Pith. sign in

REVIEW 3 major objections 5 minor 30 references

This paper claims that decentralized agents coordinate better under temporary occlusion when their messages are generated from a world model's latent dynamics—which encode observation and action history—rather than from current observations

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 11:37 UTC pith:VDRAEGVV

load-bearing objection New integration of CPC messaging with an RSSM world model, but the experiments conflate backbone and message-source changes, so the 'latent dynamics' claim is underdetermined. the 3 major comments →

arxiv 2607.19809 v1 pith:VDRAEGVV submitted 2026-07-22 cs.MA cs.LG

Dreamer-CPC: Message Learning with World Models for Decentralized Multi-agent Reinforcement Learning

Taisuke Takayama , Naoto Yoshida , Tadahiro Taniguchi This is my paper
classification cs.MA cs.LG
keywords multi-agent reinforcement learningworld modelscollective predictive codingemergent communicationdecentralized training and executionpartial observabilitylatent dynamicstemporally missing observations
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper asks whether messages between agents can carry more than what an agent currently sees, and answers yes if the messages come from a learned model of the world's dynamics. Dreamer-CPC gives each agent its own world model plus a message module, and samples messages from the model's latent state—a state that already summarizes past observations and actions. World-model learning and message learning are optimized together by a single variational objective, so no centralized training or shared reward is needed. In CatchApple, a task where the object's column is hidden at the moment the catcher must commit, Dreamer-CPC reaches an episode return about four to five times that of IPPO-CPC, a method that builds messages from current observations. The paper concludes that communication grounded in latent dynamics can support decentralized coordination when current observations alone are insufficient.

Core claim

Dreamer-CPC integrates Collective Predictive Coding-style message learning into the recurrent state-space model (RSSM) of a world model. Each agent's message module keeps its own recurrent state and samples a stochastic message from a posterior conditioned on the RSSM latent state; messages from other agents enter the RSSM transition and observation prediction, but are treated with stop-gradient, so no gradient flows between agents. The unified objective combines prediction, dynamics, representation, and message-prior KL terms, so world-model learning and message learning happen jointly. In the Observer and CatchApple tasks, the method surpasses IPPO-CPC and no-communication baselines; in Ca

What carries the argument

The recurrent state-space model (RSSM) world model—a deterministic recurrent state h_t plus a stochastic latent z_t updated from past observations, actions, and received messages—together with a recurrent message module that maintains its own state ξ_t. The message module predicts messages from ξ_t alone and infers them from ξ_t plus the RSSM state; predicted messages condition the RSSM transition, while inferred messages condition observation prediction. This arrangement is what lets a message represent temporal context rather than only the current observation.

Load-bearing premise

The paper's central explanation—that the gain comes from latent-history-grounded messages—is only indirectly supported because the winning method changes the RL algorithm (model-based Dreamer versus model-free IPPO) at the same time as it changes the message source, so the CatchApple gap cannot be cleanly attributed to the message source alone.

What would settle it

Run Dreamer-CPC with messages sampled from the current RSSM latent state but without the recurrent message module, holding all other Dreamer machinery fixed. If CatchApple return stays near 28, the history-carrying message channel is not the cause; if it drops toward IPPO-CPC's near-5 level, the paper's explanation is supported.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Decentralized agents can act on information that is not present in any current observation, as long as their world models can track the hidden state.
  • Message learning and world-model learning can share a single objective, removing the need for a separate communication reward or a centralized value function.
  • Because messages from other agents are stop-gradient inputs, the scheme introduces no inter-agent gradient paths, preserving decentralized training.
  • In CatchApple, the method turns a near-failure baseline (episode return around 5) into a functional coordinator (around 28), roughly a 4 to 5 times improvement.
  • The same latent-state message channel also helps in non-cooperative settings where agents do not share rewards.

