Pith. sign in

REVIEW 3 major objections 4 minor 57 references

Towards Language-Augmented Multi-Agent Deep Reinforcement Learning

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that agents trained to describe their observations in a human-defined language while learning a cooperative task outperform agents that develop emergent communication protocols.

desk verdict LAMARL is a competent empirical study, but the headline claim overstates what the ablations support: the oracle language leaks task-relevant supervision, and the main comparison has confounds. read the letter →

arxiv 2506.05236 v2 pith:WGDKXYEG submitted 2025-06-05 cs.MA

Maxime Toquebiau , Jae-Yun Jun , Faïz Benamar , Nicolas Bredeche This is my paper
classification cs.MA
keywords language-augmentedreinforcementlearningmulti-agentemergentcommunicationnaturallanguagegroundingCLIPcontrastiveimagecaptioningMAPPOinterpretable
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 asks whether giving multi-agent reinforcement learning agents a human-defined language to communicate in, rather than letting them invent a private protocol from scratch, makes them better learners and collaborators. The authors propose LAMARL, in which agents are trained concurrently on a cooperative task with MAPPO and on two language objectives: generating descriptions of their observations and aligning observation and language embeddings with a CLIP-style contrastive loss. Across five embodied tasks, language-augmented agents outperform emergent-communication baselines in final performance, learn faster in three tasks, and transmit far less information per message. The sympathetic reading of the claim is that supervised language grounding acts as an inductive bias that produces more structured internal representations and more transferable, interpretable communication.

What carries the argument

The load-bearing mechanism is a multi-objective training loop over an agent architecture that has both a reinforcement-learning pathway and a language pathway. A communication policy compresses the encoded observation into a context vector; a decoder turns that context into tokens of the pre-defined language, trained by a captioning cross-entropy loss; a language encoder and a visual encoder are trained jointly with the CLIP contrastive objective, which pulls matched observation-description pairs together in latent space and pushes mismatched pairs apart; and the whole system is optimized end-to-end with MAPPO for the behavior policy and value function. The complete loss is $L(\theta^i)=\beta_\pi L_\pi(\theta^i)+\beta_V L_V(\theta^i)+\beta_{\text{capt}}L_{\text{capt}}(\theta^i)+\beta_{\text{CLIP}}L_{\text{CLIP}}(\theta^i)$, with dynamic weighting to keep the losses at comparable scales. The oracle function produces the target descriptions: rule-based utterances stating, in cardinal directions, the locations of task-relevant entities such as preys, gems, or colored landmarks. This dual pathway is what carries the argument: the language objectives inject structure into the observation encoders and the communicated context, while the RL objective keeps that structure aligned with reward.

What would settle it

Run LAMARL with an oracle that describes task-irrelevant features, such as the colors of empty grid cells instead of prey locations, while keeping all other training choices fixed; if the agent still outperforms emergent-communication baselines, the claimed benefit does not depend on reward-aligned language, and if it does not, the alignment is the load-bearing factor.

Watch

Extended reading notes

Core claim

The central claim is that language augmentation is not a communication add-on but a training signal that reshapes agent representations. On five embodied tasks, LAMARL consistently outperforms the emergent-communication baselines EC, EC-AutoEncoder, and EC-LangGround, as well as the no-communication control, and it does so while sending on average about 1.19 integer tokens per message versus two or four floating-point numbers for the baselines. The authors show that the communication-policy embeddings form clean, spatially organized clusters aligned with the oracle descriptions, with a silhouette score of 0.59 for LAMARL versus 0.06 for the strongest baseline, and that the benefit persists when communication is removed, since the language-only variant beats the no-language no-communication control. In zero-shot teaming, mixing agents from different training runs degrades LAMARL's success rate far less than it degrades the emergent-communication agents, and in an interaction experiment, human-injected directional messages such as 'north' shift action probabilities in the intended direction. The authors' own summary of the discovery is that jointly optimizing captioning and CLIP objectives with MAPPO makes agents that learn faster, perform better, generalize better to new partners, and can be directed by humans.

Load-bearing premise

The oracle language is hand-designed to describe exactly the task-relevant features that determine reward; if that alignment were weaker or the language less tailored, the advantage over emergent communication could shrink or disappear.

