Pith. sign in

REVIEW 4 major objections 5 minor 52 references

Co-Creative Learning via Metropolis-Hastings Interaction between Humans and AI

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Accepting and rejecting names like a Bayesian sampler lets a human and an AI build a shared category system.

desk verdict Useful empirical extension of MHNG to human-AI dyads, but the mechanism claim rests on an in-sample fit; the behavioral contrast between conditions is the real contribution. read the letter →

arxiv 2506.15468 v1 pith:XYJKK566 submitted 2025-06-18 cs.HC cs.AIcs.LG

Ryota Okumura , Tadahiro Taniguchi , Akira Taniguchi , Yoshinobu Hagiwara This is my paper · ORCID
classification cs.HCcs.AIcs.LG
keywords co-creativelearninghuman-AIinteractionsymbolemergenceMetropolis-HastingsnaminggamejointattentiondecentralizedBayesianinferenceAIalignmenthuman-in-the-loopmachine
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 argues that human-AI collaboration can work as co-creative learning: two agents, each seeing only part of the same objects, integrate what they have by exchanging names and accepting or rejecting proposals according to the Metropolis-Hastings rule. The claim is that this simple game makes the pair behave like one decentralized Bayesian system, converging on categories that reflect both partners' partial views. An online experiment with 69 participants paired people with an MH-rule computer agent, an always-accept agent, or an always-reject agent; the MH agent ended with the highest categorization accuracy and closest convergence to a shared sign system, and participants' accept decisions tracked the theoretical MH acceptance probability. If correct, this gives a concrete, testable mechanism for AI that learns with people instead of merely being taught by them, relevant to building shared representations and aligning AI to human perception.

What carries the argument

The load-bearing mechanism is the Metropolis-Hastings naming game (MHNG), a protocol in which a listener accepts a speaker's proposed name with probability $r_{\mathrm{MH}} = \min(1, P(c_{\mathrm{Li}}|\theta_{\mathrm{Li}}, s^*)/P(c_{\mathrm{Li}}|\theta_{\mathrm{Li}}, s_{\mathrm{Li}}))$, the likelihood ratio of the listener's own inferred category under the proposed sign versus its current sign. The game converts a conversation into a distributed Metropolis-Hastings sampler with the joint posterior as its target, so local accept/reject choices implement collective Bayesian inference without any agent revealing private observations or gradients. To measure this in humans, the paper analyzes interaction logs with the Inter-GM generative model, in which each agent's observations are Gaussian mixtures over latent categories linked by shared signs, infers each participant's parameters and category assignments by Gibbs sampling, and uses the fitted model both to compute the acceptance probability and to construct the target sign posterior. The three experimental conditions instantiate three learning paradigms: always-accept as supervised learning, always-reject as unsupervised learning, and the MH rule as co-creative learning.

What would settle it

Run the same partial-observation naming game but ask participants to categorize held-out novel stimuli or rate pairwise similarities independently before, during, and after the interaction; if their accept decisions track the MH acceptance probability computed from those independently elicited beliefs as well as they track the probability computed from the game-fitted model, the finding is genuine, while a mismatch would show the reported alignment is specific to the fitted model rather than to human psychology.

Watch

Extended reading notes

Core claim

The central discovery the paper reports is that a human and a computer agent playing a joint attention naming game under partial observability achieve co-creative learning: the dyad's accepted names follow a Metropolis-Hastings update rule, so the interaction is a distributed Bayesian sampler whose stationary distribution is the posterior over shared signs conditioned on the union of both agents' observations. In the experiment, the computer agent using the MH rule reached final categorization ARI of 0.609, significantly above 0.469 for the always-accept condition and 0.404 for the always-reject condition, and its final sign distribution agreed with the combined-observation posterior at 0.765, against 0.717 and 0.469 respectively. Human accept/reject behavior was also well described by the theoretical acceptance probability, with a fitted sensitivity parameter of 0.645. The paper interprets these results as the first empirical evidence that human-AI interaction of this kind performs decentralized Bayesian inference through the MHNG mechanism.

