pith. sign in

arxiv: 2606.29126 · v1 · pith:NLFPNR72new · submitted 2026-06-28 · 💻 cs.AI

HiComm: Hierarchical Communication for Multi-agent Reinforcement Learning

Runze Zhao , Dongruo Zhou , Sumit Kumar Jha , Nathaniel D. Bastian , Ankit Shah This is my paper

Pith reviewed 2026-06-30 07:56 UTC · model grok-4.3

classification 💻 cs.AI
keywords multi-agent reinforcement learninghierarchical communicationcooperative MARLmessage passinginductive biasobservation structurediscrete selection
0
0 comments X

The pith

HiComm turns multi-agent communication into receiver-driven retrieval of specific feature slices from a sender's observation hierarchy.

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

The paper proposes that most communication protocols in cooperative multi-agent reinforcement learning send unstructured dense vectors and thereby miss an available source of structure. It introduces HiComm as a plug-in module in which the receiver issues a query that triggers a three-stage decoding process over the sender's hierarchical observations. The process first picks a group, then a sender, then an entity, and returns the matching feature slice rather than a learned vector. Experiments on tasks with varying observation structures show that this structured retrieval matches or exceeds the performance of standard learned-communication baselines while cutting the volume of data sent per receiver per episode by as much as 23 times.

Core claim

HiComm grounds messages in the sender's hierarchical observation by resolving a receiver query through sequential discrete selection of group, sender, and entity, implemented with Straight-Through Gumbel-Softmax and a shared lightweight projection so that the message becomes a retrieved feature slice instead of a transmitted dense vector.

What carries the argument

The receiver-driven three-stage decoding process that selects a group, then a sender, then an entity inside the sender's observation hierarchy and returns the corresponding feature slice.

If this is right

  • Communication volume per receiver drops sharply because only targeted feature slices are returned instead of full vectors.
  • The module attaches to existing MARL pipelines with minimal added parameters through the shared projection design.
  • Performance parity or gains hold across tasks that differ in observation structure and coordination demands.
  • Discrete selection via Straight-Through Gumbel-Softmax keeps the whole pipeline end-to-end differentiable.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same receiver-driven selection logic could be applied to other structured data sources such as spatial maps or temporal sequences without requiring an explicit group-entity hierarchy.
  • If the hierarchy must be discovered rather than given, an auxiliary loss that encourages consistent group and entity partitions might be needed to preserve the volume reduction.
  • In physical multi-robot deployments the reduction in transmitted bytes could translate directly into lower radio energy use or higher update rates under bandwidth constraints.

Load-bearing premise

Observations in cooperative environments naturally follow a hierarchy of groups and entities that supplies usable inductive bias for communication.

What would settle it

A cooperative task in which agents must exchange information that cannot be aligned with any fixed observation hierarchy and where HiComm then shows a clear performance deficit relative to flat-vector baselines.

read the original abstract

Cooperative multi-agent reinforcement learning (MARL) often relies on communication to mitigate partial observability, yet most existing protocols treat messages as flat dense vectors detached from the structure of the observations they summarize. This design overlooks an important source of inductive bias in many cooperative environments, where observations naturally follow a hierarchy such as groups and entities. We propose \textsc{HiComm}, a plug-in communication module that grounds messages in the sender's hierarchical observation. \textsc{HiComm} is receiver-driven: the receiver issues a query, and the hierarchy is resolved through a three-stage decoding process that first selects a group, then a sender, and then an entity within that group, returning the corresponding feature slice as the message. This converts communication from unstructured vector transmission into structured information retrieval over the sender's observation hierarchy. We instantiate this mechanism with Straight-Through Gumbel-Softmax for differentiable discrete selection and a lightweight shared projection design that attaches to standard MARL pipelines. Experiments across cooperative MARL tasks with different observation structures and coordination demands show that \textsc{HiComm} matches or outperforms representative learned communication baselines while reducing communication volume by up to $23\times$ per receiver per episode.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes HiComm, a plug-in communication module for cooperative multi-agent reinforcement learning. It grounds messages in the sender's hierarchical observation structure via a receiver-driven three-stage decoding process (group, sender, entity) implemented with Straight-Through Gumbel-Softmax, converting unstructured vector transmission into structured retrieval. Experiments on cooperative MARL tasks are claimed to show that HiComm matches or outperforms learned communication baselines while reducing communication volume by up to 23× per receiver per episode.