Editorial extensions

If this is right

  • LAMARL consistently outperforms all emergent-communication baselines on all five tasks and learns faster in three of them, so language augmentation is a viable alternative to emergent communication.
  • Because LAMARL sends about 1.19 integer tokens per message, the gains come with a smaller communication bandwidth, suggesting language compresses task-relevant information efficiently.
  • The embedding analysis shows language supervision creates well-separated, spatially coherent clusters in the agents' communication space, implying messages carry structured and interpretable meaning.
  • Zero-shot teaming results indicate language-trained agents retain the best success rate in mixed teams, so sharing a pre-defined language eases coordination with previously unknown partners.
  • In the interaction experiment, human-injected directional messages shift action probabilities in the intended direction, showing that a human can steer a language-augmented agent's behavior.

Reading between the lines

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

  • This is an inference beyond the paper: the advantage likely scales with how well the oracle vocabulary aligns with reward-relevant features, so a mismatched or task-irrelevant oracle should shrink or erase the reported gap.
  • Untested extension: replacing the hand-written oracle with descriptions generated by a language model would test whether the mechanism is the supervised language signal itself or the particular expert-crafted vocabulary.
  • Untested extension: because LAMARL messages are drastically smaller than baseline messages, the language bottleneck may itself contribute to the gains; varying vocabulary size and utterance length would separate compression effects from semantic grounding.
  • Untested extension: the spatial geometry of the embeddings suggests a learned coordinate system, so testing LAMARL on maps larger or differently shaped than the training grid would reveal whether this structure transfers.
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

3 major / 4 minor

Summary. The paper proposes LAMARL, a multi-agent reinforcement learning framework in which agents are trained with MAPPO together with two language objectives: a captioning loss that trains each agent to generate oracle-provided natural-language descriptions of its local observations, and a CLIP-style contrastive loss that trains a language encoder and a visual encoder to align observation and description embeddings. The method is evaluated on five tasks from MA-gym and MPE, comparing against emergent-communication baselines (EC, EC-AutoEncoder, EC-LangGround, and No Comm). The authors report that LAMARL outperforms all baselines, produces better-structured embeddings, generalizes better in zero-shot teaming, and responds to human-injected messages. An ablation study shows that language training helps both when oracle messages are used and when no communication is used. The paper concludes that language grounding improves learning, representation, and coordination.

Significance. If the reported effects are robust, the paper makes a useful contribution by showing that a pre-defined, compositional language can be used as an effective auxiliary objective and communication protocol in cooperative MARL. The submission is commendable for including several empirical components: five tasks, seven seeds, multiple ablations, representation analysis, zero-shot teaming, and a simple human-interaction test, with code released. The main results are visually clear, and the ablation design is thoughtful in isolating the language objective from the message content. However, the central claim that language per se causes the improvement is currently threatened by two confounds: the oracle language provides privileged, task-relevant feature supervision, and LAMARL uses a larger context dimension than the emergent-communication baselines. If these are addressed with additional controls, the contribution would be significant for the multi-agent RL and emergent-communication community.

