Pith. sign in

REVIEW 4 major objections 6 minor 40 references

Modeling Latent Partner Strategies for Adaptive Zero-Shot Human-Agent Collaboration

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read An agent that learns latent partner strategies and tracks them with fixed-share regret beats population-trained baselines when paired with unfamiliar humans in a timed Overcooked coordination task.

desk verdict A credible combination of latent partner modeling and online tracking for zero-shot coordination, with a human study that mostly backs the claim, but the cluster-coverage assumption is the thing to press on. read the letter →

arxiv 2507.05244 v1 pith:JSZGVZCY submitted 2025-07-07 cs.AI cs.MA

Benjamin Li , Shuyang Shi , Lucia Romero , Huao Li , Yaqi Xie , Woojun Kim , Stefanos Nikolaidis , Michael Lewis
show 2 more authors
Katia Sycara Simon Stepputtis
This is my paper · ORCID
classification cs.AIcs.MA
keywords zero-shotcoordinationhuman-agentcollaborationlatentstrategyspacevariationalautoencoderregretminimizationadhocteamworkOvercookedinference
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 tries to establish that an agent can coordinate from scratch with an unfamiliar human partner by learning a latent space of partner strategies from offline agent-agent play, partitioning it into a small set of distinct strategy types, and adapting its belief about which type is active while playing. The setting matters because ad-hoc human-agent teamwork under time pressure cannot rely on pre-arranged conventions; the agent must read the partner and adjust its own behavior in real time. The TALENTS framework trains a variational autoencoder on trajectories from diverse synthetic populations, clusters the resulting latent space, trains a strategy-conditioned cooperator, and uses a fixed-share regret update to keep switching among strategy clusters as the partner's actions arrive. In an online user study with 119 participants, the agent scored significantly higher than GAMMA and a population best-response baseline, and participants rated it higher on team fluency and trust. If correct, the method shows that partner modeling can be built without human interaction data at training time and still generalize to real people at test time.

What carries the argument

The load-bearing object is a discrete latent strategy space paired with a tracking no-regret update. A sequential variational autoencoder, trained on observation-action trajectories of synthetic agents, encodes each partner's recent behavior into a latent code; K-means clustering with silhouette analysis partitions that space into Gaussian clusters, each treated as a distinct partner type or 'expert.' During training, the decoder generates synthetic partners by sampling from cluster means, and a PPO-trained cooperator learns a per-cluster action-bias vector that tilts its action logits toward the response appropriate to that type. At test time the agent samples a latent from each cluster, predicts the partner's next action with the decoder, scores that prediction by negative log-likelihood against the partner's actual action, and updates cluster weights with a fixed-share rule that allows belief to switch between experts over time. The fixed-share rule is what turns a static strategy library into an online mechanism with tracking-regret guarantees.

What would settle it

Recruit participants who deliberately alternate between two learned strategy clusters (for example, switching every thirty seconds between a dish-washing-biased and a rice-boiling-biased style) and compare TALENTS' team score with its score against participants who stick to one cluster; a large drop for the alternating group would show that between-cluster or switching behavior breaks the discrete-cluster representation, while a negligible drop would support the fixed-share adaptation claim.

Watch

Extended reading notes

Core claim

TALENTS claims that the behavior of partners never seen during training, including people, can be compressed into a discrete latent strategy space, and that online adaptation over that space yields better coordination than simply training a cooperator on a diverse population. In a modified Overcooked environment with expiring orders, bonus rewards for fast delivery, and three cooking stations, the paper reports that TALENTS outperforms GAMMA and a population best-response (BR) cooperator with held-out Behavior Preference partners on three of four layouts and overall, with BR winning only in the Forced Coordination layout. In the human-agent study, the team-score advantage is significant ($F(2,166)=5.76$, $p=.003$), and TALENTS receives higher team-fluency and trust ratings than the BR baseline. The paper further argues the fixed-share update is essential: when a partner policy is switched mid-episode, the full agent recovers its belief and maintains reward, while a static Hedge variant does not. The intended reading is that strategy representation plus tracking-regret adaptation, not larger training populations, is what makes zero-shot coordination with novel partners work.