Significance. If the empirical results hold under rigorous verification, the work supplies a lightweight, structure-exploiting alternative to flat dense-vector communication protocols. The explicit use of observation hierarchy as inductive bias and the receiver-driven retrieval design could improve scalability in environments with natural group-entity structure, while the plug-in nature allows attachment to existing MARL pipelines without altering core algorithms.

major comments (2)
  1. [Abstract] Abstract: the central performance claim (matching or outperforming baselines with up to 23× volume reduction) is stated without reference to specific tasks, observation hierarchies, baseline implementations, number of runs, error bars, or statistical tests; these details are load-bearing for assessing whether the inductive-bias assumption actually drives the reported gains.
  2. [Abstract] The three-stage selection process is described at a high level but no equations or pseudocode are supplied for the Gumbel-Softmax parameterization, the shared projection, or the exact message construction; without these, it is impossible to verify that the discrete retrieval preserves the necessary information or that the volume reduction is achieved by construction rather than by environment-specific tuning.
minor comments (2)
  1. [Abstract] The abstract refers to 'representative learned communication baselines' without naming them or citing the corresponding papers.
  2. [Abstract] Notation for the hierarchy (groups, entities, feature slices) is introduced informally; a short diagram or formal definition would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments on the abstract below, proposing targeted revisions to improve specificity and clarity while preserving the abstract's conventional high-level nature.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central performance claim (matching or outperforming baselines with up to 23× volume reduction) is stated without reference to specific tasks, observation hierarchies, baseline implementations, number of runs, error bars, or statistical tests; these details are load-bearing for assessing whether the inductive-bias assumption actually drives the reported gains.

    Authors: We agree that greater specificity in the abstract would strengthen the central claim. In the revised manuscript we will expand the abstract to name the primary evaluation domains (e.g., StarCraft II micromanagement scenarios and other cooperative tasks with explicit group-entity hierarchies), list the representative baselines (CommNet, TarMAC, and IC3Net), state that results are averaged over 5–10 independent seeds with error bars, and note that statistical significance was assessed via paired t-tests. The full experimental protocol, including hierarchy definitions and volume calculations, remains in Section 4. revision: yes

  2. Referee: [Abstract] The three-stage selection process is described at a high level but no equations or pseudocode are supplied for the Gumbel-Softmax parameterization, the shared projection, or the exact message construction; without these, it is impossible to verify that the discrete retrieval preserves the necessary information or that the volume reduction is achieved by construction rather than by environment-specific tuning.

    Authors: Abstracts conventionally omit equations. The complete three-stage decoding (group, sender, entity) is formalized in Section 3 with the Straight-Through Gumbel-Softmax parameterization (Eqs. 4–6), the shared linear projection, and the exact message-construction rule that returns a single feature slice. Algorithm 1 provides the corresponding pseudocode. The volume reduction follows directly from the discrete selection of one entity vector rather than a dense message; this is independent of any particular environment. We will add a parenthetical reference in the abstract directing readers to Section 3 for the technical details. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper introduces HiComm as an independent plug-in module for standard MARL pipelines, grounding communication in an explicitly assumed hierarchical observation structure via a three-stage receiver-driven selection process instantiated with Straight-Through Gumbel-Softmax. No load-bearing equations, fitted parameters renamed as predictions, or self-citation chains are present in the provided description; the central empirical claims rest on experiments across cooperative tasks rather than reducing to definitional equivalences or prior author results by construction. The inductive bias assumption is stated outright rather than smuggled in, making the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that observations possess exploitable hierarchical structure and that differentiable discrete selection via Gumbel-Softmax can be attached to existing MARL pipelines without further justification.