major comments (3)
  1. [Sec. 4.4, Appendix D, Sec. 6.3] The oracle language used to supervise LAMARL is hand-designed to describe exactly the reward-relevant variables (e.g., prey cardinal direction and type in Predator-Prey, resource colors in Foraging, landmark colors in Coordinated Placement). Consequently LAMARL's captioning loss (Eq. 1) and CLIP loss (Eqs. 2-3) provide dense, direct gradient supervision about these features into the observation encoder and communication policy, whereas the EC and EC-AutoEncoder baselines must discover the same features from scalar reward signals alone. The ablation in Sec. 6.3 (Lang+Oracle vs No Lang+Oracle) holds the content of the oracle messages fixed, but it does not hold the representation-level supervision fixed: only the language-trained variant receives gradients that explicitly name the features. To attribute the advantage to language structure rather than to auxiliary feature supervision, the authors should add a non-linguistic control that receives the same privileged information, for example a classifier or auxiliary regression head trained to predict the oracle labels (prey direction/type, resource color, landmark color) from the communication-policy output or observation embedding, with the same context dimension and capacity as LAMARL. This is load-bearing because the abstract and Sec. 6.1 generalize to 'language-augmented agents outperform emergent communication baselines' without this control.
  2. [Sec. 6.1, Appendix C (Table 2)] The comparison in Fig. 3 uses different context dimensions C for the communication-policy output: C=16 for LAMARL, versus C=2 for EC and EC-AutoEncoder and C=4 for EC-LangGround (Table 2). The paper argues that LAMARL uses a smaller bandwidth because it transmits on average 1.19 discrete tokens, while EC baselines transmit 2 or 4 floating-point numbers. However, the context dimension is the size of the internal representation produced by the communication policy before decoding, so LAMARL has substantially more representational capacity in this module. Since this capacity is not matched across methods, the superiority of LAMARL in Fig. 3 could partly reflect a larger bottleneck/network capacity rather than the language grounding. The authors should either match C across all methods (e.g., C=16 everywhere) or provide an additional ablation in which EC and EC-LangGround use C=16, to separate the effect of the language objective from the effect of communication-policy capacity.
  3. [Sec. 6.1, Fig. 3] The central quantitative claim that 'LAMARL consistently outperforms all baselines across all tasks' is supported only by plots of the median and 95% confidence intervals over 7 runs; no statistical significance tests are reported. Given that some pairs of curves appear close (e.g., LAMARL vs EC in Coordinated Placement in the described results), the reader cannot assess whether the final-performance differences are statistically reliable. The authors should report per-task significance tests on final performance (e.g., Mann-Whitney U test or bootstrap confidence intervals on the difference of medians) or provide effect sizes. This is necessary to back the 'consistently outperforms' formulation.
minor comments (4)
  1. [Sec. 4.3, Eq. (3)] The sentence immediately after Eq. (3) incorrectly describes the CLIP loss: for associated pairs (j=k) the loss term is -cosim, which is minimized by maximizing the cosine similarity, and for unassociated pairs the term is +cosim, which is minimized by decreasing the cosine similarity. Please correct 'minimizes the cosine similarity between related representations, and maximizes the unrelated ones' to reflect the actual objective.
  2. [Appendix C, Table 2] In the row 'Context dimension C' the values appear as '- 2 2 416'; this appears to be a typo for '- 2 2 4 16'. Please fix the formatting so the per-column values are clear.
  3. [Sec. 6.4, Fig. 6 caption] The caption states that the results show the success rate over 1.2×10^6 episodes, while the text says each team composition is evaluated over 25,000 episodes. Please clarify whether 1.2×10^6 is the total across all conditions or correct the number to match the per-condition evaluation.
  4. [Sec. 6.2, Fig. 4] The t-SNE visualizations are qualitative; reporting silhouette scores is useful, but additional quantitative measures of clustering quality (e.g., adjusted Rand index or k-nearest-neighbor accuracy against oracle labels) would make the representation-learning claim more robust.

Circularity Check

1 steps flagged · score 2.0 of 10

Central task-performance claim is empirically self-contained; only the representation-cluster evidence in Section 6.2 is partly enforced by the captioning objective itself.

  1. self definitional [Section 6.2 (Internal representations), Figure 4 discussion]
    "By explicitly learning to generate language utterances, LAMARL learns better structured representation, as shown by the more distinct clusters and the silhouette scores: 0.59 for LAMARL and 0.06 for EC-LangGround."

    The plotted points are the communication-policy outputs that feed the decoder trained with the captioning cross-entropy loss (Eq. 1) to emit the oracle tokens “North/South/East/West/Center”, and the t-SNE colors are exactly those captioning labels. An embedding supervised so that a downstream decoder can recover a label must be separable by that label, so high silhouette with respect to the supervised descriptions is the captioning objective being realized rather than an independent discovery about representation quality. The EC-LangGround embedding (dimension 4) is only pulled toward pretrained language embeddings and is not trained to be decoded into the same vocabulary, so the contrast partly compares a supervised classifier’s latent space with a weaker grounding objective.

full rationale

The central claim that LAMARL outperforms emergent communication baselines is an empirical comparison against external baselines (EC, EC-AutoEncoder, EC-LangGround, No Comm), and the ablations in Section 6.3 (Lang+No Comm vs No Comm; Lang+Oracle vs No Lang+Oracle) isolate the language-training objective from the communication content, so the task-performance result does not reduce to a fit or to a self-citation. The hand-designed oracle (Section 4.4) intentionally names reward-relevant features, which creates an information asymmetry relative to the EC baselines, but that is an experimental-design boundary condition rather than circularity: the paper fits no parameter and then predicts a closely related quantity. The MAPPO, CLIP, and auto-encoding components are standard external results, and no load-bearing self-citation chain or imported uniqueness theorem appears in the derivation. One mild circular flavor remains in Section 6.2: the t-SNE clusters are labeled by the very oracle descriptions that the captioning loss (Eq. 1) trains the embedding to encode, so the measured cluster structure is partly the training objective itself. I score this 2 because the flagged step is supporting analysis, not the central performance claim.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central claim rests on a small set of hand-designed choices: the oracle vocabulary, the context dimension, and the language embedding size. None of these is justified by a sensitivity analysis, and the oracle in particular is a task-specific construction that directly encodes the features that determine reward. The RL and representation-learning assumptions are standard for the field, and no new physical or mathematical entities are introduced.