Load-bearing premise

A finite set of Gaussian clusters learned from synthetic-agent trajectories is a sufficient stand-in for the full space of human partner strategies; if a real person's behavior falls between or outside those clusters, the fixed-share update can only converge to the nearest training cluster, and the regret bound does not by itself guarantee good coordination with that person.

Editorial extensions

If this is right

  • The fixed-share update lets the agent track a partner whose strategy changes mid-episode, whereas a static Hedge update keeps its belief pinned to the original strategy and loses cumulative reward after the switch.
  • The strategy space is learned offline from synthetic populations (FCP, MEP, and BP trajectories), so the framework needs no human interaction data during training while still coordinating with unfamiliar human partners at test time.
  • In agent-agent tests on held-out Behavior Preference partners, TALENTS scores higher than GAMMA and the population best-response baseline overall and on most layouts; the best-response baseline wins only in the Forced Coordination layout.
  • In human-agent play, TALENTS produces significantly higher team scores and higher subjective team fluency and trust than the best-response baseline.
  • The tracking-regret framing implies that performance with a genuinely novel partner is not worse than the best response to the closest trained cluster, up to the algorithm's regret bound.

Reading between the lines

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

  • Because test-time inference picks among the learned clusters, a human whose behavior lies between two clusters will be matched to the nearest one; a natural extension is continuous or interpolated latent adaptation, which the paper does not test.
  • The framework's reach into real human play depends on how well the synthetic populations' diversity spans human behavior; deliberately recruiting unusual or adversarial human strategies would clarify where that boundary lies.
  • The conditioning mechanism is a per-cluster action-bias on the actor's logits, so the inferred partner type shifts action preference rather than the policy's state-dependent reasoning; richer conditioning may matter when the correct response to a strategy is situation-specific.
  • The same fixed-share machinery could apply per teammate in larger teams, but the single-belief update assumes one active strategy at a time, so multi-partner teams would require factored beliefs or per-partner expert streams.
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

4 major / 6 minor

Summary. The paper introduces TALENTS, a zero-shot coordination agent for human-agent teamwork. The method learns a variational-autoencoder latent strategy space from offline trajectories of FCP, MEP, and BP synthetic partners, partitions the latent space with K-means clustering, trains a PPO cooperator conditioned on cluster identity, and at test time uses a fixed-share tracking-regret update to infer and adapt to the partner's strategy online. Experiments in a customized Overcooked domain compare TALENTS against GAMMA and a population best-response agent in both agent-agent and human-agent settings; the online human study reports significantly higher team scores and better subjective fluency and trust ratings for TALENTS than for the baselines.

Significance. The problem—adapting to heterogeneous human partners in zero-shot coordination—is important, and the paper's empirical scope is substantial: four layouts, three synthetic partner populations, an online human study, and an ablation comparing fixed-share adaptation to static regret. The agent-agent results support the method on three of four layouts and in aggregate, and the human study is a potentially convincing demonstration if the statistical reporting is correct. The main contribution, combining a learned latent strategy space with fixed-share expert switching for online partner inference, is a reasonable extension of existing HiP-MDP and no-regret ideas. The missing link is evidence that the discrete cluster support learned from synthetic populations covers real human strategies; without that evidence, the central human-facing claim is not fully established.

