Pith. sign in

REVIEW 5 major objections 4 minor 29 references

Structured LLM Reasoning for Zero-Shot Human--Robot Coordination Under Hidden Goals

T0 review · 5 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a Dec-POMDP-structured LLM with theory-of-mind inference beats offline RL and no-ToM LLM baselines in zero-shot human-robot coordination.

desk verdict A coherent architecture and a promising pilot, but the ToM benefit is statistically unproven: n=5, no inferential tests, and an unexplained 0% LLM/RL simulation result. read the letter →

arxiv 2608.04309 v1 pith:VOCLZ5NN submitted 2026-08-05 cs.RO cs.SYeess.SY

Dong Hae Mangalindan , Anand Gokhale , Francesco Bullo , Vaibhav Srivastava This is my paper · ORCID
classification cs.ROcs.SYeess.SY
keywords human-robotcoordinationtheoryofmindlargelanguagemodelsDec-POMDPzero-shotplanningprivategoalstrustinHRILaBocaconstructiontask
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

The paper tries to establish that a robot can coordinate zero-shot with a human teammate when each has a private goal, if the robot's decisions are decomposed along the lines of a decentralized POMDP and LLMs handle the intractable inference and planning pieces. In a cooperative construction game inspired by La Boca, where human and robot each see only their own 2D projection of a 3D structure, the proposed architecture keeps a belief over the human's hidden target, plans hierarchically, interprets speech, verifies proposed actions with rules, and replans on failure. In a five-participant pilot, this LLM+ToM system required fewer interaction steps (mean 5.2) than an LLM ablation without ToM (6.4) and an offline-trained RL policy (10.2), and scored higher on post-interaction trust (84.9 vs 81.1 and 57.1). The import is that the bottleneck in human-robot coordination under hidden information may be decision-theoretic structure, and LLMs can serve as tractable surrogates for computations that are formally intractable.

What carries the argument

The central mechanism is the action-conditioned Theory-of-Mind inference loop: an LLM generates goal-conditioned action predictions, a second LLM scores their compatibility with observed human behavior, and these scores instantiate the human-response likelihood in the Bayes update of Eq. (1). This loop is embedded in a hierarchical planner that first selects a subgoal and then a candidate physical action, a rule-based verifier that rejects infeasible proposals, and a replanner that revises rejected actions using structured failure feedback. A conversation module extracts task-relevant intent from human speech and feeds it back into both inference and planning. The Dec-POMDP common-information view supplies the organizing principle: keep a belief over the human's private goal and act on an approximate version of that belief.

What would settle it

Run the same task with a held-out set of human goal views while recording the robot's goal belief after every human turn; if the probability assigned to the true goal does not rise on average across participants, or if an ablation with randomly shuffled ToM scores matches the full system on steps and trust, then the ToM module is not the active ingredient. A direct probe would compare the LLM's predicted human action distribution for a candidate goal against actual placements made by humans assigned that goal.

Watch

Extended reading notes

Core claim

Under the Dec-POMDP formulation, the robot maintains a belief $b^H_t(g)$ over the human's private target view, updated via Bayes' rule when a human action is observed. The LLM-based ToM module instantiates the needed human-response model: for each candidate goal $g$, an LLM generates plausible human actions and a second LLM compares those predictions to the observed action history, producing a compatibility score $\hat{q}^H_t(g)$ that plays the role of the likelihood in the belief update. The paper reports that, in human experiments, the average belief assigned to the true goal rises after each update, and that the full architecture completed tasks in fewer steps and earned higher trust than both the no-ToM LLM ablation and the offline RL policy. The claim is that explicit partner-goal inference, rather than merely generating physically feasible actions, is what makes coordination efficient and trustworthy.

Load-bearing premise

The load-bearing premise is that the LLM compatibility scores, used in place of the true human-response likelihood in the Bayes update, genuinely reflect how a human acts under each candidate goal; if they do not, the robot's belief about the hidden goal is noise and the ToM module provides no coordination benefit.