free parameters (3)
  • Context dimension C = LAMARL: 16; EC and EC-AutoEncoder: 2; EC-LangGround: 4
    Chosen by hand (Appendix C, Table 2). LAMARL has a larger communication bottleneck before decoding, so the comparison between methods does not control for communication channel capacity.
  • Language embedding dimension = 4
    Chosen by hand for the language modules (Appendix C). No sensitivity analysis is reported, and this dimension directly controls the capacity of the CLIP and decoder heads.
  • Task-specific oracle vocabulary and rules = e.g., V={Prey, North, South, East, West, Center}
    The language and the oracle mapping from observations to descriptions are hand-designed per task (Section 4.4, Appendix D). These rules determine exactly what information agents are supervised to communicate, so they are a choice made by the authors rather than an external given.
assumptions (5)
  • domain assumption The Dec-POMDP with broadcasting communication is an appropriate formal model for the studied tasks.
    Used throughout the preliminaries (Section 3) and not validated against alternative models; this is standard practice in the field.
  • domain assumption The oracle maps observations to descriptions deterministically and these descriptions contain the information needed for coordination.
    Section 4.4 defines the oracle for each task; the sufficiency of these descriptions for solving the task is assumed, not proven.
  • domain assumption The constructed vocabulary is compositional and structured enough to serve as a proxy for natural language.
    Section 4.4 states the language is structured, combinatorial, and compositional, but this is a qualitative claim about a toy vocabulary and is not validated against human language properties.
  • domain assumption MAPPO provides a stable and sufficient RL backbone for all variants.
    The paper follows Yu et al. [54] and uses the same PPO update for all agents; no convergence analysis is given.
  • standard math t-SNE embeddings preserve the structure relevant to the clustering claims.
    Section 6.2 uses t-SNE for visualization and silhouette scores for quantification; t-SNE can distort global geometry, though the high-level clustering patterns are plausible.
invented entities (1)
  • Hand-designed oracle language (e.g., 'Prey North')
    purpose: Provides supervised targets for the captioning and CLIP losses, and defines the communication protocol agents learn to generate and interpret.
    The vocabulary and rules are invented by the authors for each task (Appendix D). There is no external evidence that this language is natural or general; it is constructed to encode task-relevant spatial information, so its benefit is partly built in.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Language-Augmented Multi-Agent Deep Reinforcement Learning." pith.science (2026). https://pith.science/paper/WGDKXYEG

@misc{pith2026250605236,
  author       = {Pith},
  title        = {Pith review of: Towards Language-Augmented Multi-Agent Deep Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WGDKXYEG}},
  note         = {Machine review of arXiv:2506.05236}
}
read the original abstract

Most prior works on communication in multi-agent reinforcement learning have focused on emergent communication, which often results in inefficient and non-interpretable systems. Inspired by the role of language in natural intelligence, we investigate how grounding agents in a human-defined language can improve the learning and coordination of embodied agents. We propose a framework in which agents are trained not only to act but also to produce and interpret natural language descriptions of their observations. This language-augmented learning serves a dual role: enabling efficient and interpretable communication between agents, and guiding representation learning. We demonstrate that language-augmented agents outperform emergent communication baselines across various tasks. Our analysis reveals that language grounding leads to more informative internal representations, better generalization to new partners, and improved capability for human-agent interaction. These findings demonstrate the effectiveness of integrating structured language into multi-agent learning and open avenues for more interpretable and capable multi-agent systems.

Figures

Figures reproduced from arXiv: 2506.05236 by the authors.