major comments (4)
  1. [Section III-C, Algorithm 1] The transfer from tracking regret to team-performance degradation is asserted, not derived. The sentence at the end of Section III-C, 'This provides regret bounds ensuring that encountering truly novel strategies only degrades performance to the best-response of the closest observed training strategy,' does not follow from the fixed-share bound: Herbster and Warmuth's fixed-share algorithm bounds a cumulative prediction loss over experts, not the team reward obtained under a policy that lies between or outside the K learned clusters. Because Algorithm 1 restricts online adaptation to the K cluster-conditioned policies and action-bias vectors b(c*), a human strategy not represented by a cluster center will not receive an appropriate best response. Please supply either a formal argument or empirical coverage evidence, e.g., project human-agent trajectories through q_phi(z|tau) and report distances to the learned cluster centers, or report team-score regret relative to an oracle that knows the human's true latent strategy.
  2. [Section IV-B, Results] The human-team-score ANOVA reporting is incomplete and appears inconsistent with the described design. For a mixed design with three within-subject agent levels and three between-subject layout groups, N=119 participants would give within-subject error degrees of freedom (119-3)*2=232, not the reported 166. The reported F(2,166) implies a final sample of N=86 after exclusions, but the paper does not state the final N, the number excluded, the layout cell sizes, or the exclusion timing relative to the questionnaires. Please report these values along with effect sizes, and state whether agent-round order was counterbalanced and whether agent type interacted with layout.
  3. [Section IV-A, Table I] The agent-agent results do not support the unqualified 'state-of-the-art zero-shot performance' language in the abstract and conclusion. Table I shows that the population best-response agent is significantly better than TALENTS on the Forced Coordination layout (69.70 vs. 28.26 and 41.10), and the text's own hypothesis in Section IV-A concedes a structural failure mode. Because the human study includes Forced Coordination as one of the three layouts, the paper should either report layout-specific human scores and discuss whether this failure transfers, or qualify the overall claims so that they are consistent with the presented evidence.
  4. [Section III-B and Algorithm 1] The strategy-conditioned policy is not described precisely enough to be reproducible. In Section III-B, the cooperator is trained using sampled cluster-mean latents plus an action-bias vector, but Algorithm 1 samples z_c from each cluster Gaussian for prediction, then at step 9 defines pi_theta(·|o_t, c*) = pi_theta(·|o_t) + b(c*), while step 10 writes pi_theta(·|o_t, c*) as though the policy takes c* as input. It is unclear whether the actual policy input is the discrete cluster index, the sampled z_c, or neither, and how the decoder's z_c relates to b(c*). Please clarify the exact network inputs and the role of the sampled z_c at execution time.
minor comments (6)
  1. [Abstract and Conclusion] The phrase 'state-of-the-art zero-shot performance' is too strong for a single environment with two baselines; consider qualifying it as 'strong performance' or 'improvement over the evaluated baselines.'
  2. [Section IV-A, Fig. 4] The ablation figure lacks error bars, confidence intervals, and a statement of how many seeds or episodes underlie the cumulative-reward curves; it also does not define 'second half of the episode' in timesteps.
  3. [Section III-A] The chosen number of clusters K from silhouette analysis is never reported, although K is a load-bearing hyperparameter of Algorithm 1; report K for each training population or for the combined population.
  4. [Section III-A] The VAE context length h and prediction horizon H are not specified; these affect what the strategy representation captures and should be stated.
  5. [Section IV-B] The subjective-rating ANOVAs are reported with different denominator degrees of freedom (e.g., F(2,122)) than the team-score ANOVA; please explain the source of the discrepancy, since it affects the reader's ability to verify the mixed-model analysis.
  6. [Section IV-B] The agent action-rate matching to 'average human players' is not quantified; report the actual rate and how it was measured.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; evaluation is external to training and no fitted value is relabeled as a prediction.

full rationale

TALENTS's derivation chain is empirical and self-contained rather than circular. The latent strategy space is learned by maximizing a VAE ELBO over offline trajectory data (Sec. III-A), clusters are chosen by K-means with silhouette analysis, the cooperator is trained with PPO conditioned on cluster samples, and test-time adaptation uses the fixed-share algorithm with VAE action-prediction losses (Algorithm 1). None of these components is defined in terms of the evaluation metric (team score), and no evaluation result is used to fit cluster centers, VAE parameters, or action-bias vectors. The human-agent study is external to training: human participants were unseen during model fitting, and no human data are fed back into the latent space or policy. The synthetic agent-agent evaluation uses distinct held-out policies from the training populations, and the fixed-share algorithm is a standard external tracking-regret method; its cited regret guarantees are not taken from the authors' own prior work. The paper's self-citations appear only in related-work discussions and are not load-bearing for the central claim. The statement that novel strategies 'only degrade performance to the best-response of the closest observed training strategy' is an asserted coverage assumption rather than a circular derivation; whether real human strategies project near the learned synthetic clusters is a validity risk, not a circularity, because the human team scores are measured directly rather than derived from the clusters. No equation is equivalent to another by construction, and no fitted parameter is renamed as a prediction. Accordingly, the paper does not meet the evidence bar for any circular step.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the assumption that synthetic agent populations trained with reward shaping, especially BP agents, are a sufficient proxy for human partner strategies, so the VAE latent clusters learned from them transfer to real humans. Additional assumptions are that a discrete cluster assignment and next-action predictive loss capture enough of partner behavior for conditioning, and that the fixed-share weight update converges to a useful best response. These are domain assumptions, not standard mathematical facts.

