosen to match each baseline's originally reported setting. With n=10, a 100% result has a wide binomial interval, and matchup selection further weakens the comparison. Please report more games, all HIMA races, and interval estimates, or temper the claim that HIMA 'outperforms all previous approaches.'
  3. [Section 3.2, Fig. 4(c), App. G] Key framework hyperparameters—planning horizon Delta=3 min, number of agents k=3, enemy threshold tau=10—are selected based on the same win-rate numbers that are later reported as results. For example, Fig. 4(c) shows win rate peaking at 3-4 minutes and App. G shows k=3 as the empirical optimum, both on the evaluation set. This selection-on-the-benchmark makes the 'optimal' design claim circular and inflates the apparent effectiveness of the components. Please provide a tuning/validation split, report sensitivity curves with error bars, or state explicitly that these values were chosen on the evaluation set and should be treated as fitted parameters.
  4. [Abstract, Section 1, Section 4] The abstract and contributions claim 'strategic clarity, adaptability, and computational efficiency,' but clarity and adaptability are not defined or measured. Win rate is the main metric; Appendix F reports APU, RUR, PBR, and TR, but these are macro-management statistics, not direct measures of strategic clarity or adaptability. Please either define and quantitatively measure these constructs or remove them from the headline claims.
  5. [Table 6] The efficiency claim compares HIMA's LLM call counts and total call time against TextStarCraft and HEP numbers taken from prior papers. The comparison is not controlled: different environments, different game patches, different LLM versions, and different hardware/API latencies may explain part of the difference. Please re-measure the baselines in the same harness or report the comparison as a literature-number illustration rather than a direct experimental result.
minor comments (5)
  1. [Abstract] Typo: 'exisiting' should be 'existing.' Also 'TEXT SCII-A LL' spacing is inconsistent throughout; choose one form.
  2. [Table 5 header] 'Clutering' should be 'Clustering.'
  3. [Appendix J] 'instuction-tuning' should be 'instruction-tuning.'
  4. [Section 4.4 vs. Tables 4-5] The text says 'running 20 randomly sampled matches' for detailed analysis, but Table 4's last row uses 50 games and Table 5's rows are not annotated with n. Please state the number of games for each row/column consistently.
  5. [Figure 4(c)] The horizontal axis label is cut off in the rendering; ensure the caption clearly states that the x-axis is planning time horizon Delta in minutes and that the left axis is win rate while the right is API calls.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's claims are empirical comparisons, not derivations that reduce to their inputs.

full rationale

The paper contains no formal derivation chain: there are no equations whose consequents are identical to their antecedents, and no parameter is fitted and then renamed as a prediction. The main empirical claim (HIMA outperforms SoTAs) rests on win-rate comparisons in Tables 2, 3, and 6. These comparisons are vulnerable to a validity threat—baseline numbers are quoted from prior papers (Ma et al., 2024; Shao et al., 2024; Wu & Hu, 2025; Li et al., 2025b) and were not re-run in the proposed TEXTSCII-ALL environment (Section 4.3, Table 2 dashes indicate untested conditions). That is a benchmarking / comparability concern, not circularity: the baseline numbers are external evidence, and no step in the paper reduces to those numbers by definition. Similarly, hyperparameters such as the time window (Fig. 4c), number of agents (Appendix G), and enemy threshold (Sec. 3.2, τ=10 heuristically determined) were selected on the same benchmark, which can inflate reported win rates, but this is overfitting/selection, not definitional equivalence. The framework's components (imitation agents, NGT, t-CoT) are evaluated via ablations (Table 4, Fig. 4) that support the hierarchy claim independently. No self-citations are load-bearing; all cited prior work is external. The paper's own limitation statements—e.g., Terran underperformance due to micro demands and inability to compare EpicStar due to code unavailability—are transparent and do not establish circularity. Therefore, the circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

All free parameters were tuned against the same win-rate metric used for final results, which inflates the apparent performance. The main domain assumptions concern the validity of the replay dataset, the simulator, and the small-sample win-rate estimates.