Editorial extensions

If this is right

  • If the architecture is right, zero-shot coordination with hidden goals does not require training over goal pairs; a structured LLM pipeline with explicit goal inference can match or beat a purpose-trained RL policy in task efficiency.
  • Explicit ToM inference is the source of the coordination gain: the LLM ablation without it took more steps (6.4 vs 5.2) and scored lower on trust (81.1 vs 84.9).
  • Human trust tracks interpretability, not just feasibility: the RL policy produced only feasible actions, yet four of five participants terminated their trials in frustration and trust fell below pre-experiment levels.
  • Separating semantic planning from rule-based verification lets approximate LLM-generated actions be corrected before execution, so the LLM does not need to know the exact admissible action set.
  • The belief update can move the posterior toward the true goal with only two observations in a short task, suggesting that even sparse evidence is useful for hidden-goal inference.

Reading between the lines

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

  • A testable extension would be to calibrate the LLM compatibility scores against measured human behavior under each candidate goal; the paper does not validate the belief against independent observations, and with only two belief updates per task the posterior could be tracking noise.
  • The trust gap between the LLM conditions and RL may stem partly from the LLM's conversational ability rather than goal inference alone, so a follow-up could ablate speech while keeping ToM to isolate the mechanism.
  • Because the RL baseline was trained in self-play, a fairer comparison might train it with human demonstrations or a longer curriculum; the 100% simulated RL/RL success suggests the policy fails under distribution shift to human partners, not that RL is inherently unsuitable.
  • If the belief-update effect is real, then a confidence-triggered clarification mechanism, where the robot asks only when the top two goal hypotheses are close, could cut steps further; the paper lists this as future work but does not test it.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper presents a structured large-language-model (LLM) architecture for zero-shot human-robot coordination in a cooperative construction task in which the human and robot hold private target views. The architecture decomposes decision-making into action-conditioned Theory-of-Mind (ToM) inference, hierarchical planning, conversation interpretation, rule-based action verification, and feedback-based replanning, with the Dec-POMDP formulation as the motivating framework. The authors report a human-participant pilot study (n=5) comparing the full architecture (LLM+ToM), an ablation without ToM (LLM), and an offline multi-agent RL policy, finding that LLM+ToM required fewer interaction steps and yielded higher post-interaction trust scores than both baselines. The paper also reports simulation results for RL/RL, LLM/RL, and LLM/LLM configurations and concludes that structured LLM decomposition can improve both task coordination and human experience.

Significance. If the empirical claims were statistically validated, the paper would make a useful contribution: it offers a concrete instantiation of a Dec-POMDP-guided modular LLM architecture for human-robot coordination, with a clear separation between approximate semantic reasoning and exact feasibility verification. The architecture is well motivated, the module decomposition is sensible, and the use of a rule-based verifier to constrain LLM-generated actions is a practical and defensible design choice. The paper is also honest in calling the human study a pilot. However, the central claim that the proposed method improves coordination efficiency and trust over the ablation rests on a very small sample without inferential statistics, and the simulation results contain an unexplained failure mode that currently undermines, rather than supports, the general claim. The architectural contribution is interesting, but the evidence base is not yet sufficient for the abstract's strong conclusions.