free parameters (4)
  • K (number of strategy clusters) = not reported
    Chosen via silhouette analysis in Section III-A; directly defines the granularity of partner types used for inference and conditioning.
  • beta (VAE regularization weight) = not reported
    Weights the KL term in the ELBO; not reported, but it shapes the latent strategy space.
  • alpha (fixed-share switching parameter) = not reported
    Controls how quickly the belief can switch among clusters in Algorithm 1; affects online adaptation speed.
  • eta (expert weight learning rate) = not reported
    Determines the update step size for cluster weights in Algorithm 1; not specified.
assumptions (4)
  • domain assumption Behavior Preference (BP) agent populations trained with reward shaping approximate the diversity of human team strategies.
    Justifies training the VAE and cooperator on synthetic populations and transferring to humans; supported only by a citation to Wang et al. [36], not by evidence in this paper.
  • domain assumption A discrete set of K Gaussian clusters in the VAE latent space covers the strategies of novel human partners.
    Test-time inference in Algorithm 1 selects among these clusters; if a human strategy lies outside them, adaptation can only approach the closest training cluster.
  • domain assumption Next-action predictive loss is a sufficient statistic for identifying a partner's strategy.
    The VAE is trained to predict the next high-level action, and cluster losses in Algorithm 1 use only log pθ(a_t|z,o_t). This ignores other behavioral cues that might matter for coordination.
  • ad hoc to paper Fixed-share tracking regret implies that novel partner strategies degrade performance only to the best response of the closest observed training strategy.
    Stated in Section III-C without proof and not a direct consequence of the cited fixed-share regret bound, which bounds tracking regret against the best expert, not performance against arbitrary novel policies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Modeling Latent Partner Strategies for Adaptive Zero-Shot Human-Agent Collaboration." pith.science (2026). https://pith.science/paper/JSZGVZCY

@misc{pith2026250705244,
  author       = {Pith},
  title        = {Pith review of: Modeling Latent Partner Strategies for Adaptive Zero-Shot Human-Agent Collaboration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JSZGVZCY}},
  note         = {Machine review of arXiv:2507.05244}
}
read the original abstract

In collaborative tasks, being able to adapt to your teammates is a necessary requirement for success. When teammates are heterogeneous, such as in human-agent teams, agents need to be able to observe, recognize, and adapt to their human partners in real time. This becomes particularly challenging in tasks with time pressure and complex strategic spaces where the dynamics can change rapidly. In this work, we introduce TALENTS, a strategy-conditioned cooperator framework that learns to represent, categorize, and adapt to a range of partner strategies, enabling ad-hoc teamwork. Our approach utilizes a variational autoencoder to learn a latent strategy space from trajectory data. This latent space represents the underlying strategies that agents employ. Subsequently, the system identifies different types of strategy by clustering the data. Finally, a cooperator agent is trained to generate partners for each type of strategy, conditioned on these clusters. In order to adapt to previously unseen partners, we leverage a fixed-share regret minimization algorithm that infers and adjusts the estimated partner strategy dynamically. We assess our approach in a customized version of the Overcooked environment, posing a challenging cooperative cooking task that demands strong coordination across a wide range of possible strategies. Using an online user study, we show that our agent outperforms current baselines when working with unfamiliar human partners.

Figures

Figures reproduced from arXiv: 2507.05244 by the authors.