Figure 1
Figure 1. Language-augmented agent architecture. Each module represents a neural network with a specific purpose. Encoder modules receive incoming information (observations or communication) and embed it in latent representations used in further modules, with dashed self-pointing arrows indicating the use of recurrent neural networks to allow some memorization over multiple time steps. The communication policy selects what in… view at source ↗
Figure 2
Figure 2. Illustration of the oracle’s process for describing observations. In the top-left part is a screenshot of the MA-gym Predator-Prey task (agents in blue and preys in red). Agents observe their position in the grid and the objects in their 5 × 5 observation range (shown in dark blue). The oracle generates a language description describing the location of observed preys. To generate language examples, an oracle functio… view at source ↗
Figure 3
Figure 3. Training performance of LAMARL agents against baselines (7 runs each, with median and 95% confidence interval). LAMARL agents consistently beat the emergent communication baselines. communication baselines. We first look at training performance in classical multi-agent problems and then provide a deeper analysis of the impacts of language learning on our agents. Important hyperpa￾rameters and our code for running al… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Visualizations of embeddings produced by the communication policy in LAMARL (left) and EC-LangGround (right). By explicitly learning to generate language utterances, LAMARL learns better structured representation, as shown by the more distinct clusters and the silhouet…
Figure 5
Figure 5. Figure 5: Training performance of the ablated version on Predator-Prey 18 × 18 (7 runs each, with median and 95% confidence interval). LAMARL agents achieve similar performance as Oracle agents, showing that they succeed in using the language properly. All other ablations are in…
Figure 6
Figure 6. Figure 6: Evaluation performance in the zero-shot teaming experiment. Letters indicate the team composition. Results show the success rate over the evaluation run (1.2 × 106 episodes). and Lang+Oracle are both better than their variant that do not learn language, i.e., No Comm a…
Figure 8
Figure 8. Figure 8: Visualizations of embeddings produced by subsequent layers in the agent architecture in LAMARL (left) and EC-LangGround (right). The top row shows the embedding space after the input multi-layer perceptron of the observation encoder. The middle row shows the embedding …
Figure 9
Figure 9. Figure 9: Action probabilities of agents evaluated in the interaction experiment, when no message is received [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Action probabilities of agents evaluated in the interaction experiment, when corresponding messages are received. The changes in action probabilities displayed in [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 55 canonical work pages

  1. [1]

    In:Science378.6624 (2022), pp

    Anton Bakhtin et al.Human-level play in the game of Diplo- macy by combining language models with strategic rea- soning. In:Science378.6624 (2022), pp. 1067–1074.ISSN: 1095-9203

    work page 2022

  2. [2]

    In:Proceedings of the 2018 Conference on Empiri- cal Methods in Natural Language Processing

    Diane Bouchacourt and Marco Baroni.How agents see things: On visual representations in an emergent language game. In:Proceedings of the 2018 Conference on Empiri- cal Methods in Natural Language Processing. Association for Computational Linguistics, 2018, pp. 981–985

    work page 2018

  3. [3]

    In:Proceedings of Machine Learning Research

    Thomas Carta et al.Grounding Large Language Mod- els in Interactive Environments with Online Reinforcement Learning. In:Proceedings of Machine Learning Research. V ol. 202. 3676-3713. PMLR, 2023

    work page 2023

  4. [4]

    In:Advances in Neural Information Process- ing Systems

    Rahma Chaabouni et al.Anti-efficient encoding in emergent communication. In:Advances in Neural Information Process- ing Systems. Ed. by H. Wallach et al. V ol. 32. Curran Asso- ciates, Inc., 2019

    work page 2019

  5. [5]

    In:Proceedings of the 58th Annual Meeting of the Association for Computational Lin- guistics

    Rahma Chaabouni et al.Compositionality and Generaliza- tion In Emergent Languages. In:Proceedings of the 58th Annual Meeting of the Association for Computational Lin- guistics. Online: Association for Computational Linguistics, 2020, pp. 4427–4442

    work page 2020

  6. [6]

    In:Advances in Neu- ral Information Processing Systems

    Cédric Colas et al.Language as a Cognitive Tool to Imagine Goals in Curiosity Driven Exploration. In:Advances in Neu- ral Information Processing Systems. Ed. by H. Larochelle et al. V ol. 33. Curran Associates, Inc., 2020, pp. 3761–3774

    work page 2020

  7. [7]

    In:Proceedings of the 36th International Con- ference on Machine Learning

    Abhishek Das et al.TarMAC: Targeted Multi-Agent Com- munication. In:Proceedings of the 36th International Con- ference on Machine Learning. Ed. by Kamalika Chaudhuri and Ruslan Salakhutdinov. V ol. 97. Proceedings of Machine Learning Research. PMLR, 2019, pp. 1538–1546

    work page 2019

  8. [8]

    Ferry, Susan J

    Alissa L. Ferry, Susan J. Hespos, and Sandra R. Waxman. Categorization in 3- and 4-Month-Old Infants: An Advantage of Words Over Tones. In:Child Development81.2 (2010), pp. 472–479.ISSN: 1467-8624

    work page 2010

Show all 57 references
  1. [9]

    Foerster et al.Learning to Communicate with Deep Multi-Agent Reinforcement Learning

    Jakob N. Foerster et al.Learning to Communicate with Deep Multi-Agent Reinforcement Learning. In:Proceedings of the 30th International Conference on Neural Information Pro- cessing Systems. NIPS’16. Red Hook, NY , USA: Curran As- sociates Inc., 2016, pp. 2145–2153