Where Pith is reading between the lines

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

  • Beyond the paper: if messages carry latent history, the same channel should let agents anticipate not only occluded states but also other agents' future actions, since the RSSM latent is predictive; a direct test would be a task with longer, variable occlusion gaps.
  • Beyond the paper: the message-prior KL weights are very small (1e-3 and 1e-4), which raises the possibility that the message channel acts as a free-form bottleneck; varying these weights could reveal how much of the gain comes from message content as opposed to the world model's own predictions.
  • Beyond the paper: the two-agent, custom-environment evidence suggests that scaling to more agents would require messages to become compressed summaries of latent state rather than copies; a natural next step is measuring how performance degrades as the number of agents grows.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces Dreamer-CPC, a decentralized multi-agent RL method that adds a stochastic message module to each agent's DreamerV3-style RSSM world model. Each agent maintains its own world model and message module; at every step, a message is inferred from the agent's latent state (deterministic recurrent state h and stochastic latent z, Eq. 10), exchanged with other agents, and used in the next transition (Eq. 11) and observation prediction (Eq. 12). Messages from other agents are detached to keep learning decentralized. World-model and message learning are trained jointly by minimizing a variational lower bound (Eqs. 15-19) with KL-balanced free-bits terms (Eqs. 20-24); actor and critic are trained as in DreamerV3, with imagination rollouts that include synchronous message exchange. The method is evaluated in two custom two-agent tasks: Observer, where one stationary agent observes the reward location and a second must dig at that location, and CatchApple, where the object's trajectory is temporarily occluded and the catcher must anticipate the landing column from messages. Dreamer-CPC outperforms IPPO-CPC and no-communication baselines in both tasks (e.g., CatchApple IQM 28.53 vs 5.52 for IPPO-CPC), suggesting that latent-dynamics-grounded messages help when current observations are insufficient.

Significance. If the mechanism claim is upheld, the paper makes a useful contribution: an extension of CPC-based messaging from current observations to latent histories, with a principled derivation and plausible empirical gains. Concrete strengths: the ELBO derivation (Eqs. 15-18) is standard and correctly carried out; the DTDE setup is clearly described; and the evaluation uses 10 independent runs with IQM and bootstrap CIs rather than a single run. The two main weaknesses are (i) the experiments do not isolate the claimed mechanism and (ii) an implementation-critical detail (the discrete-message gradient estimator) is unstated. Both are addressable in revision.

major comments (3)
  1. [Section 3.3, Fig. 6] The headline comparison confounds two variables. Dreamer-CPC and IPPO-CPC differ simultaneously in the RL backbone (DreamerV3 model-based vs IPPO model-free) and the message source (RSSM latent (h^k_t,z^k_t), Eq. 10, vs current observation). The low no-comm DreamerV3 baselines (0.48 Observer, 5.93 CatchApple) show a channel is necessary, but not that latent-dynamics grounding is the operative factor. Observer has no occlusion - the full state is reconstructible from current observations every step - yet Dreamer-CPC (39.85) far exceeds IPPO-CPC (7.31); this gap is more plausibly ascribed, at least partly, to backbone differences. A within-backbone ablation (DreamerV3-CPC with current-observation messages, and/or IPPO with recurrent-state messages) is required before the title claim is supported.
  2. [Section 2.3, Eqs. (10)-(12), (23)-(24); Appendix A] Messages are one-hot (Section 3), but no gradient estimator for the discrete samples is specified. The message posterior is trained via the ELBO (Eq. 17), i.e., through the decoder input (Eq. 12) and the message KL terms. For one-hot samples, no gradient reaches q_psi through the sample without a straight-through/Gumbel-Softmax/REINFORCE estimator. Moreover, the KL terms are nearly inactive: free-bits threshold 0.5 and weights beta_m-dyn=1e-3, beta_m-rep=1e-4. Without stating the estimator, the 'unified objective' claim (contribution 2) is unverifiable and the experiments are not reproducible. Please specify the gradient path and report sensitivity to these weights.
  3. [Section 4 (limitations)] The paper explicitly states that what information the learned messages carry was not analyzed. This is load-bearing because the central claim is that latent-dynamics grounding lets messages carry temporal history, and because Fig. 6's comparison is confounded (see major comment 1). I recommend closing this limitation with at least a probe (e.g., decoding the object trajectory from messages) or a control with random messages, to show the gains come from information-bearing communication rather than from the model-based backbone alone. The limitation is acknowledged by the authors; I weigh it as requiring action before publication, not merely a scope note.
minor comments (5)
  1. [Eqs. (15)-(18)] The derivation does not spell out that h_t in Eq. (11) uses the prior-predicted \hat m_t while the variational q in Eq. (10) conditions on h_t; given the deterministic recurrence, a brief clarification of the amortized variational family and the role of \hat m_t would prevent apparent circularity.
  2. [Section 2.3] No InfoNCE/contrastive loss appears in the objective; the connection between the ELBO (Eq. 17) and 'Collective Predictive Coding' (Refs. [22,23]) should be spelled out, since 'CPC' typically denotes a contrastive objective.
  3. [Fig. 6] The curves are hard to read; a table of final IQMs with 95% CIs for each method and message configuration would make the numbers in Section 3.3 easier to verify. Also, the legend is small.
  4. [Section 2.2 heading; Section 3] The heading 'W orld Model' has a typo. The obs-shared baseline is not precisely defined: what exactly is shared (raw observations or a concatenated vector) and does the policy remain independent?
  5. [General] No code or artifact link is provided; releasing code would improve reproducibility, especially given the unstated discrete-message gradient details.