Figure 1
Figure 1. The TALENTS framework: Given an unknown team [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The four Overcooked layouts used in experiments. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Human-agent teamwork evaluation results are measured using team score and participants’ subjective ratings of their [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Accumulated reward in the first and second halves [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 35 canonical work pages

  1. [1]

    On the utility of learning about humans for human-ai coordina- tion

    Micah Carroll, Rohin Shah, Mark K Ho, Tom Griffiths, Sanjit Seshia, Pieter Abbeel, and Anca Dragan. On the utility of learning about humans for human-ai coordina- tion. Advances in neural information processing systems, 32, 2019. 2, 3

    work page 2019

  2. [2]

    Coordina- tion in multiagent reinforcement learning: A bayesian approach

    Georgios Chalkiadakis and Craig Boutilier. Coordina- tion in multiagent reinforcement learning: A bayesian approach. In Proceedings of the second international joint conference on Autonomous agents and multiagent systems, pages 709–716, 2003. 2

    work page 2003

  3. [3]

    Generating diverse cooperative agents by learning incompatible policies

    Rujikorn Charakorn, Poramate Manoonpong, and Nat Dilokthanakul. Generating diverse cooperative agents by learning incompatible policies. In The Eleventh International Conference on Learning Representations ,

    work page

  4. [4]

    Is independent learning all you need in the starcraft multi-agent chal- lenge? 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 chal- lenge? arXiv preprint arXiv:2011.09533 , 2020. 3

    arXiv 2011

  5. [5]

    Hidden parameter markov decision processes: A semiparamet- ric regression approach for discovering latent task parametrizations

    Finale Doshi-Velez and George Konidaris. Hidden parameter markov decision processes: A semiparamet- ric regression approach for discovering latent task parametrizations. In IJCAI: proceedings of the confer- ence, volume 2016, page 1432, 2016. 2

    work page 2016

  6. [6]

    Learning with opponent- learning awareness

    J Foerster, R Chen, M Al-Shedivat, S Whiteson, P Abbeel, and I Mordatch. Learning with opponent- learning awareness. Autonomous Agents and Multi-Agent Systems (AAMAS 2018) , 2018. 2

    work page 2018

  7. [7]

    Bayesian action de- coder for deep multi-agent reinforcement learning

    Jakob Foerster, Francis Song, Edward Hughes, Neil Burch, Iain Dunning, Shimon Whiteson, Matthew Botvinick, and Michael Bowling. Bayesian action de- coder for deep multi-agent reinforcement learning. In International Conference on Machine Learning , pages 1942–1951. PMLR, 2019. 2

    work page 1942

  8. [8]

    A decision-theoretic generalization of on-line learning and an application to boosting

    Yoav Freund and Robert E Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. Journal of computer and system sciences , 55 (1):119–139, 1997. 3

    work page 1997

Show all 40 references
  1. [9]

    Learning policy represen- tations in multiagent systems

    Aditya Grover, Maruan Al-Shedivat, Jayesh Gupta, Yuri Burda, and Harrison Edwards. Learning policy represen- tations in multiagent systems. In International conference on machine learning, pages 1802–1811. PMLR, 2018. 2

    2018

  2. [10]

    Development of NASA-TLX (task load index): Results of empirical and theoretical research

    Sandra G Hart and Lowell E Staveland. Development of NASA-TLX (task load index): Results of empirical and theoretical research. In Advances in Psychology , Advances in psychology, pages 139–183. Elsevier, 1988. 4

    1988

  3. [11]

    Opponent modeling in deep reinforcement learning

    He He, Jordan Boyd-Graber, Kevin Kwok, and Hal Daum´e III. Opponent modeling in deep reinforcement learning. In International conference on machine learn- ing, pages 1804–1813. PMLR, 2016. 2

    2016

  4. [12]

    Fictitious self-play in extensive-form games

    Johannes Heinrich, Marc Lanctot, and David Silver. Fictitious self-play in extensive-form games. In Interna- tional conference on machine learning , pages 805–813. PMLR, 2015. 2

    2015

  5. [13]

    Tracking the best expert

    Mark Herbster and Manfred K Warmuth. Tracking the best expert. Machine learning, 32(2):151–178, 1998. 3