    2016

  2. [10]

    Lukas Galke, Yoav Ram, and Limor Raviv.Emergent Com- munication for Understanding Human Language Evolution: What’s Missing?In:Emergent Communication Workshop at ICLR 2022. 2022

    2022

  3. [11]

    Greene and B.R

    J.O. Greene and B.R. Burleson.Handbook of Communication and Social Interaction Skills. ISSN. Taylor & Francis, 2003. ISBN: 9781135664107

    2003

  4. [12]

    In:33rd International Joint Conference on Artificial Intelligence (IJCAI 2024)

    T Guo et al.Large Language Model based Multi-Agents: A Survey of Progress and Challenges. In:33rd International Joint Conference on Artificial Intelligence (IJCAI 2024). 2024

    2024

  5. [13]

    In:Advances in Neural Information Processing Systems

    Abhinav Gupta, Marc Lanctot, and Angeliki Lazaridou.Dy- namic population-based meta-learning for multi-agent com- munication with natural language. In:Advances in Neural Information Processing Systems. Ed. by A. Beygelzimer et al. 2021

    2021

  6. [14]

    In:Proceedings of the 2023 International Conference on Autonomous Agents and Multiagent Systems

    Shuai Han, Mehdi Dastani, and Shihan Wang.Model-Based Sparse Communication in Multi-Agent Reinforcement Learn- ing. In:Proceedings of the 2023 International Conference on Autonomous Agents and Multiagent Systems. AAMAS ’23. London, United Kingdom: International Foundation for ...

    2023

  7. [15]

    Hanjie, Victor Y Zhong, and Karthik Narasimhan

    Austin W. Hanjie, Victor Y Zhong, and Karthik Narasimhan. Grounding Language to Entities and Dynamics for Gener- alization in Reinforcement Learning. In:Proceedings of the 38th International Conference on Machine Learning. Ed. by Marina Meila and Tong Zhang. V ol. 139. Proceed...

    2021

  8. [16]

    In:Proceedings of the 31st Interna- tional Conference on Neural Information Processing Sys- tems

    Serhii Havrylov and Ivan Titov.Emergence of Language with Multi-Agent Games: Learning to Communicate with Se- quences of Symbols. In:Proceedings of the 31st Interna- tional Conference on Neural Information Processing Sys- tems. NIPS’17. Long Beach, California, USA: Curran Asso...

    2017

  9. [17]

    In:Advances in Neural Information Pro- cessing Systems

    Simao Herdade et al.Image Captioning: Transforming Ob- jects into Words. In:Advances in Neural Information Pro- cessing Systems. Ed. by H. Wallach et al. V ol. 32. Curran As- sociates, Inc., 2019

    2019

  10. [18]

    In:International Conference on Learning Representations

    Felix Hill et al.Grounded Language Learning Fast and Slow. In:International Conference on Learning Representations. 2021

    2021

  11. [19]

    In:Ad- vances in Neural Information Processing Systems

    Hengyuan Hu et al.Hierarchical Decision Making by Gener- ating and Following Natural Language Instructions. In:Ad- vances in Neural Information Processing Systems. Ed. by H. Wallach et al. V ol. 32. Curran Associates, Inc., 2019

    2019

  12. [20]

    In:Proceed- ings of The 6th Conference on Robot Learning

    Wenlong Huang et al.Inner Monologue: Embodied Reason- ing through Planning with Language Models. In:Proceed- ings of The 6th Conference on Robot Learning. Ed. by Karen Liu, Dana Kulic, and Jeff Ichnowski. V ol. 205. Proceed- ings of Machine Learning Research. PMLR, 2023, pp. 1...

    2023

  13. [21]

    In:Proceed- ings of the 36th International Conference on Machine Learn- ing

    Natasha Jaques et al.Social Influence as Intrinsic Motivation for Multi-Agent Deep Reinforcement Learning. In:Proceed- ings of the 36th International Conference on Machine Learn- ing. V ol. 97. 2019, pp. 3040–3049

    2019

  14. [22]

    In:Proceedings of the 2023 Inter- national Conference on Autonomous Agents and Multiagent Systems

    Seth Karten, Siva Kailas, and Katia Sycara.Emergent Com- positional Concept Communication through Mutual Informa- tion in Multi-Agent Teams. In:Proceedings of the 2023 Inter- national Conference on Autonomous Agents and Multiagent Systems. AAMAS ’23. , London, United Kingdom, ...