Circularity Check

0 steps flagged

No significant circularity: the central result is an external reward comparison; only minor non-load-bearing self-citations are present.

full rationale

Dreamer-CPC's message construction does make the statement that messages 'reflect history' true by design: Eq. (10) defines m_t^k ~ q(h_t^k, z_t^k, ξ_t^k), and h_t is recurrent, so the message is a function of accumulated latent state. However, this is an architectural design property, not a fitted prediction or a claimed first-principles result. The paper's headline claim is empirical: Dreamer-CPC outperforms IPPO-CPC and no-communication baselines on episode return measured in the environments. That comparison is against external task rewards, not against the definition of the message. The ELBO in Eq. (17) is a standard Jensen lower bound derived in the text; the KL-balanced surrogate follows DreamerV3 by citation, but DreamerV3 is external and not the target of the paper's claim. The main weakness is a confound, not circularity: Dreamer-CPC and IPPO-CPC differ in both backbone and message source, and the paper's own limitation section admits that the learned message content was not analyzed. This undermines the attribution of the performance gain to latent-dynamics grounding, but it does not reduce the result to its inputs by construction. Self-citations to CPC and MARL-CPC are present and contextual, but no load-bearing uniqueness theorem is imported and the derivation is self-contained apart from standard external technique. Therefore the circularity score is low.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The central method rests on standard ELBO math plus several hand-set weighting coefficients and a design choice (messages from RSSM state with stop-gradient). No independent external evidence for the message representation is provided beyond task returns on two custom environments.

free parameters (4)
  • β_pred, β_dyn, β_rep (state KL weights) = 1, 1, 0.1
    Hand-set coefficients in the world-model loss (Eq. 19); values follow DreamerV3 defaults but are still chosen weights with no sensitivity analysis.
  • β_m-dyn (message prior KL weight) = 1×10^-3
    Hand-set weight for Eq. 23 controlling how predictable messages must be from message history; no sensitivity study.
  • β_m-rep (message representation KL weight) = 1×10^-4
    Hand-set weight for Eq. 24; very small, so messages may carry little predictable structure beyond reconstruction.
  • Free-bits thresholds for state and message KL = 1.0 (state), 0.5 (message)
    Thresholds in the max() operations of Eq. 21-24; hand-set, inherited from free-bits / DreamerV3 practice.
axioms (5)
  • domain assumption The environment is a partially observable Markov game and each agent optimizes its own discounted return (Eq. 1).
    Problem statement; excludes shared-reward cooperative settings and non-Markovian state structure.
  • domain assumption An RSSM/DreamerV3 world model with deterministic recurrent state h_t and stochastic latent z_t is sufficient to capture task-relevant history.
    Central modeling choice in Section 2.2; if false, messages have no useful history to extract.
  • standard math The variational lower bound via Jensen's inequality (Eq. 17) is a valid objective, and KL-balanced/free-bits surrogates preserve the optimum direction.
    Standard ELBO derivation; the surrogate modifications in Eq. 19-24 are heuristic.
  • ad hoc to paper Messages from other agents can be treated as fixed inputs via stop-gradient, preserving decentralized learning.
    Algorithmic design choice in Section 2.3 and Algorithm 1; not derived from first principles.
  • domain assumption Reward and continuation predictions do not condition on messages (Eq. 13).
    Modeling simplification; messages affect only observation prediction and state transition.
invented entities (1)
  • Recurrent message state ξ_t and stochastic message m_t no independent evidence
    purpose: Auxiliary latent variables that summarize and transmit history encoded in the RSSM state
    Introduced for the method; there is no external falsifiable prediction beyond task returns on two custom environments.