free parameters (3)
  • Planning horizon Delta = 3 minutes
    Chosen based on win-rate vs. horizon experiments (Fig. 4c); peak at 3-4 minutes.
  • Number of specialist agents k = 3
    Validated empirically in Appendix G; 3 clusters gave highest win rates.
  • Enemy-force alarm threshold tau = 10
    Heuristically determined in Sec. 3.2 for triggering re-planning.
assumptions (5)
  • domain assumption SC2EGSet professional replays are representative of expert play and cover the defined action space.
    Section 3.1 and Appendix E.1 extract state-action pairs from SC2EGSet; if replays are unrepresentative, imitation agents cannot learn valid strategies.
  • domain assumption The TEXT SCII-ALL executor faithfully implements StarCraft II prerequisites and race-specific semantics.
    Section 3.2 and Appendix D rely on rule-based prerequisite checks; if the simulator is wrong, win rates reflect simulator bugs rather than strategy.
  • domain assumption Win rate over 20 or 50 matches is a stable estimate of agent strength.
    All tables report point estimates without confidence intervals; small sample sizes may produce noisy rankings.
  • domain assumption GPT-4o-mini generated tactical rationales are accurate and beneficial for imitation training.
    Appendix J compares with/without rationales, but the rationales were machine-generated and not validated against expert annotation.
  • ad hoc to paper Unit-composition clusters correspond to coherent, distinct strategic personas that the Strategic Planner can arbitrate.
    This is the core design premise of the multi-agent splitting (Sec. 3.1, Appendix E.2); Table 5 shows it outperforms alternatives empirically, but the persona interpretation is not independently verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Society of Mind Meets Real-Time Strategy: A Hierarchical Multi-Agent Framework for Strategic Reasoning." pith.science (2026). https://pith.science/paper/N7CMZH6G

@misc{pith2026250806042,
  author       = {Pith},
  title        = {Pith review of: Society of Mind Meets Real-Time Strategy: A Hierarchical Multi-Agent Framework for Strategic Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N7CMZH6G}},
  note         = {Machine review of arXiv:2508.06042}
}
read the original abstract

Large Language Models (LLMs) have recently demonstrated impressive action sequence prediction capabilities but often struggle with dynamic, long-horizon tasks such as real-time strategic games. In a game such as StarCraftII (SC2), agents need to manage resource constraints and adapt to evolving battlefield situations in a partially observable environment. This often overwhelms exisiting LLM-based approaches. To address these challenges, we propose a hierarchical multi-agent framework that employs specialized imitation learning agents under a meta-controller called Strategic Planner (SP). By expert demonstrations, each specialized agent learns a distinctive strategy, such as aerial support or defensive maneuvers, and produces coherent, structured multistep action sequences. The SP then orchestrates these proposals into a single, environmentally adaptive plan that ensures local decisions aligning with long-term strategies. We call this HIMA (Hierarchical Imitation Multi-Agent). We also present TEXTSCII-ALL, a comprehensive SC2 testbed that encompasses all race match combinations in SC2. Our empirical results show that HIMA outperforms state of the arts in strategic clarity, adaptability, and computational efficiency, underscoring the potential of combining specialized imitation modules with meta-level orchestration to develop more robust, general-purpose AI agents.

Figures

Figures reproduced from arXiv: 2508.06042 by the authors.