    1998

  6. [14]

    Learning to influence human behavior with offline reinforcement learning

    Joey Hong, Sergey Levine, and Anca Dragan. Learning to influence human behavior with offline reinforcement learning. Advances in Neural Information Processing Systems, 36:36094–36105, 2023. 2

    2023

  7. [15]

    other-play

    Hengyuan Hu, Adam Lerer, Alex Peysakhovich, and Jakob Foerster. “other-play” for zero-shot coordination. In International Conference on Machine Learning, pages 4399–4410. PMLR, 2020. 2

    2020

  8. [16]

    Human-level performance in 3d multiplayer games with population-based reinforcement learning

    Max Jaderberg, Wojciech M Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio Garcia Castaneda, Charles Beattie, Neil C Rabinowitz, Ari S Morcos, Avra- ham Ruderman, et al. Human-level performance in 3d multiplayer games with population-based reinforcement learning. Scien...

    2019

  9. [17]

    Auto-encoding variational bayes, 2013

    Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. 2

    2013

  10. [18]

    Theory of mind for multi-agent col- laboration via large language models

    Huao Li, Yu Quan Chong, Simon Stepputtis, Joseph Campbell, Dana Hughes, Charles Michael Lewis, and Katia P Sycara. Theory of mind for multi-agent col- laboration via large language models. In The 2023 Conference on Empirical Methods in Natural Language Processing. 2

    2023

  11. [19]

    Individualized mutual adaptation in human- agent teams

    Huao Li, Tianwei Ni, Siddharth Agrawal, Fan Jia, Suhas Raja, Yikang Gui, Dana Hughes, Michael Lewis, and Ka- tia Sycara. Individualized mutual adaptation in human- agent teams. IEEE Transactions on Human-Machine Systems, 51(6):706–714, 2021. doi: 10.1109/THMS. 2021.3107675. 2

    2021

  12. [20]

    Theory of mind modeling in search and rescue teams

    Huao Li, Ini Oguntola, Dana Hughes, Michael Lewis, and Katia Sycara. Theory of mind modeling in search and rescue teams. In 2022 31st IEEE International Conference on Robot and Human Interactive Communi- cation (RO-MAN), pages 483–489, 2022. doi: 10.1109/ RO-MAN53752.2022.9900572. 2

    2022

  13. [21]

    Learning to cooperate with humans using generative agents

    Yancheng Liang, Daphne Chen, Abhishek Gupta, Si- mon Shaolei Du, and Natasha Jaques. Learning to cooperate with humans using generative agents. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2, 3

  14. [22]

    Trajectory diversity for zero-shot coordination

    Andrei Lupu, Brandon Cui, Hengyuan Hu, and Jakob Foerster. Trajectory diversity for zero-shot coordination. In International conference on machine learning , pages 7204–7213. PMLR, 2021. 2

    2021

  15. [23]

    Follow-the-regularized-leader and mirror descent: Equivalence theorems and l1 regular- ization

    Brendan McMahan. Follow-the-regularized-leader and mirror descent: Equivalence theorems and l1 regular- ization. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 525–533. JMLR Workshop and Conference Proceedings,

  16. [24]

    Efficient model learning from joint-action demonstrations for human-robot collaborative tasks

    Stefanos Nikolaidis, Ramya Ramakrishnan, Keren Gu, and Julie Shah. Efficient model learning from joint-action demonstrations for human-robot collaborative tasks. In Proceedings of the tenth annual ACM/IEEE international conference on human-robot interaction , pages 189–196,

  17. [25]

    Theory of mind as intrinsic motivation for multi-agent reinforcement learning

    Ini Oguntola, Joseph Campbell, Simon Stepputtis, and Katia P Sycara. Theory of mind as intrinsic motivation for multi-agent reinforcement learning. In First Workshop on Theory of Mind in Communicating Agents . 2

  18. [26]

    Variational autoencoders for opponent modeling in multi-agent sys- tems

    Georgios Papoudakis and Stefano V Albrecht. Variational autoencoders for opponent modeling in multi-agent sys- tems. arXiv preprint arXiv:2001.10829 , 2020. 2