pith-pipeline@v1.3.0-alltime-deepseek · 10608 in / 13228 out tokens · 122898 ms · 2026-08-01T11:37:55.038055+00:00 · methodology

0 comments
read the original abstract

In multi-agent reinforcement learning (MARL), inter-agent communication is effective for improving performance under partial observability. Representation learning-based approaches enable decentralized agents to learn messages grounded in their own observations, but they rely only on current observations and cannot convey information accumulated over time. We propose Dreamer-CPC, a decentralized model-based MARL method that integrates message learning based on Collective Predictive Coding (CPC) into the world model of DreamerV3. Each agent independently maintains a world model and a message module, and infers and exchanges messages from the latent states of the world model that reflect the history of past observations and actions. We evaluated Dreamer-CPC in two environments: Observer, a non-cooperative information-sharing task, and CatchApple, a newly introduced task in which task-relevant observations are temporarily missing. In both environments, Dreamer-CPC outperformed IPPO-CPC, an existing CPC-based method that generates messages from current observations, as well as no-communication baselines. In particular, in CatchApple, Dreamer-CPC achieved 4 to 5 times the episode return of IPPO-CPC, demonstrating effective coordination where other methods fail due to missing observations. These results suggest that communication grounded in the latent dynamics of world models can support decentralized decision-making when current observations alone are insufficient.

Figures

Figures reproduced from arXiv: 2607.19809 by Naoto Yoshida, Tadahiro Taniguchi, Taisuke Takayama.

Figure 1
Figure 1. Figure 1: Overview of Dreamer-CPC. Each agent maintains an RSSM-based world model and a message module. Messages are inferred from the latent states of the world model and exchanged between agents. and ∆(S) denotes the set of probability distributions over S. At each time step t, each agent k selects an action a k t ∈ Ak according to its policy π k , using only its available local history. Given the joint action at … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of world-model learning in Dreamer-CPC. Each agent learns an RSSM state from its local observation history and action history. The message module pre￾dicts a message from its recurrent state and infers a message from the RSSM state. The learned model reconstructs observations using the inferred message and received messages, and predicts rewards and continuation from the RSSM state. Messages from … view at source ↗
Figure 3
Figure 3. Figure 3: Overview of actor and critic learning with message exchange during imagi￾nation. At each imagined step, agents synchronously apply the message generation and inference procedure, share messages, and update their RSSM states using the exchanged messages. Actor and critic losses are then computed on the resulting imagi￾nation rollout. Stop gradient labels indicate that received messages are detached during o… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the Ob￾server environment. Agent 1 observes the reward location, shown as the colored cell, but cannot move, whereas Agent 2 observes only its own grid po￾sition and must locate the re￾ward through communication. This environment is a non-cooperative MARL en￾vironment in which multiple agents receive asym￾metric observations, the complete state informa￾tion of the environment can be reconstruct… view at source ↗
Figure 5
Figure 5. Figure 5: Overview of CatchApple environment. (A) The environment contains an 8 × 8 falling object area and one ground row. Agent 1 observes the falling object trajectory and receives no reward, whereas Agent 2 controls the catcher on the ground row and observes its own column. The hatched region indicates an occluded region in which the object position is replaced by a hidden flag. (B) The object moves downward, po… view at source ↗
Figure 6
Figure 6. Figure 6: Learning curves in Observer and CatchApple. Curves show the IQM of episode return over 10 independent runs. Shaded regions show 95% confidence intervals esti￾mated by paired nonparametric bootstrap over runs. The left and right panels show Observer and CatchApple, respectively. tests whether such message grounding supports coordination under temporally occluded observations. 3.3 Results [PITH_FULL_IMAGE:f… view at source ↗

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

30 extracted references · 2 linked inside Pith

  1. [1]

    MIT Press (2024)

    Albrecht, S.V., Christianos, F., Sch¨ afer, L.: Multi-Agent Reinforcement Learning: Foundations and Modern Approaches. MIT Press (2024)

    2024

  2. [2]

    In: 2023 15th International Congress on Advanced Applied Informatics Winter (IIAI-AAI-Winter)

    Ebara, H., Nakamura, T., Taniguchi, A., Taniguchi, T.: Multi-agent reinforcement learning with emergent communication using discrete and indifferentiable message. In: 2023 15th International Congress on Advanced Applied Informatics Winter (IIAI-AAI-Winter). pp. 366–371 (2023)