Load-bearing premise

The load-bearing premise is that the statistical category model fitted to each participant's interaction log faithfully reproduces how that person actually categorized the stimuli, since both the acceptance probability and the target sign distribution come from that same fitted model.

Editorial extensions

If this is right

  • Dyads using the MH rule converge to a shared sign system that encodes information from both sensors, so the computer's categories can improve beyond what the human's labels alone or its own observations alone would support.
  • The acceptance rule is not a neutral interface: an AI that accepts or rejects proposals probabilistically shapes human categorization as well, making the interaction a two-way influence rather than one-way teaching.
  • Because the mechanism never requires sharing raw observations or model gradients, co-creative learning of this kind could operate in privacy-sensitive settings where centralized training is impossible.
  • Human acceptance behavior being predictable from the MH probability suggests that the same algorithm could be used to design AI agents that communicate in ways people find natural to accept or reject.

Reading between the lines

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

  • The argument should extend from dyads to larger teams: because the MHNG update uses only local messages, a network of humans and AIs with different sensors could in principle converge to one shared naming scheme without any central coordinator. The paper does not test this.
  • A sharper, independently grounded test would elicit participants' category beliefs outside the naming game, for example by asking them to sort held-out novel stimuli before and after the session, and then check whether the MH acceptance curve predicts those independent judgments; the current experiment derives the target posterior from the same model family it uses to score participants.
  • One can derive a quantitative prediction for the acceptance curve: shifting category overlap or prior uncertainty in the stimulus generation should shift the fitted slope and baseline in a specific direction, which would distinguish genuine Bayesian updating from a generic tendency to accept plausible names.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes 'co-creative learning' as a paradigm in which a human and an AI integrate partial observations through decentralized Bayesian inference, instantiated by the Metropolis-Hastings naming game (MHNG) with an Interpersonal Gaussian Mixture (Inter-GM) model. The authors formalize this as minimization of collective free energy and report an online joint-attention naming game experiment (N=69 after exclusions) with three computer-agent conditions (MH-based, always-accept, always-reject). They report that the MH condition yields significantly higher computer categorization accuracy (ARI) and sign-posterior agreement than controls, and that human accept/reject decisions are positively correlated with the model-derived MH acceptance probability r_MH. The paper interprets this as first empirical evidence for co-creative learning in human-AI dyads.

Significance. If the empirical claims held, the paper would be a meaningful step: it extends experimental semiotics and MHNG from human-human dyads to human-AI dyads, connects the mechanism to collective predictive coding and bidirectional alignment, and releases its data. The theoretical formalization of co-creative learning via collective free energy is useful, and the three-condition design is a sensible way to contrast co-creative, supervised-like, and unsupervised-like interaction. However, the central mechanism evidence is in-sample, the sign-agreement target is model-derived from the same data, and the human-side advantages over the always-accept control are not statistically significant; these issues currently limit the strength of the headline claim.