    2001 arXiv

  19. [27]

    Machine theory of mind

    Neil Rabinowitz, Frank Perbet, Francis Song, Chiyuan Zhang, SM Ali Eslami, and Matthew Botvinick. Machine theory of mind. In International conference on machine learning, pages 4218–4227. PMLR, 2018. 2

    2018

  20. [28]

    Modeling others using oneself in multi-agent reinforcement learning

    Roberta Raileanu, Emily Denton, Arthur Szlam, and Rob Fergus. Modeling others using oneself in multi-agent reinforcement learning. In International conference on machine learning, pages 4257–4266. PMLR, 2018. 2

    2018

  21. [29]

    Silhouettes: a graphical aid to the interpretation and validation of cluster analysis

    Peter J Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics , 20:53–65,

  22. [30]

    A. L. Samuel. Some studies in machine learning using the game of checkers. IBM Journal of Research and Development, 3(3):210–229, 1959. doi: 10.1147/rd.33

    1959 doi

  23. [31]

    Proximal policy optimization algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 , 2017. 3

    2017 arXiv

  24. [32]

    Mastering the game of go without human knowledge

    David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without human knowledge. nature, 550(7676):354–359, 2017. 2

    2017

  25. [33]

    Ad hoc autonomous agent teams: Collab- oration without pre-coordination

    Peter Stone, Gal Kaminka, Sarit Kraus, and Jeffrey Rosenschein. Ad hoc autonomous agent teams: Collab- oration without pre-coordination. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 24, pages 1504–1509, 2010. 1

    2010

  26. [34]

    Collaborating with humans without human data

    DJ Strouse, Kevin McKee, Matt Botvinick, Edward Hughes, and Richard Everett. Collaborating with humans without human data. Advances in Neural Information Processing Systems, 34:14502–14515, 2021. 2, 3

    2021

  27. [35]

    Grandmaster level in starcraft ii using multi-agent reinforcement learning

    Oriol Vinyals, Igor Babuschkin, Wojciech M Czarnecki, Micha¨el Mathieu, Andrew Dudzik, Junyoung Chung, David H Choi, Richard Powell, Timo Ewalds, Petko Georgiev, et al. Grandmaster level in starcraft ii using multi-agent reinforcement learning. nature, 575(7782): 350–354, 2019. 2

    2019

  28. [36]

    Zsc-eval: An evaluation toolkit and benchmark for multi- agent zero-shot coordination

    Xihuai Wang, Shao Zhang, Wenhao Zhang, Wentao Dong, Jingxiao Chen, Ying Wen, and Weinan Zhang. Zsc-eval: An evaluation toolkit and benchmark for multi- agent zero-shot coordination. Advances in Neural Infor- mation Processing Systems , 37:47344–47377, 2024. 2, 3

    2024

  29. [37]

    Learning latent representations to influence multi-agent interaction

    Annie Xie, Dylan Losey, Ryan Tolsma, Chelsea Finn, and Dorsa Sadigh. Learning latent representations to influence multi-agent interaction. In Conference on robot learning, pages 575–588. PMLR, 2021. 2

    2021

  30. [38]

    Coordination with humans via strategy matching

    Michelle Zhao, Reid Simmons, and Henny Admoni. Coordination with humans via strategy matching. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 9116–9123. IEEE, 2022. 2

    2022

  31. [39]

    Maximum entropy population-based training for zero- shot human-ai coordination

    Rui Zhao, Jinming Song, Yufeng Yuan, Haifeng Hu, Yang Gao, Yi Wu, Zhongqian Sun, and Wei Yang. Maximum entropy population-based training for zero- shot human-ai coordination. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 37, pages 6145–6153, 2023. 2, 3

    2023

  32. [40]

    Deep interactive bayesian reinforcement learning via meta-learning

    Luisa Zintgraf, Sam Devlin, Kamil Ciosek, Shimon Whiteson, and Katja Hofmann. Deep interactive bayesian reinforcement learning via meta-learning. In Proceedings of the 20th International Conference on Autonomous Agents and MultiAgent Systems, pages 1712–1714, 2021. 2

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.