Figure 1
Figure 1. Comparison of short-horizon (Ma et al., 2024) vs. proposed long-horizon (HIMA) planning in SC2. Fewer LLM calls, more coherent build orders, and higher win rates in HIMA. (a) Existing methods frequently produce one-step actions at short intervals (e.g., 00:00, 00:01, 00:02, . . . ), causing repeated invalid actions (X) due to unmet prerequisites. In contrast, HIMA produces structured multi-step plans with fewer quer… view at source ↗
Figure 2
Figure 2. Overview of the proposed hierarchical imitation multi-agent (HIMA) frame￾work. Each specialized imitation agent (e.g., Air Superiority, Infrastructure, Ground Superi￾ority) receives up-to-date player state information (resources, units, buildings) and produces a multi-step action plan along with a rationale explaining its strategic intent. A high-level meta-controller (i.e., Strategic Planner) then merges these agen… view at source ↗
Figure 3
Figure 3. Dataset construction pipeline for generating structured action sequence and environment-aware action orchestration in HIMA. (a) We extract state-action pairs {St , At} from SC2EGSet (Białecki et al., 2023) and prompt an LLM to generate a Tacti￾cal Rationale (TR) for each multi-step action sequence, At:t+∆. This TR is appended to form an instruction-tuning dataset, capturing what actions occur and why they’re chosen.… view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Detailed analyses of the proposed HIMA. (a) Enhanced aggregation methods improve performance, (b) t-CoT effectively bridges short- and long-term strategies and (c) optimal planning horizon balances reactivity and efficiency. The opening strategy uses public metadata1 i…
Figure 5
Figure 5. Figure 5: Overview of data construction process. We use SC2EGSET to obtain player’s current state and corresponding action sequence. In particular, the number of units in each current state is determined by separately recording unit counts at each time step. In order to convert …
Figure 6
Figure 6. Figure 6: Action space of three races in StarCraft II. Each race has actions related to unit production, building construction, and technology development. Both the imitation agent and the strategic planner form sequences of these race-specific actions to issue commands. 17 [PI…
Figure 7
Figure 7. Figure 7: Prompt Details for Rational Generation in an Instruction-Tuning Dataset. We use GPT-4o-mini to generate the rationale behind the chosen action sequence for each situation. Subsequently, the generated rationale and the corresponding action sequence are combined into a s…
Figure 8
Figure 8. Figure 8: Number of actions generated per time interval. We compare the number of actions generated per time interval across 68 games from the Star League 7 tournament held in 2021. We observe that early stages produce fewer actions due to limited resources and tech prerequisite…
Figure 9
Figure 9. Figure 9: Agent response diversity across different clustering criteria. We measure the diversity of multi-agent outputs across opening strategy, advancement tempo, and unit composition criteria. J Performance Impact of Rationales in the Instruction-Tuning Dataset In [PITH_FULL…
Figure 10
Figure 10. Figure 10: Prompt for generating strategic objectives and strategic objective prompt for each race. We identify the underlying strategic objective of each imitation agent based on its unit ratio. L Various Open-source Models for Imitation Learning Agent We conduct additional exp…
Figure 11
Figure 11. Figure 11: System prompt for the Protoss race. It consists of two parts: one that guides each imitation agent’s strategic objective, and another that supports the reasoning process of the strategic planner. For other races, the system prompt remains unchanged, except for the lis…
Figure 12
Figure 12. Figure 12: Input example for the Protoss race. The imitation agent’s input consists solely of the current game state. The strategic planner’s input includes not only the current game state but also the imitation agent’s output and feedback from the environment. 25 [PITH_FULL_IM…
Figure 13
Figure 13. Figure 13: Output example for the Protoss race. We demonstrate that the strategic planner’s output currently includes only its decisions, but by modifying the system prompt, we can incorporate the rationale into the output as well. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_13.png]
Figure 14
Figure 14. Figure 14: shows an infeasible action being received, while [PITH_FULL_IMAGE:figures/full_fig_p027_14.png]
Figure 15
Figure 15. Figure 15: Overall Strategy Generation Pipeline in a Defense Scenario. Strategy planner (SP) urgently produces counter units (Void Rays, Zealots) in response to the enemy’s attack, and also constructs defensive buildings (Photon Cannons) to prepare for subsequent assaults. 28 […

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Weak-Link Optimization for Multi-Agent Reasoning and Collaboration

    cs.AI 2026-04 unverdicted novelty 7.0 of 10

    WORC improves multi-agent LLM reasoning to 82.2% average accuracy by predicting and compensating for the weakest agent via targeted extra sampling rather than uniform reinforcement.