major comments (5)
  1. [Section IV.B, Fig. 7] The central comparison between LLM+ToM and the LLM ablation is reported as 5.2 vs 6.4 interaction steps and 84.9 vs 81.1 trust points, but these means are based on n=5 with no per-condition variance, no confidence intervals, and no significance tests. The claim that explicit ToM inference improves coordination efficiency and trust is therefore not statistically established. Please report participant-level data, paired effect sizes, and appropriate nonparametric or bootstrap inference (e.g., Wilcoxon signed-rank test or permutation test on the paired differences).
  2. [Section IV.B, RL condition] The RL baseline is confounded in two ways. First, four of five participants terminated the RL condition before completing the task, so the reported mean of 10.2 interaction steps mixes one completed trajectory with four partial ones and cannot be interpreted as a completion-time comparison. Second, the RL policy has no conversation channel, whereas the LLM conditions do, so the lower trust and higher steps in the RL condition may be driven by the absence of communication rather than by the policy quality. To support the comparison, either provide a matched RL baseline with communication or restrict the claims to the LLM conditions only.
  3. [Section IV.C, Table I] The simulation results show that LLM/RL achieves 0% success while RL/RL achieves 100%, and LLM/LLM only 60%—yet the human experiments show the LLM architecture outperforming RL. This discrepancy is not explained. If LLM-based planners are ineffective when paired with RL partners, the claim that the structured LLM architecture provides a generally effective surrogate for multi-agent coordination is called into question. Please provide an analysis of why LLM/RL fails (e.g., action distribution, equilibrium mismatch, or incompatibility of communication) or remove the simulation claim from the paper's support.
  4. [Section III.B, Eq. (1)] The Bayesian update in Eq. (1) requires the likelihood q_H^t(e_H^t | g, C_t, g_R) of the current evidence e_H^t, but the LLM_compare module computes compatibility scores over the entire observed human-action history h_H^t and these scores are then used directly in place of the likelihood. This is not an exact implementation of the stated Bayesian update, and the approximation is not analyzed. Moreover, the ToM belief shown in Fig. 8 is the module's own output; the paper does not validate the belief against any independent measurement of the human's behavior or goal. Please clarify the approximation and provide external validation, such as prediction of held-out human actions or comparison with an independent behavioral model.
  5. [Section IV.B, Fig. 8] The direct evidence for ToM inference consists of only three averaged belief states (initial plus two updates), with no participant-level trajectories, no uncertainty estimates, and no quantitative criterion for 'progressive inference' (e.g., rank of the true goal or calibration of the probabilities). Since the ToM module runs once every two steps and tasks average 5.2 steps, the evidence is minimal. Please provide per-participant belief trajectories with error bars and a formal evaluation metric.
minor comments (4)
  1. [Section IV.B, Fig. 8 caption] The text says the ToM module was executed 'every two trials' and that participants completed the task in '5.2 trials,' but the task measure is logged interaction steps, not trials. Please replace 'trials' with 'steps' or 'interaction steps' for consistency with Section IV.A and Fig. 7.
  2. [Section III.B, last paragraph] There is a sentence fragment and punctuation error: 'outputs a belief distribution representing the probability that the human is pursuing each possible goal. which is supplied to both the planner and conversation modules.' Please revise to '... each possible goal; this belief is supplied to both the planner and conversation modules.'
  3. [Section IV.A, Goal views] Fig. 6 shows three human goal views plus an unassigned fourth view, and the results state that the initial belief is uniform over all four. Please clarify how the set of candidate goal views G_H(g_R) is generated and why the unassigned view is included as a candidate when it can never be the true goal.
  4. [Abstract and Section V] The conclusion states that the architecture 'can improve both task coordination and the human experience,' but the supporting evidence is a five-participant pilot with no significance tests. Please temper the wording to 'may improve' or 'shows preliminary evidence for improving' to match the reported results.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central efficiency and trust claims rest on external human-participant measurements, not on the architecture's own outputs.

full rationale

The paper's central claims—fewer interaction steps and higher post-interaction trust under LLM+ToM—are measured externally in a human-participant study (Section IV.B) against a ToM-free LLM ablation and an offline RL baseline, so they do not reduce to the architecture's inputs. The belief update in Eq. (1) is a stated Bayesian update; Section III.B replaces the true human-response likelihood with LLM-generated compatibility scores, but this is an explicitly approximate surrogate, not a parameter fitted to the outcome it is later used to predict. No parameter is fitted to the experimental results, and the LLM modules are zero-shot. The only self-citation, reference [16] (LogicGuard, with overlapping authors), appears in a related-work enumeration and is not load-bearing. Figure 8 plots the ToM module's own posterior probability assigned to the true goal; this is an internal consistency check rather than independent validation, but it is not circular because the belief could in principle have failed to track the true goal. Statistical weaknesses (n=5, no inferential statistics, RL terminations) are validity threats, not circularity. Therefore no load-bearing step reduces to its own inputs.