    2023

  15. [23]

    In:International Con- ference on Learning Representations

    Daewoo Kim et al.Learning to Schedule Communication in Multi-agent Reinforcement Learning. In:International Con- ference on Learning Representations. 2019

    2019

  16. [24]

    In:Cognition141 (2015), pp

    Simon Kirby et al.Compression and communication in the cultural evolution of linguistic structure. In:Cognition141 (2015), pp. 87–102.ISSN: 0010-0277

    2015

  17. [25]

    Anurag Koul.ma-gym: Collection of multi-agent environ- ments based on OpenAI gym.2019

    2019

  18. [26]

    Multi-agent Communication meets Natural Language: Syn- ergies between Functional and Structural Language Learn- ing

    Angeliki Lazaridou, Anna Potapenko, and Olivier Tieleman. Multi-agent Communication meets Natural Language: Syn- ergies between Functional and Structural Language Learn- ing. In:Proceedings of the 58th Annual Meeting of the As- sociation for Computational Linguistics. Online: ...

    2020

  19. [27]

    In:The Thirty-eighth Annual Conference on Neural Informa- tion Processing Systems

    Huao Li et al.Language Grounded Multi-agent Reinforce- ment Learning with Human-interpretable Communication. In:The Thirty-eighth Annual Conference on Neural Informa- tion Processing Systems. 2024

    2024

  20. [28]

    In: (2023)

    Jessy Lin et al.Learning to Model the World with Language. In: (2023). arxiv:2308.01399. eprint: 2308.01399

    2023 arXiv

  21. [29]

    In:Advances in Neural Information Processing Systems

    Toru Lin et al.Learning to Ground Multi-Agent Communica- tion with Autoencoders. In:Advances in Neural Information Processing Systems. 2021

    2021

  22. [30]

    Davison.End-To- End Multi-Task Learning With Attention

    Shikun Liu, Edward Johns, and Andrew J. Davison.End-To- End Multi-Task Learning With Attention. In:Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2019

    2019

  23. [31]

    In:Advances in Neural Information Processing Systems

    Ryan Lowe et al.Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments. In:Advances in Neural Information Processing Systems. V ol. 30. Long Beach, California, USA, 2017

    2017

  24. [32]

    In:International Con- ference on Learning Representations

    Ryan Lowe et al.On the interaction between supervision and self-play in emergent communication. In:International Con- ference on Learning Representations. 2020

    2020

  25. [33]

    In:Proceedings of the 18th International Conference on Autonomous Agents and MultiAgent Systems

    Ryan Lowe et al.On the Pitfalls of Measuring Emergent Communication. In:Proceedings of the 18th International Conference on Autonomous Agents and MultiAgent Systems. AAMAS ’19. Montreal QC, Canada: International Founda- tion for Autonomous Agents and Multiagent Systems, 2019, ...

    2019

  26. [34]

    In:The Psychology of Learn- ing and Motivation

    Gary Lupyan.What Do Words Do? Toward a Theory of Language-Augmented Thought. In:The Psychology of Learn- ing and Motivation. Ed. by Brian H. Ross. V ol. 57. Psy- chology of Learning and Motivation. Academic Press, 2012, pp. 255–297

    2012

  27. [35]

    In:Journal of Machine Learning Research 9 (2008), pp

    Laurens van der Maaten and Geoffrey Hinton.Visualizing data using t-SNE. In:Journal of Machine Learning Research 9 (2008), pp. 2579–2605

    2008

  28. [36]

    Springer, 2016

    Frans A Oliehoek and Christopher Amato.A concise intro- duction to decentralized POMDPs. Springer, 2016

    2016

  29. [37]

    In:Proceedings of the 36th An- nual ACM Symposium on User Interface Software and Tech- nology

    Joon Sung Park et al.Generative Agents: Interactive Simu- lacra of Human Behavior. In:Proceedings of the 36th An- nual ACM Symposium on User Interface Software and Tech- nology. UIST ’23. , San Francisco, CA, USA, Association for Computing Machinery, 2023

    2023

  30. [38]

    In: (2024)

    Jérémy Perez et al.Cultural evolution in populations of Large Language Models. In: (2024). arxiv:2403.08882

    2024 arXiv

  31. [39]

    Piaget.La formation du symbole chez l’enfant: Imitation, jeu et rêve, image et représentation

    J. Piaget.La formation du symbole chez l’enfant: Imitation, jeu et rêve, image et représentation. Originally published in

  32. [40]

    In:Behavioral and Brain Sciences13.4 (1990), pp

    Steven Pinker and Paul Bloom.Natural language and natu- ral selection. In:Behavioral and Brain Sciences13.4 (1990), pp. 707–727.ISSN: 1469-1825