Reference graph

Works this paper leans on

50 extracted references · 27 canonical work pages · cited by 1 Pith paper

  1. [1]

    Apprenticeship learning via inverse reinforcement learning

    Pieter Abbeel and Andrew Y Ng. Apprenticeship learning via inverse reinforcement learning. In ICML, 2004

    work page 2004

  2. [2]

    Avalon: A benchmark for rl generalization using procedurally generated worlds

    Joshua Albrecht, Abraham Fetterman, Bryden Fogelman, Ellie Kitanidis, Bartosz Wr \'o blewski, Nicole Seo, Michael Rosenthal, Maksis Knutins, Zack Polizzi, James Simon, et al. Avalon: A benchmark for rl generalization using procedurally generated worlds. In NeurIPS, 2022

    work page 2022

  3. [3]

    The arcade learning environment: An evaluation platform for general agents

    Marc G Bellemare, Yavar Naddaf, Joel Veness, and Michael Bowling. The arcade learning environment: An evaluation platform for general agents. Journal of artificial intelligence research, 2013

    work page 2013

  4. [4]

    Sc2egset: Starcraft ii esport replay and game-state dataset

    Andrzej Białecki, Natalia Jakubowska, Paweł Dobrowolski, Piotr Białecki, Leszek Krupiński, Andrzej Szczap, Robert Białecki, and Jan Gajewski. Sc2egset: Starcraft ii esport replay and game-state dataset. Scientific Data, 2023

    work page 2023

  5. [5]

    Starcraft II

    Blizzard Entertainment . Starcraft II . Video game (PC), 2010

    work page 2010

  6. [6]

    Do as i can, not as i say: Grounding language in robotic affordances

    Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, et al. Do as i can, not as i say: Grounding language in robotic affordances. In CORL, 2022

    work page 2022

  7. [7]

    Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020

    work page 2020

  8. [8]

    Understanding individual agent importance in multi-agent system via counterfactual reasoning

    Jianming Chen, Yawen Wang, Junjie Wang, Xiaofei Xie, Qing Wang, Fanjiang Xu, et al. Understanding individual agent importance in multi-agent system via counterfactual reasoning. arXiv preprint arXiv:2412.15619, 2024 a

    arXiv 2024

Show all 50 references
  1. [9]

    Reconcile: Round-table conference improves reasoning via consensus among diverse llms

    Justin Chen, Swarnadeep Saha, and Mohit Bansal. Reconcile: Round-table conference improves reasoning via consensus among diverse llms. In ACL, 2024 b

    2024

  2. [10]

    Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors in agents

    Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chen Qian, Chi-Min Chan, Yujia Qin, Yaxi Lu, Ruobing Xie, et al. Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors in agents. arXiv preprint arXiv:2308.10848, 2023

    2023 arXiv

  3. [11]

    Delbecq and Andrew H

    Andr \'e L. Delbecq and Andrew H. Van de Ven. A group process model for problem identification and program planning. Journal of Applied Behavioral Science, 1971

    1971

  4. [12]

    PaLM-E : An embodied multimodal language model

    Danny Driess, Fei Xia, Mehdi Sajjadi, Corey Lynch, Aakanksha Chowdhery, Jonathan Hoffman, Yue Hu, Sergey Levine, Vincent Vanhoucke, Quan Vuong, et al. PaLM-E : An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023

    2023 arXiv

  5. [13]

    Tenenbaum, and Igor Mordatch

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. In ICML, 2023

    2023

  6. [14]

    MineDojo : Building open-ended embodied agents with internet-scale knowledge

    Linxi Fan, Alane Suhr Xie, Ziyu Jiang, Alice H To, Shuang Li Yao, Marios Skreta, Vincent Yu, Yitao Bai, Zifan Wang, Kurt Shuster, et al. MineDojo : Building open-ended embodied agents with internet-scale knowledge. In NeurIPS, 2022

    2022

  7. [15]

    Kanishk Gandhi, Dorsa Sadigh, and Noah D. Goodman. Strategic reasoning with language models. arXiv preprint arXiv:2305.19165, 2023

    2023 arXiv

  8. [16]

    Pokellmon: A human-parity agent for pokemon battles with large language models

    Sihao Hu, Tiansheng Huang, and Ling Liu. Pokellmon: A human-parity agent for pokemon battles with large language models. arXiv preprint arXiv:2402.01118, 2024