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

The paper has no numeric free parameters fitted to data; its central claims rest on modeling assumptions about human behavior and on the validity of LLM-generated belief scores, which are not independently benchmarked.

assumptions (4)
  • domain assumption The human response model q_H^t(e|g,C_t,g_R) can be approximated by LLM compatibility scores.
    The belief update in Eq. (1) requires a goal-conditioned model of human behavior; the paper substitutes LLM-generated predictions and comparisons (Section III.B), but provides no validation that these scores behave like likelihoods, for example calibration or discriminability.
  • domain assumption The candidate goal set G_H(g_R) is complete and exhaustive; the human's private view is always one of the four views in the experiment.
    The ToM module iterates only over G_H(g_R); if the true goal is outside this set, the belief is undefined (Section II.B).
  • domain assumption Physical feasibility is fully captured by the rule-based verifier V(x_t, a~_t) in Section III.A.
    The verifier checks piece availability, boundaries, overlap, support, and operation rules; the paper assumes this deterministic function exactly matches the admissible set A_feas(x_t).
  • domain assumption Participants act approximately rationally and their actions convey information about their private goal.
    ToM inference relies on observed human actions being evidence about g_H; the paper does not measure human policy deviations or irrationality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Structured LLM Reasoning for Zero-Shot Human--Robot Coordination Under Hidden Goals." pith.science (2026). https://pith.science/paper/VOCLZ5NN

@misc{pith2026260804309,
  author       = {Pith},
  title        = {Pith review of: Structured LLM Reasoning for Zero-Shot Human--Robot Coordination Under Hidden Goals},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VOCLZ5NN}},
  note         = {Machine review of arXiv:2608.04309}
}
read the original abstract

We present a structured large-language-model (LLM) architecture for zero-shot human--robot coordination in a cooperative construction task with private goal views. Guided by a Dec-POMDP formulation, the architecture decomposes decision-making into (i) action-conditioned Theory-of-Mind (ToM) inference, (ii) hierarchical planning, (iii) conversation interpretation, (iv) action verification, and (v) feedback-based replanning. We compare the proposed method with an ablation without ToM inference and a multi-agent reinforcement-learning policy trained offline over many goal pairs. In human-participant experiments, the proposed method required fewer interaction steps and yielded higher post-interaction trust ratings than both baselines. These results suggest that systematically decomposing the team decision problem, using LLMs as tractable surrogates for otherwise intractable inference and planning computations, and retaining conventional verification for physical feasibility can improve both task coordination and the human experience.

Figures

Figures reproduced from arXiv: 2608.04309 by the authors.

Figure 1
Figure 1. The nine construction pieces used in the experiment. The pieces differ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Structured LLM architecture for human–robot coordination under [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Hierarchical planning with verified replanning. The LLM first selects [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Action-conditioned Theory-of-Mind inference. For each candidate [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: Goal-view configurations used in the task: (a) is the target goal assigned [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 8
Figure 8. Figure 8: Evolution of belief in the target goal view. The belief assigned to the [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 7
Figure 7. Figure 7: Pilot human-participant results. a Logged interaction steps under [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 20 canonical work pages

  1. [1]

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

    M. Ahn, A. Brohan, N. Brown, Y . Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gopalakrishnan, K. Hausman, A. Herzog, D. Ho, J. Hsu, J. Ibarz, B. Ichter, A. Irpan, E. Jang, R. J. Ruano, K. Jeffrey, S. Jesmonth, N. J. Joshi, R. Julian, D. Kalashnikov, Y . Kuang, K.-H. Lee, S. Levine, Y . Lu, L. Luu, C. Parada, P. Pastor, J. Quiambao, K. Rao, J. Retting...

    work page 2023

  2. [2]

    Inner monologue: Embod- ied reasoning through planning with language models,

    W. Huang, F. Xia, T. Xiao, H. Chan, J. Liang, P. Florence, A. Zeng, J. Tompson, I. Mordatch, Y . Chebotaret al., “Inner monologue: Embod- ied reasoning through planning with language models,”arXiv preprint arXiv:2207.05608, 2022

    arXiv 2022

  3. [3]

    Code as policies: Language model programs for embodied control,

    J. Liang, W. Huang, F. Xia, P. Xu, K. Hausman, B. Ichter, P. Florence, and A. Zeng, “Code as policies: Language model programs for embodied control,” inIEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 9493–9500

    work page 2023

  4. [4]

    Progprompt: Generating situated robot task plans using large language models,

    I. Singh, V . Blukis, A. Mousavian, A. Goyal, D. Xu, J. Tremblay, D. Fox, J. Thomason, and A. Garg, “Progprompt: Generating situated robot task plans using large language models,”arXiv preprint arXiv:2209.11302, 2022

    arXiv 2022

  5. [5]