    1990

  33. [41]

    In:Proceedings of the 38th International Conference on Machine Learn- ing

    Alec Radford et al.Learning Transferable Visual Mod- els From Natural Language Supervision. In:Proceedings of the 38th International Conference on Machine Learn- ing. V ol. 139. Proceedings of Machine Learning Research. PMLR, 2021, pp. 8748–8763

    2021

  34. [42]

    In: (2016)

    John Schulman et al.High-Dimensional Continuous Control Using Generalized Advantage Estimation. In: (2016). Ed. by Yoshua Bengio and Yann LeCun

    2016

  35. [43]

    In: (2017)

    John Schulman et al.Proximal Policy Optimization Algo- rithms. In: (2017)

    2017

  36. [44]

    Searcy and Stephen Nowicki.The Evolution of Animal Communication: Reliability and Deception in Sig- naling Systems

    William A. Searcy and Stephen Nowicki.The Evolution of Animal Communication: Reliability and Deception in Sig- naling Systems. Princeton University Press, 2010.ISBN: 9781400835720

    2010

  37. [45]

    Learning when to Communicate at Scale in Multiagent Coop- erative and Competitive Tasks

    Amanpreet Singh, Tushar Jain, and Sainbayar Sukhbaatar. Learning when to Communicate at Scale in Multiagent Coop- erative and Competitive Tasks. In:International Conference on Learning Representations. 2019

    2019

  38. [46]

    Smith and D

    J.M. Smith and D. Harper.Animal Signals. Animal Signals. OUP Oxford, 2003.ISBN: 9780198526858

    2003

  39. [47]

    In:Evolution of Communication1.1 (1997), pp

    Luc Steels.The Synthetic Modeling of Language Origins. In:Evolution of Communication1.1 (1997), pp. 1–34.ISSN: 1569-9757

    1997

  40. [48]

    In: Advances in Neural Information Processing Systems

    Sainbayar Sukhbaatar, Arthur Szlam, and Rob Fergus.Learn- ing Multiagent Communication with Backpropagation. In: Advances in Neural Information Processing Systems. 2016, pp. 2244–2252

    2016

  41. [49]

    Harvard university press, 2009

    Michael Tomasello.The cultural origins of human cognition. Harvard university press, 2009

    2009

  42. [50]

    In:Advances in Neural Information Process- ing Systems

    Mycal Tucker et al.Emergent Discrete Communication in Se- mantic Spaces. In:Advances in Neural Information Process- ing Systems. Ed. by A. Beygelzimer et al. 2021

    2021

  43. [51]

    In:IEEE Transactions on Pattern Analysis and Machine Intelligence(2021).ISSN: 1939-3539

    Simon Vandenhende et al.Multi-Task Learning for Dense Prediction Tasks: A Survey. In:IEEE Transactions on Pattern Analysis and Machine Intelligence(2021).ISSN: 1939-3539

    2021

  44. [52]

    Vygotsky.Thought and Language

    Lev S. Vygotsky.Thought and Language. MIT Press, 1934

    1934

  45. [53]

    In:Proceed- ings of the 2023 International Conference on Autonomous Agents and Multiagent Systems

    Nathaniel Weir et al.One-Shot Learning from a Demon- stration with Hierarchical Latent Language. In:Proceed- ings of the 2023 International Conference on Autonomous Agents and Multiagent Systems. AAMAS ’23. London, United Kingdom: International Foundation for Autonomous Agents...

    2023

  46. [54]

    In: 35 (2021)

    Chao Yu et al.The Surprising Effectiveness of PPO in Coop- erative, Multi-Agent Games. In: 35 (2021). Ed. by S. Koyejo et al., pp. 24611–24624

    2021

  47. [55]

    In:Advances in Neural Information Process- ing Systems

    Sai Qian Zhang, Qi Zhang, and Jieyu Lin.Efficient Commu- nication in Multi-Agent Reinforcement Learning via Variance Based Control. In:Advances in Neural Information Process- ing Systems. Ed. by H. Wallach et al. V ol. 32. Curran Asso- ciates, Inc., 2019

    2019

  48. [56]

    predator

    Changxi Zhu, Mehdi Dastani, and Shihan Wang.A survey of multi-agent deep reinforcement learning with communica- tion. In:Autonomous Agents and Multi-Agent Systems38.1 (2024).ISSN: 1573-7454. Appendices A Proximal Policy Optimization Loss Here we provide details on the training...

    2024

  49. [1945]

    Delachaux et Niestlé, 1978.ISBN: 9791036909528

    1978

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.