    2023

  3. [3]

    In: 21st International Conference on Autonomous Agents and Multiagent Sys- tems, AAMAS 2022, Auckland, New Zealand, May 9-13, 2022

    Egorov, V., Shpilman, A.: Scalable multi-agent model-based reinforcement learn- ing. In: 21st International Conference on Autonomous Agents and Multiagent Sys- tems, AAMAS 2022, Auckland, New Zealand, May 9-13, 2022. pp. 381–390 (2022)

    2022

  4. [4]

    In: Advances in Neural Information Processing Systems

    Foerster, J.N., Assael, Y.M., de Freitas, N., Whiteson, S.: Learning to communicate with deep multi-agent reinforcement learning. In: Advances in Neural Information Processing Systems. vol. 29, pp. 2137–2145 (2016)

    2016

  5. [5]

    Ar- tificial Intelligence Review55(2), 895–943 (2022)

    Gronauer, S., Diepold, K.: Multi-agent deep reinforcement learning: A survey. Ar- tificial Intelligence Review55(2), 895–943 (2022)

    2022

  6. [6]

    In: Advances in Neural Information Processing Systems (2018)

    Ha, D., Schmidhuber, J.: Recurrent world models facilitate policy evolution. In: Advances in Neural Information Processing Systems (2018)

    2018

  7. [7]

    In: Proceedings of the 36th International Conference on Machine Learning

    Hafner, D., Lillicrap, T., Fischer, I., Villegas, R., Ha, D., Lee, H., Davidson, J.: Learning latent dynamics for planning from pixels. In: Proceedings of the 36th International Conference on Machine Learning. pp. 2555–2565 (2019)

    2019

  8. [8]

    In: International Conference on Learning Representations (2021)

    Hafner, D., Lillicrap, T.P., Norouzi, M., Ba, J.: Mastering Atari with discrete world models. In: International Conference on Learning Representations (2021)

    2021

  9. [9]

    Nature640(8059), 647–653 (2025)

    Hafner, D., Pasukonis, J., Ba, J., Lillicrap, T.: Mastering diverse control tasks through world models. Nature640(8059), 647–653 (2025)

    2025

  10. [10]

    In: Proceedings of the Nineteenth National Confer- ence on Artificial Intelligence

    Hansen, E.A., Bernstein, D.S., Zilberstein, S.: Dynamic programming for partially observable stochastic games. In: Proceedings of the Nineteenth National Confer- ence on Artificial Intelligence. pp. 709–715 (2004)

    2004

  11. [11]

    Acta Mathematica30(1), 175–193 (1906)

    Jensen, J.L.W.V.: Sur les fonctions convexes et les in´ egalit´ es entre les valeurs moyennes. Acta Mathematica30(1), 175–193 (1906)

    1906

  12. [12]

    Artificial Intelligence101(1–2), 99–134 (1998)

    Kaelbling, L.P., Littman, M.L., Cassandra, A.R.: Planning and acting in partially observable stochastic domains. Artificial Intelligence101(1–2), 99–134 (1998)

    1998

  13. [13]

    In: 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings (2014)

    Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings (2014)

    2014

  14. [14]

    In: Advances in Neural Information Processing Systems (2016)

    Kingma, D.P., Salimans, T., Jozefowicz, R., Chen, X., Sutskever, I., Welling, M.: Improved variational inference with inverse autoregressive flow. In: Advances in Neural Information Processing Systems (2016)

    2016

  15. [15]

    In: Advances in Neural Information Processing Systems

    Lin, T., Huh, J., Stauffer, C., Lim, S., Isola, P.: Learning to ground multi-agent communication with autoencoders. In: Advances in Neural Information Processing Systems. pp. 15230–15242 (2021)

    2021

  16. [16]

    In: Advances in Neu- ral Information Processing Systems

    Lowe, R., Wu, Y., Tamar, A., Harb, J., Abbeel, P., Mordatch, I.: Multi-agent actor-critic for mixed cooperative-competitive environments. In: Advances in Neu- ral Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)

    2017

  17. [17]

    IEEE Transactions on Automatic Control58(7), 1644–1658 (2013)