major comments (4)
  1. [§3.2, §4, Eq. (2), Fig. 4] The human-acceptance alignment is computed in-sample: r_MH uses Inter-GM parameters and category assignments inferred by Gibbs sampling from the same interaction stream on which the human accept/reject decisions were recorded. A flexible model fit to the data will assign high r_MH to proposals consistent with the participant's learned label-category mapping, exactly the proposals the participant tends to accept, so the positive fitted slope a=0.645 does not independently verify that the human is sampling according to the MH target. Please add an out-of-sample evaluation, for example fitting the Inter-GM on the initial categorization or the first half of each participant's interactions and predicting accept/reject decisions in the held-out half, and compare prediction accuracy against simple baselines (e.g., always accept, accept-when-label-matches-current-category). Absent such a test, the claim that the human half of the dyad performs Bayesian inference is not supported.
  2. [§4, Evaluation Metrics, Table 1] The sign-posterior agreement target is obtained by Gibbs sampling the full Inter-GM model on the same dyad's observations and sign sequences, and the empirical sign distributions being compared are generated by the same model family. High agreement can therefore partly reflect the fitted model's ability to reconstruct the data rather than convergence to the true integrated posterior. Please validate the target using held-out data or an independent target (for instance, posterior over ground-truth categories or a model fit on one half of the interactions), and report agreement against human-only and AI-only targets as a reference.
  3. [§4, Participants] The analysis excludes 21 of 90 participants based on 'inactivity or failure to follow instructions,' but the exclusion criteria are not pre-registered or quantified (e.g., no threshold for number of identical responses), and no information is given about excluded participants by condition. Because the MH-vs-AA computer ARI difference is marginal (p=0.046), the results may be sensitive to these exclusions. Please specify the exact exclusion rules, report counts per condition, and provide a sensitivity analysis including all participants.
  4. [§5, Table 1] The human-side outcomes do not show a significant MH advantage: human ARI in MH (0.490) is not significantly different from AA (0.473; p=0.799), and human sign agreement in MH (0.729) is not significantly different from AA (0.722; p=0.762). The evidence for 'mutual' integration is therefore confined to the computer agent's metrics. Please either provide analyses demonstrating a human-side benefit (e.g., individual learning curves, within-dyad alignment over time) or temper the claim that the dyad co-creatively learns beyond what the always-accept baseline achieves.
minor comments (5)
  1. [Fig. 4] The figure contains Japanese text in its axis labels and title; these should be translated to English for the intended readership.
  2. [Table 1] The column layout of Table 1 is difficult to parse, especially the placement of Initial, Final, and the agreement columns; please reformat so that each phase and metric is clearly labeled.
  3. [§5] For the Welch t-tests, please report effect sizes and confidence intervals in addition to p-values, particularly for the marginally significant MH-vs-AA comparison.
  4. [§2.2] The statement that the collective free energy decrease follows from detailed balance via the Data Processing Inequality is loose; please cite the standard Metropolis-Hastings convergence argument or prove the KL decrease directly.
  5. [Data Availability] The data link is a Google Drive folder; please provide a stable archival identifier (e.g., a DOI) and include the analysis code to make the in-sample/out-of-sample distinction reproducible.

Circularity Check

2 steps flagged · score 6.0 of 10

Human-acceptance alignment is an in-sample fit to model parameters inferred from the same interaction stream, and the sign-agreement target is the sampler's own stationary distribution, so the MHNG-mechanism claim is only partially supported.

  1. fitted input called prediction [Section 3.2 and Section 4 (Modeling Human Acceptance Behavior); results in Section 5]
    "The MH acceptance probability r_MH_n (Eq. 2) is calculated using these inferred parameters. ... For each instance where the human acted as listener, r_MH_n was calculated using the human's inferred Inter-GM parameters (ΘA) and category assignment (cA_n) via Gibbs sampling. The parameters a (sensitivity to r_MH_n) and b (baseline acceptance) were estimated via MLE across all relevant interactions for all participants using gradient descent."

    The alignment claim is the only direct evidence that the human half of the dyad is performing MH-style Bayesian inference. However, r_MH is not a fixed theoretical prediction: it is computed from Inter-GM parameters and category assignments inferred by Gibbs sampling from the very interaction stream that the human's accept/reject decisions helped produce (the accepted sign sequence conditions the inference). The linear sensitivity a and baseline b are then fitted by MLE to those same accept/reject outcomes. The reported positive a therefore measures how well the post-hoc fitted model reconstructs the participant's own labeling behavior; it is an in-sample correlation, not an out-of-sample verification that humans sample from the MH target.