    Language models as zero-shot planners: Extracting actionable knowledge for embodied agents,

    W. Huang, P. Abbeel, D. Pathak, and I. Mordatch, “Language models as zero-shot planners: Extracting actionable knowledge for embodied agents,” inInternational Conference on Machine Learning. PMLR, 2022, pp. 9118–9147

    2022

  6. [6]

    VOY AGER: An open-ended embodied agent with large language models,

    G. Wang, Y . Xie, Y . Jiang, A. Mandlekar, C. Xiao, Y . Zhu, L. Fan, and A. Anandkumar, “VOY AGER: An open-ended embodied agent with large language models,”arXiv preprint arXiv:2305.16291, 2023

    arXiv 2023

  7. [7]

    Cognitive architectures for language agents,

    T. R. Sumers, S. Yao, K. Narasimhan, and T. L. Griffiths, “Cognitive architectures for language agents,”arXiv preprint arXiv:2309.02427, 2023

    arXiv 2023

  8. [8]

    Adapt: As-needed decomposition and planning with language models,

    A. Prasad, A. Koller, M. Hartmann, P. Clark, A. Sabharwal, M. Bansal, and T. Khot, “Adapt: As-needed decomposition and planning with language models,” inFindings of the Association for Computational Linguistics: NAACL 2024, 2024, pp. 4226–4252

    work page 2024

Show all 29 references
  1. [9]

    KGLAMP: knowledge graph-guided language model for adaptive multi-robot plan- ning and replanning,

    C. L. Shek, F. M. Tariq, S. Bae, D. Isele, and P. Gupta, “KGLAMP: knowledge graph-guided language model for adaptive multi-robot plan- ning and replanning,”arXiv preprint arXiv:2602.04129, 2026

    2026 arXiv

  2. [10]

    Language-driven coordination and learning in multi-agent simulation environments,

    Z. Li, S. Campos, and N. Wang, “Language-driven coordination and learning in multi-agent simulation environments,”arXiv preprint arXiv:2506.04251, 2025

    2025

  3. [11]

    Collaborating action by action: A multi-agent LLMframework for embodied reasoning,

    I. White, K. Nottingham, A. Maniar, M. Robinson, H. Lillemark, M. Maheshwari, L. Qin, and P. Ammanabrolu, “Collaborating action by action: A multi-agent LLMframework for embodied reasoning,”arXiv preprint arXiv:2504.17950, 2025

    2025 arXiv

  4. [12]

    Language grounded multi-agent rein- forcement learning with human-interpretable communication,

    H. Li, H. N. Mahjoub, B. Chalaki, V . Tadiparthi, K. Lee, E. Moradi- Pari, M. Lewis, and K. Sycara, “Language grounded multi-agent rein- forcement learning with human-interpretable communication,”Advances in Neural Information Processing Systems, vol. 37, pp. 87 908–87 933, 2024

    2024

  5. [13]

    LAMI: Large language models for multi-modal human-robot interaction,

    C. Wang, S. Hasler, D. Tanneberg, F. Ocker, F. Joublin, A. Ceravola, J. Deigmoeller, and M. Gienger, “LAMI: Large language models for multi-modal human-robot interaction,” inExtended Abstracts of the CHI Conference on Human Factors in Computing Systems, 2024, pp. 1–10

    2024

  6. [14]

    Safe LLM- controlled robots with formal guarantees via reachability analysis,

    A. Hafez, A. N. Akhormeh, A. Hegazy, and A. Alanwar, “Safe LLM- controlled robots with formal guarantees via reachability analysis,”arXiv preprint arXiv:2503.03911, 2025

    2025 arXiv

  7. [15]

    Safety guardrails for llm-enabled robots,

    Z. Ravichandran, A. Robey, V . Kumar, G. J. Pappas, and H. Hassani, “Safety guardrails for llm-enabled robots,”IEEE Robotics and Automa- tion Letters, 2026

    2026

  8. [16]