axioms (1)
  • domain assumption Observations in many cooperative environments naturally follow a hierarchy such as groups and entities.
    Invoked in the abstract as the overlooked source of inductive bias that flat protocols ignore.

pith-pipeline@v0.9.1-grok · 5750 in / 1275 out tokens · 32555 ms · 2026-06-30T07:56:11.689442+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 4 canonical work pages · 2 internal anchors

  1. [1]

    Feudal multi-agent hierarchies for cooperative reinforcement learning

    S Ahilan and P Dayan. Feudal multi-agent hierarchies for cooperative reinforcement learning. InWorkshop on Structure & Priors in Reinforcement Learning (SPiRL 2019) at ICLR 2019, pages 1–11, 2019

    2019

  2. [2]

    Tarmac: Targeted multi-agent communication

    Abhishek Das, Théophile Gervet, Joshua Romoff, Dhruv Batra, Devi Parikh, Mike Rabbat, and Joelle Pineau. Tarmac: Targeted multi-agent communication. InInternational Conference on machine learning, pages 1538–1546. PMLR, 2019

    2019

  3. [3]

    Is independent learning all you need in the starcraft multi-agent challenge?arXiv preprint arXiv:2011.09533, 2020

    Christian Schroeder De Witt, Tarun Gupta, Denys Makoviichuk, Viktor Makoviychuk, Philip HS Torr, Mingfei Sun, and Shimon Whiteson. Is independent learning all you need in the starcraft multi-agent challenge?arXiv preprint arXiv:2011.09533, 2020

    work page arXiv 2011

  4. [4]

    Multi-agent coordination via multi-level communication.Advances in Neural Information Processing Systems, 37:118513–118539, 2024

    Ziluo Ding, Zeyuan Liu, Zhirui Fang, Kefan Su, Liwen Zhu, and Zongqing Lu. Multi-agent coordination via multi-level communication.Advances in Neural Information Processing Systems, 37:118513–118539, 2024

    2024

  5. [5]

    Smacv2: An improved benchmark for cooperative multi-agent reinforcement learning.Advances in Neural Information Processing Systems, 36:37567–37593, 2023

    Benjamin Ellis, Jonathan Cook, Skander Moalla, Mikayel Samvelyan, Mingfei Sun, Anuj Mahajan, Jakob Foerster, and Shimon Whiteson. Smacv2: An improved benchmark for cooperative multi-agent reinforcement learning.Advances in Neural Information Processing Systems, 36:37567–37593, 2023

    2023

  6. [6]

    Learning to communicate with deep multi-agent reinforcement learning

    Jakob N Foerster, Yannis M Assael, Nando de Freitas, and Shimon Whiteson. Learning to communicate with deep multi-agent reinforcement learning. InProceedings of the 30th International Conference on Neural Information Processing Systems, pages 2145–2153, 2016

    2016

  7. [7]

    Cong Guan, Feng Chen, Lei Yuan, Zongzhang Zhang, and Yang Yu. Efficient communication via self-supervised information aggregation for online and offline multiagent reinforcement learning.IEEE Transactions on Neural Networks and Learning Systems, 36(5):9044–9056, 2024

    2024

  8. [8]

    Learning multi-agent communication from graph modeling perspective

    Shengchao Hu, Li Shen, Ya Zhang, and Dacheng Tao. Learning multi-agent communication from graph modeling perspective. InThe Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=Qox9rO0kN0

    2024

  9. [9]

    Categorical reparameterization with gumbel-softmax

    Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In International Conference on Learning Representations, 2017. URL https://openreview.net/forum? id=rkE3y85ee

    2017

  10. [10]

    Learning attentional communication for multi-agent cooperation

    Jiechuan Jiang and Zongqing Lu. Learning attentional communication for multi-agent cooperation. Advances in neural information processing systems, 31, 2018

    2018

  11. [11]