  2. self definitional [Section 3.1 and Section 4 (Evaluation Metrics)]
    "The MHNG process functions as a distributed implementation of an MCMC algorithm ... targeting the posterior distribution over the shared signs sn conditioned on the distributed observations xA_n and xB_n. ... the target distribution was approximated by the sign posterior derived from Gibbs sampling on the full Inter-GM model given both human and computer observations {xA_n, xB_n}."

    The metric used to claim 'stronger convergence toward a shared sign system' compares each agent's sign distribution to the sign posterior of the full Inter-GM model. The MHNG is itself defined as a sampler that targets exactly this posterior given the distributed observations. Consequently, the significant improvement in the computer agent's agreement in the MH condition is the algorithm converging to its own stationary distribution; it is a self-consistency property of the designed sampler rather than an independent empirical demonstration of human-AI co-creative learning. Human agreement differences were not significant (MH vs AA, p=0.762), so this self-referential metric carries much of the shared-sign-system conclusion.

full rationale

The paper's headline ARI results are grounded in ground-truth category labels and are not circular: the MH-agent's higher final computer ARI and the comparison against AA/AR are independent empirical outcomes. The theoretical MHNG mechanism is also externally established in prior computational work. However, two evaluation steps introduce partial circularity. First, the human-acceptance alignment—the key evidence that the human half is doing Bayesian inference—is computed from Inter-GM parameters inferred from the same interaction data that the accept/reject decisions shaped, and the sensitivity parameter a is fitted to those same decisions. This is an in-sample description rather than an independent prediction, so it cannot bear the weight of the mechanism claim. Second, the sign-posterior agreement target is the stationary distribution of the very MHNG sampler being tested, making the computer's convergence to it partly definitional. These issues do not destroy the empirical ARI contribution, but they mean the paper's central claim that human-AI dyads perform decentralized Bayesian inference through MHNG is only partially supported by the evidence as presented.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim relies on the MHNG convergence theorem (from prior literature) and on the assumption that the Inter-GM model accurately represents human internal states. The empirical evaluation adds two fitted parameters (a,b) for human acceptance and a hand-chosen stimulus distribution. No new physical entities are introduced.

free parameters (3)
  • a (sensitivity to r_MH) = 0.645 ± 0.300
    Fitted by MLE to human acceptance decisions in the constrained linear model P(accept|r_MH)=a*r_MH+b; used as evidence that humans align with MH probability.
  • b (baseline acceptance) = 0.201 ± 0.187
    Fitted alongside a; represents baseline acceptance rate independent of r_MH.
  • Ground truth Gaussian parameters (mu_k, Sigma) = not reported in main text
    Chosen by hand so that categories overlap under partial observation; these directly determine task difficulty and therefore the magnitude of ARI improvements.
assumptions (4)
  • standard math The Metropolis-Hastings transition kernel of MHNG satisfies detailed balance with respect to the joint posterior p(s | X_AI, X_Human, Theta).
    Stated in Section 2.2 and used to assert monotone decrease of collective free energy; established in Hagiwara et al. 2019 and Taniguchi et al. 2023b.
  • standard math Data Processing Inequality implies KL divergence decreases under the MH update.
    Invoked in Section 2.2 as the mechanism for monotone convergence; standard information theory.
  • domain assumption The Inter-GM generative model adequately captures human category learning and perception.
    Used to infer human internal parameters and to compute r_MH in Eq. (2); the validity of the alignment result depends on this model being a faithful description of human behavior.
  • domain assumption Participants maintain joint attention and follow the naming game protocol as instructed.
    The JA-NG paradigm assumes both agents attend to the same object; the authors list perfect joint attention as an assumption and limitation in Section 6.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Co-Creative Learning via Metropolis-Hastings Interaction between Humans and AI." pith.science (2026). https://pith.science/paper/XYJKK566

@misc{pith2026250615468,
  author       = {Pith},
  title        = {Pith review of: Co-Creative Learning via Metropolis-Hastings Interaction between Humans and AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XYJKK566}},
  note         = {Machine review of arXiv:2506.15468}
}
read the original abstract