    2024 arXiv

  9. [17]

    Camel: Communicative agents for" mind" exploration of large language model society

    Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for" mind" exploration of large language model society. In NeurIPS, 2023

    2023

  10. [18]

    Jarvis-vla: Post-training large-scale vision language models to play visual games with keyboards and mouse

    Muyao Li, Zihao Wang, Kaichen He, Xiaojian Ma, and Yitao Liang. Jarvis-vla: Post-training large-scale vision language models to play visual games with keyboards and mouse. arXiv preprint arXiv:2503.16365, 2025 a

    2025 arXiv

  11. [19]

    Hierarchical expert prompt for large-language-model: An approach defeat elite ai in textstarcraft ii for the first time

    Zongyuan Li, Chang Lu, Xiaojie Xu, Runnan Qi, Yanan Ni, Lumin Jiang, Xiangbei Liu, Xuebo Zhang, Yongchun Fang, Kuihua Huang, and Xian Guo. Hierarchical expert prompt for large-language-model: An approach defeat elite ai in textstarcraft ii for the first time. arXiv preprint ar...

    2025 arXiv

  12. [20]

    Multi-expert prompting improves reliability, safety and usefulness of large language models

    Do Long, Duong Yen, Luu Anh Tuan, Kenji Kawaguchi, Min-Yen Kan, and Nancy Chen. Multi-expert prompting improves reliability, safety and usefulness of large language models. In EMNLP, 2024

    2024

  13. [21]

    Large language models play starcraft ii: Benchmarks and a chain of summarization approach

    Weiyu Ma, Qirui Mi, Yongcheng Zeng, Xue Yan, Runji Lin, Yuqiao Wu, Jun Wang, and Haifeng Zhang. Large language models play starcraft ii: Benchmarks and a chain of summarization approach. In NeurIPS, 2024

    2024

  14. [22]

    The Society of Mind

    Marvin Minsky. The Society of Mind. Simon and Schuster, New York, NY, USA, 1986. ISBN 978-0-671-65713-0

    1986

  15. [23]

    Gpt-4o mini: Advancing cost-efficient intelligence

    OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/, 2024. Accessed: 2025-03-20

    2024

  16. [24]

    Training language models to follow instructions with human feedback

    Xiang Ouyang, Jeffrey Wu, Xu Jiang, Diana Almeida, Carolyn Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Frederick Kelton, Lilian Miller, Christopher Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan...

    2022

  17. [25]

    No-press diplomacy: Modeling multi-agent gameplay

    Philip Paquette, Yuchen Lu, Seton Steven Bocco, Max Smith, Satya O-G, Jonathan K Kummerfeld, Joelle Pineau, Satinder Singh, and Aaron C Courville. No-press diplomacy: Modeling multi-agent gameplay. In NeurIPS, 2019

    2019

  18. [26]

    Multi-agent reinforcement learning for cooperative air transportation services in city-wide autonomous urban air mobility

    Chanyoung Park, Gyu Seon Kim, Soohyun Park, Soyi Jung, and Joongheon Kim. Multi-agent reinforcement learning for cooperative air transportation services in city-wide autonomous urban air mobility. IEEE Transactions on Intelligent Vehicles, 2023

    2023

  19. [27]

    Civrealm: A learning and reasoning odyssey in civilization for decision-making agents

    Siyuan Qi, Shuo Chen, Yexin Li, Xiangyu Kong, Junqi Wang, Bangcheng Yang, Pring Wong, Yifan Zhong, Xiaoyuan Zhang, Zhaowei Zhang, et al. Civrealm: A learning and reasoning odyssey in civilization for decision-making agents. arXiv preprint arXiv:2401.10568, 2024

    2024 arXiv

  20. [28]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 2020

    2020

  21. [29]

    Swarmbrain: Embodied agent for real-time strategy game starcraft ii via large language models

    Xiao Shao, Weifu Jiang, Fei Zuo, and Mengqing Liu. Swarmbrain: Embodied agent for real-time strategy game starcraft ii via large language models. arXiv preprint arXiv:2401.11749, 2024