    Nayyar, A., Mahajan, A., Teneketzis, D.: Decentralized stochastic control with partial history sharing: A common information approach. IEEE Transactions on Automatic Control58(7), 1644–1658 (2013)

    2013

  18. [18]

    In: 2025 IEEE International Conference on Development and Learning (ICDL)

    Nomura, K., Aoki, T., Taniguchi, T., Horii, T.: Decentralized collective world model for emergent communication and coordination. In: 2025 IEEE International Conference on Development and Learning (ICDL). pp. 1–8 (2025) Title Suppressed Due to Excessive Length 15

    2025

  19. [19]

    arXiv preprint arXiv:1807.03748 (2018)

    van den Oord, A., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

    Pith/arXiv arXiv 2018

  20. [20]

    In: Advances in Neural Information Processing Systems (2016)

    Sukhbaatar, S., Szlam, A., Fergus, R.: Learning multiagent communication with backpropagation. In: Advances in Neural Information Processing Systems (2016)

    2016

  21. [21]

    In: Proceed- ings of the 17th International Conference on Autonomous Agents and MultiAgent Systems

    Sunehag, P., Lever, G., Gruslys, A., Czarnecki, W.M., Zambaldi, V., Jaderberg, M., Lanctot, M., Sonnerat, N., Leibo, J.Z., Tuyls, K., Graepel, T.: Value-decomposition networks for cooperative multi-agent learning based on team reward. In: Proceed- ings of the 17th International Conference on Autonomous Agents and MultiAgent Systems. pp. 2085–2087 (2018)

    2085

  22. [22]

    Frontiers in Robotics and AI11(2024)

    Taniguchi, T.: Collective predictive coding hypothesis: Symbol emergence as de- centralized bayesian inference. Frontiers in Robotics and AI11(2024)

    2024

  23. [23]

    Artificial Life31(4), 465–496 (2025)

    Taniguchi, T., Hirai, Y., Suzuki, M., Murata, S., Horii, T., Tanaka, K.: System 0/1/2/3: Quad-process theory for multitimescale embodied collective cognitive sys- tems. Artificial Life31(4), 465–496 (2025)

    2025

  24. [24]

    Advanced Robotics37(19), 1266–1282 (2023)

    Taniguchi, T., Yoshida, Y., Matsui, Y., Hoang, N.L., Taniguchi, A., Hagiwara, Y.: Emergent communication through metropolis-hastings naming game with deep generative models. Advanced Robotics37(19), 1266–1282 (2023)

    2023

  25. [25]

    de Witt, C.S., Gupta, T., Makoviichuk, D., Makoviychuk, V., Torr, P.H.S., Sun, M., Whiteson, S.: Is independent learning all you need in the StarCraft multi-agent challenge? arXiv preprint arXiv:2011.09533 (2020)

    Pith/arXiv arXiv 2011

  26. [26]

    In: Proceedings of the AAAI Conference on Artificial Intelligence (2023)

    Wu, Z., Yu, C., Chen, C., Hao, J., Zhuo, H.H.: Models as agents: Optimizing multi- step predictions of interactive local models in model-based multi-agent reinforce- ment learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2023)

    2023

  27. [27]

    In: Advances in Neural Information Processing Systems (2022)

    Xu, Z., Li, D., Zhang, B., Zhan, Y., Bai, Y., Fan, G.: Mingling foresight with imag- ination: Model-based cooperative multi-agent reinforcement learning. In: Advances in Neural Information Processing Systems (2022)

    2022

  28. [28]

    In: Neural Information Processing

    Yoshida, N., Taniguchi, T.: Reward-independent messaging for decentralized multi- agent reinforcement learning. In: Neural Information Processing. Lecture Notes in Computer Science, vol. 16309, pp. 367–382. Springer (2026)

    2026

  29. [29]

    Transactions on Machine Learning Research (2025)

    Zhang, Y., Bai, C., Zhao, B., Yan, J., Li, X., Li, X.: Decentralized Transformers with Centralized Aggregation Are Sample-Efficient Multi-Agent World Models. Transactions on Machine Learning Research (2025)

    2025

  30. [30]

    Autonomous Agents and Multi-Agent Systems38(1) (2024)

    Zhu, C., Dastani, M., Wang, S.: A survey of multi-agent deep reinforcement learn- ing with communication. Autonomous Agents and Multi-Agent Systems38(1) (2024)

    2024