    Graph convolutional reinforcement learning

    Jiechuan Jiang, Chen Dun, Tiejun Huang, and Zongqing Lu. Graph convolutional reinforcement learning. InInternational Conference on Learning Representations, 2020. URL https://openreview.net/ forum?id=HkxdQkSYDB

    2020

  12. [12]

    Cage challenge 4: A scalable multi-agent reinforcement learning gym for autonomous cyber defence.AI Magazine, 46(3):e70021, 2025

    Mitchell Kiely, Metin Ahiskali, Etienne Borde, Benjamin Bowman, David Bowman, Dirk Van Bruggen, KC Cowan, Prithviraj Dasgupta, Erich Devendorf, Ben Edwards, et al. Cage challenge 4: A scalable multi-agent reinforcement learning gym for autonomous cyber defence.AI Magazine, 46(3):e70021, 2025

    2025

  13. [13]

    Exploring the efficacy of multi-agent reinforcement learning for autonomous cyber defence: A cage challenge 4 perspective

    Mitchell Kiely, Metin Ahiskali, Etienne Borde, Benjamin Bowman, David Bowman, Dirk Van Bruggen, KC Cowan, Prithviraj Dasgupta, Erich Devendorf, Ben Edwards, et al. Exploring the efficacy of multi-agent reinforcement learning for autonomous cyber defence: A cage challenge 4 perspective. InProceedings of the AAAI Conference on Artificial Intelligence, volum...

    2025

  14. [14]

    Learning to Schedule Communication in Multi-agent Reinforcement Learning

    Daewoo Kim, Sangwoo Moon, David Hostallero, Wan Ju Kang, Taeyoung Lee, Kyunghwan Son, and Yung Yi. Learning to schedule communication in multi-agent reinforcement learning.arXiv preprint arXiv:1902.01554, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1902

  15. [15]

    Google research football: A novel reinforcement learning environment

    Karol Kurach, Anton Raichuk, Piotr Sta ´nczyk, Michał Zaj ˛ ac, Olivier Bachem, Lasse Espeholt, Carlos Riquelme, Damien Vincent, Marcin Michalski, Olivier Bousquet, et al. Google research football: A novel reinforcement learning environment. InProceedings of the AAAI conference on artificial intelligence, volume 34, pages 4501–4510, 2020

    2020

  16. [16]

    Deep implicit coordina- tion graphs for multi-agent reinforcement learning

    Sheng Li, Jayesh K Gupta, Peter Morales, Ross Allen, and Mykel J Kochenderfer. Deep implicit coordina- tion graphs for multi-agent reinforcement learning. InProceedings of the 20th International Conference on Autonomous Agents and MultiAgent Systems, pages 764–772, 2021

    2021

  17. [17]

    Context-aware communication for multi-agent reinforcement learning

    Xinran Li and Jun Zhang. Context-aware communication for multi-agent reinforcement learning. In Proceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems, pages 1156–1164, 2024

    2024

  18. [18]

    When2com: Multi-agent perception via communication graph grouping

    Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. InProceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4106–4115, 2020

    2020

  19. [19]

    Who2com: Collaborative perception via learnable handshake communication

    Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com: Collaborative perception via learnable handshake communication. In2020 IEEE International Conference on Robotics and Automation (ICRA), pages 6876–6883. IEEE, 2020

    2020

  20. [20]

    Deep hierarchical com- munication graph in multi-agent reinforcement learning

    Zeyang Liu, Lipeng Wan, Xue Sui, Zhuoran Chen, Kewu Sun, and Xuguang Lan. Deep hierarchical com- munication graph in multi-agent reinforcement learning. InProceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 208–216, 2023

    2023

  21. [21]

    Hierarchical message-passing policies for multi-agent reinforcement learning.arXiv preprint arXiv:2507.23604, 2025

    Tommaso Marzi, Cesare Alippi, and Andrea Cini. Hierarchical message-passing policies for multi-agent reinforcement learning.arXiv preprint arXiv:2507.23604, 2025

    work page arXiv 2025

  22. [22]