    LogicGuard: Improving embodied LLM agents through temporal logic based critics,

    A. Gokhale, V . Srivastava, and F. Bullo, “LogicGuard: Improving embodied LLM agents through temporal logic based critics,”arXiv preprint arXiv:2507.03293, 2025

    2025

  9. [17]

    Action understanding as inverse planning,

    C. L. Baker, R. Saxe, and J. B. Tenenbaum, “Action understanding as inverse planning,”Cognition, vol. 113, no. 3, pp. 329–349, 2009

    2009

  10. [18]

    A framework for sequential planning in multi-agent settings,

    P. J. Gmytrasiewicz and P. Doshi, “A framework for sequential planning in multi-agent settings,”Journal of Artificial Intelligence Research, vol. 24, pp. 49–79, 2005

    2005

  11. [19]

    Autonomous agents modelling other agents: A comprehensive survey and open problems,

    S. V . Albrecht and P. Stone, “Autonomous agents modelling other agents: A comprehensive survey and open problems,”Artificial Intelligence, vol. 258, pp. 66–95, 2018

    2018

  12. [20]

    Legibility and predictabil- ity of robot motion,

    A. D. Dragan, K. C. Lee, and S. S. Srinivasa, “Legibility and predictabil- ity of robot motion,” inACM/IEEE International Conference on Human- Robot Interaction (HRI), 2013, pp. 301–308

    2013

  13. [21]

    The complexity of decentralized control of Markov decision processes,

    D. S. Bernstein, R. Givan, N. Immerman, and S. Zilberstein, “The complexity of decentralized control of Markov decision processes,” Mathematics of Operations Research, vol. 27, no. 4, pp. 819–840, 2002

    2002

  14. [22]

    F. A. Oliehoek, C. Amatoet al.,A Concise Introduction to Decentralized POMDPs. Springer, 2016, vol. 1

    2016

  15. [23]

    Multi-agent actor-critic for mixed cooperative-competitive environ- ments,

    R. Lowe, Y . I. Wu, A. Tamar, J. Harb, O. Pieter Abbeel, and I. Mordatch, “Multi-agent actor-critic for mixed cooperative-competitive environ- ments,”Advances in Neural Information Processing Systems, vol. 30, 2017

    2017

  16. [24]

    QMIX: Monotonic value function factorisation for deep multi-agent reinforcement learning,

    T. Rashid, M. Samvelyan, C. Schroeder, G. Farquhar, J. Foerster, and S. Whiteson, “QMIX: Monotonic value function factorisation for deep multi-agent reinforcement learning,” inInternational Conference on Machine Learning, 2018, pp. 4295–4304

    2018

  17. [25]

    The surprising effectiveness of PPO in cooperative multi-agent games,

    C. Yu, A. Velu, E. Vinitsky, J. Gao, Y . Wang, A. Bayen, and Y . Wu, “The surprising effectiveness of PPO in cooperative multi-agent games,”Ad- 8 vances in Neural Information Processing Systems, vol. 35, pp. 24 611– 24 624, 2022

    2022

  18. [26]

    Decentralized stochastic control with partial history sharing: A common information approach,

    A. Nayyar, A. Mahajan, and D. Teneketzis, “Decentralized stochastic control with partial history sharing: A common information approach,” IEEE Transactions on Automatic Control, vol. 58, no. 7, pp. 1644–1658, 2013

    2013

  19. [27]

    A general reinforcement learning algorithm that masters chess, shogi, and go through self-play,

    D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre, D. Kumaran, T. Graepelet al., “A general reinforcement learning algorithm that masters chess, shogi, and go through self-play,”Science, vol. 362, no. 6419, pp. 1140–1144, 2018

    2018

  20. [28]

    Fictitious self-play in extensive- form games,

    J. Heinrich, M. Lanctot, and D. Silver, “Fictitious self-play in extensive- form games,” inInternational Conference on Machine Learning. PMLR, 2015, pp. 805–813

    2015

  21. [29]

    Measuring trust in human robot interactions: Develop- ment of the “trust perception scale-HRI

    K. E. Schaefer, “Measuring trust in human robot interactions: Develop- ment of the “trust perception scale-HRI”,” inRobust Intelligence and Trust in Autonomous Systems. Springer, 2016, pp. 191–218

    2016

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.