    2024

  22. [30]

    Alfworld: Aligning text and embodied environments for interactive learning

    Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Cote, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. Alfworld: Aligning text and embodied environments for interactive learning. In ICLR, 2021

    2021

  23. [31]

    Tenenbaum, Antonio Torralba, Shuang Li, and Igor Mordatch

    Vighnesh Subramaniam, Yilun Du, Joshua B. Tenenbaum, Antonio Torralba, Shuang Li, and Igor Mordatch. Multiagent finetuning: Self improvement with diverse reasoning chains. In ICLR, 2025

    2025

  24. [32]

    Mujoco: A physics engine for model-based control

    Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In IROS, 2012

    2012

  25. [33]

    Behavioral cloning from observation

    Faraz Torabi, Garrett Warnell, and Peter Stone. Behavioral cloning from observation. In IJCAI, 2018

    2018

  26. [34]

    LLaMA : Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Edward Lockhart, et al. LLaMA : Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  27. [35]

    Two tales of persona in llms: A survey of role-playing and personalization

    Yu-Min Tseng, Yu-Chao Huang, Teng-Yun Hsiao, Wei-Lin Chen, Chao-Wei Huang, Yu Meng, and Yun-Nung Chen. Two tales of persona in llms: A survey of role-playing and personalization. arXiv preprint arXiv:2406.01171, 2024

    2024 arXiv

  28. [36]

    Ilbit: Imitation learning for robot using position and torque information based on bilateral control with transformer

    Yusuke Urakami, Kazuya Yoshida, and Takashi Tsuji. Ilbit: Imitation learning for robot using position and torque information based on bilateral control with transformer. arXiv preprint arXiv:2401.16653, 2024

    2024 arXiv

  29. [37]

    Grandmaster level in starcraft ii using multi-agent reinforcement learning

    Oriol Vinyals, Igor Babuschkin, Junyoung Chung, Micha \"e l Mathieu, Max Jaderberg, Wojciech Marian Czarnecki, Andrew Dudzik, Brandon Houghton, Tobias Pohlen, Valentin Dalibard, et al. Grandmaster level in starcraft ii using multi-agent reinforcement learning. Nature, 2019

    2019

  30. [38]

    Voyager: An open-ended embodied agent with large language models

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. Transactions on Machine Learning Research, 2024

    2024

  31. [39]

    Scienceworld: Is your agent smarter than a 5th grader? In EMNLP, 2022

    Ruoyao Wang, Peter Jansen, Marc-Alexandre C \^o t \'e , and Prithviraj Ammanabrolu. Scienceworld: Is your agent smarter than a 5th grader? In EMNLP, 2022

    2022

  32. [40]

    Le, Ed H

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V. Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. In ICLR, 2023

    2023

  33. [41]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, 2022

    2022

  34. [42]

    Robocopilot: Human-in-the-loop interactive imitation learning for robot manipulation

    Philipp Wu, Yide Shentu, Qiayuan Liao, Ding Jin, Menglong Guo, Koushil Sreenath, Xingyu Lin, and Pieter Abbeel. Robocopilot: Human-in-the-loop interactive imitation learning for robot manipulation. arXiv preprint arXiv:2503.07771, 2025

    2025 arXiv

  35. [43]

    Llms are not good strategists, yet memory-enhanced agency boosts reasoning

    Yi Wu and Zhimin Hu. Llms are not good strategists, yet memory-enhanced agency boosts reasoning. In ICLR Workshop on Reasoning and Planning for LLMs, 2025

    2025

  36. [44]

    Qwen2 technical report

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze...

    2024 arXiv

  37. [45]

    Qwen3 technical report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jia...

    2025 arXiv

  38. [46]

    Multi-agent reinforcement learning: A selective overview of theories and algorithms

    Kaiqing Zhang, Zhuoran Yang, and Tamer Başar. Multi-agent reinforcement learning: A selective overview of theories and algorithms. arXiv preprint arXiv:1911.10635, 2019

    1911 arXiv

  39. [47]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  40. [48]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  41. [49]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  42. [50]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.