    Multi-agent graph-attention communication and teaming

    Yaru Niu, Rohan Paleja, and Matthew Gombolay. Multi-agent graph-attention communication and teaming. InProceedings of the 20th International Conference on Autonomous Agents and MultiAgent Systems, pages 964–973, 2021

    2021

  23. [23]

    Learning multi-agent communication through structured attentive reasoning.Advances in Neural Information Processing Systems, 33:10088–10098, 2020

    Murtaza Rangwala and Ryan Williams. Learning multi-agent communication through structured attentive reasoning.Advances in Neural Information Processing Systems, 33:10088–10098, 2020

    2020

  24. [24]

    Multi-agent actor-critic with hierarchical graph attention network

    Heechang Ryu, Hayong Shin, and Jinkyoo Park. Multi-agent actor-critic with hierarchical graph attention network. InProceedings of the AAAI conference on artificial intelligence, volume 34, pages 7236–7243, 2020

    2020

  25. [25]

    The starcraft multi-agent challenge

    Mikayel Samvelyan, Tabish Rashid, Christian Schroeder de Witt, Gregory Farquhar, Nantas Nardelli, Tim GJ Rudner, Chia-Man Hung, Philip HS Torr, Jakob Foerster, and Shimon Whiteson. The starcraft multi-agent challenge. InProceedings of the 18th International Conference on Autonomous Agents and MultiAgent Systems, pages 2186–2188, 2019

    2019

  26. [26]

    Learning structured communication for multi-agent reinforcement learning.Autonomous Agents and Multi-Agent Systems, 36(2):50, 2022

    Junjie Sheng, Xiangfeng Wang, Bo Jin, Junchi Yan, Wenhao Li, Tsung-Hui Chang, Jun Wang, and Hongyuan Zha. Learning structured communication for multi-agent reinforcement learning.Autonomous Agents and Multi-Agent Systems, 36(2):50, 2022

    2022

  27. [27]

    Hierarchical multi-agent reinforcement learning for cyber network defense

    Aditya Vikram Singh, Ethan Rathbun, Emma Graham, Lisa Oakley, Simona Boboila, Peter Chin, and Alina Oprea. Hierarchical multi-agent reinforcement learning for cyber network defense. InProceedings of the 24th International Conference on Autonomous Agents and Multiagent Systems, pages 2747–2749, 2025

    2025

  28. [28]

    Learning when to Communicate at Scale in Multiagent Cooperative and Competitive Tasks

    Amanpreet Singh, Tushar Jain, and Sainbayar Sukhbaatar. Learning when to communicate at scale in multiagent cooperative and competitive tasks.arXiv preprint arXiv:1812.09755, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  29. [29]

    Code: Communication delay-tolerant multi-agent collaboration via dual alignment of intent and timeliness

    Shoucheng Song, Youfang Lin, Sheng Han, Chang Yao, Hao Wu, Shuo Wang, and Kai Lv. Code: Communication delay-tolerant multi-agent collaboration via dual alignment of intent and timeliness. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 23304–23312, 2025

    2025

  30. [30]

    Boosting studies of multi-agent reinforcement learning on google research football environment: the past, present, and future

    Y Song, H Jiang, H Zhang, Z Tian, W Zhang, and J Wang. Boosting studies of multi-agent reinforcement learning on google research football environment: the past, present, and future. InProceedings of the International Joint Conference on Autonomous Agents and Multiagent Systems, AAMAS, volume 2024, pages 1772–1781. Association for Computing Machinery (ACM)...

    2024

  31. [31]

    Learning multiagent communication with backprop- agation

    Sainbayar Sukhbaatar, Arthur Szlam, and Rob Fergus. Learning multiagent communication with backprop- agation. InProceedings of the 30th International Conference on Neural Information Processing Systems, pages 2252–2260, 2016

    2016

  32. [32]

    T2mac: Targeted and trusted multi-agent communication through selective engagement and evidence- driven integration

    Chuxiong Sun, Zehua Zang, Jiabao Li, Jiangmeng Li, Xiao Xu, Rui Wang, and Changwen Zheng. T2mac: Targeted and trusted multi-agent communication through selective engagement and evidence- driven integration. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 15154–15163, 2024

    2024

  33. [33]

    Learning nearly decomposable value functions via communication minimization

    Tonghan Wang, Jianhao Wang, Chongyi Zheng, and Chongjie Zhang. Learning nearly decomposable value functions via communication minimization. InInternational Conference on Learning Representations,

  34. [34]

    URLhttps://openreview.net/forum?id=HJx-3grYDB

  35. [35]

    Context-aware sparse deep coordination graphs

    Tonghan Wang, Liang Zeng, Weijun Dong, Qianlan Yang, Yang Yu, and Chongjie Zhang. Context-aware sparse deep coordination graphs. InInternational Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=wQfgfb8VKTn

    2022

  36. [36]

    Subgoal-based hierarchical reinforcement learning for multiagent collaboration.IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2026

    Cheng Xu, Yuchen Shi, Changtian Zhang, Ran Wang, Shihong Duan, Yadong Wan, and Xiaotong Zhang. Subgoal-based hierarchical reinforcement learning for multiagent collaboration.IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2026

    2026

  37. [37]

    The surpris- ing effectiveness of ppo in cooperative multi-agent games.Advances in neural information processing systems, 35:24611–24624, 2022

    Chao Yu, Akash Velu, Eugene Vinitsky, Jiaxuan Gao, Yu Wang, Alexandre Bayen, and Yi Wu. The surpris- ing effectiveness of ppo in cooperative multi-agent games.Advances in neural information processing systems, 35:24611–24624, 2022

    2022

  38. [38]

    Multi-agent incentive communication via decentralized teammate modeling

    Lei Yuan, Jianhao Wang, Fuxiang Zhang, Chenghe Wang, Zongzhang Zhang, Yang Yu, and Chongjie Zhang. Multi-agent incentive communication via decentralized teammate modeling. InProceedings of the AAAI conference on artificial intelligence, volume 36, pages 9466–9474, 2022

    2022

  39. [39]

    G-designer: Architecting multi-agent communication topologies via graph neural networks

    Guibin Zhang, Yanwei Yue, Xiangguo Sun, Guancheng Wan, Miao Yu, Junfeng Fang, Kun Wang, Tianlong Chen, and Dawei Cheng. G-designer: Architecting multi-agent communication topologies via graph neural networks. InForty-second International Conference on Machine Learning, 2025. URL https: //openreview.net/forum?id=LpE54NUnmO

    2025

  40. [40]

    Efficient communication in multi-agent reinforcement learning via variance based control

    Sai Qian Zhang, Qi Zhang, and Jieyu Lin. Efficient communication in multi-agent reinforcement learning via variance based control. InAdvances in Neural Information Processing Systems, pages 3235–3244, 2019

    2019

  41. [41]

    Succinct and robust multi-agent communication with temporal message control.Advances in Neural Information Processing Systems, 33:17271–17282, 2020

    Sai Qian Zhang, Qi Zhang, and Jieyu Lin. Succinct and robust multi-agent communication with temporal message control.Advances in Neural Information Processing Systems, 33:17271–17282, 2020

    2020

  42. [42]

    higher is better

    Zhuohui Zhang, Bin He, Bin Cheng, and Gang Li. Bridging training and execution via dynamic directed graph-based communication in cooperative multi-agent systems. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 23395–23403, 2025. 12 A Additional Details for Experiments All experiments were conducted on a single node equipp...

    2025

  43. [43]

    owned by us

    On the 5v5 scenarios this resolves (Ne, Na−1) to (5,4) , so the concrete obs_segs we hand to HICOMMis (1,4),(5,|F |),(4,|F |),(1, o a),(1,11),(1,5) , with |F | ∈ {8,9} and oa ∈ {4,5} chosen by race as above. The action space is a discrete head of size Ne + 6 (Ne unit targeted attacks plus no-op, stop, and the four cardinal moves), with a per step validity...