We propose co-creative learning as a novel paradigm where humans and AI, i.e., biological and artificial agents, mutually integrate their partial perceptual information and knowledge to construct shared external representations, a process we interpret as symbol emergence. Unlike traditional AI teaching based on unilateral knowledge transfer, this addresses the challenge of integrating information from inherently different modalities. We empirically test this framework using a human-AI interaction model based on the Metropolis-Hastings naming game (MHNG), a decentralized Bayesian inference mechanism. In an online experiment, 69 participants played a joint attention naming game (JA-NG) with one of three computer agent types (MH-based, always-accept, or always-reject) under partial observability. Results show that human-AI pairs with an MH-based agent significantly improved categorization accuracy through interaction and achieved stronger convergence toward a shared sign system. Furthermore, human acceptance behavior aligned closely with the MH-derived acceptance probability. These findings provide the first empirical evidence for co-creative learning emerging in human-AI dyads via MHNG-based interaction. This suggests a promising path toward symbiotic AI systems that learn with humans, rather than from them, by dynamically aligning perceptual experiences, opening a new venue for symbiotic AI alignment.

Figures

Figures reproduced from arXiv: 2506.15468 by the authors.

Figure 1
Figure 1. A typical probabilistic graphical model (left) assumed in CPC and its decomposition for [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Experimental setup: (a) Stimuli under partial observation and (b) JA-NG procedure flow. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Mean ARI (± SE) over 200 steps for Human and Computer agents under MH, Always Accept (AA), and Always Reject (AR) conditions [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Screenshot of the experimental interface used in the JA-NG task. The graphical user [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 28 canonical work pages

  1. [1]

    S. Ali, N. Devasia, H. W. Park, and C. Breazeal. Social robots as creativity eliciting agents. Frontiers in Robotics and AI, 8, 2021. doi:10.3389/frobt.2021.674389

    work page arXiv 2021

  2. [2]

    Balzan, J

    F. Balzan, J. Campbell, K. Friston, M. J. Ramstead, D. Friedman, and A. Constant. Distributed science - the scientific process as multi-scale active inference, 2023. URL https://osf.io/preprints/dnw5k. preprint

    work page 2023

  3. [3]

    Brandizzi

    N. Brandizzi. Toward more human-like ai communication: A review of emergent communication research. IEEE Access, 11: 0 142317--142340, 2023

    work page 2023

  4. [4]

    Brinkmann, F

    L. Brinkmann, F. Baumann, J.-F. Bonnefon, M. Derex, T. F. M^^c3^^bcller, A.-M. Nussberger, A. Czaplicka, A. Acerbi, T. L. Griffiths, J. Henrich, J. Z. Leibo, R. McElreath, P.-Y. Oudeyer, J. Stray, and I. Rahwan. Machine culture. Nature Human Behaviour, 7 0 (11): 0 1855--1868, 2023. doi:10.1038/s41562-023-01653-4

    work page doi:10.1038/s41562-023-01653-4 2023

  5. [5]

    Cangelosi and M

    A. Cangelosi and M. Schlesinger. Developmental Robotics: From Babies to Robots. The MIT Press, Cambridge, MA, 2015. ISBN 9780262028668. doi:10.7551/mitpress/9780262028668.001.0001

    arXiv 2015

  6. [6]

    E. G. Carayannis, J. Morawska-Jancelewicz, and D. Meissner. Co-creation knowledge processes in a robotizing knowledge economy. Journal of the Knowledge Economy, 12: 0 240--263, 2021. doi:10.1007/s13132-020-00679-8

    work page doi:10.1007/s13132-020-00679-8 2021

  7. [7]

    Constant, M

    A. Constant, M. J. Ramstead, S. P. Veissiere, J. O. Campbell, and K. J. Friston. A variational approach to niche construction. Journal of the Royal Society Interface, 15 0 (141): 0 20170685, 2018

    work page 2018

  8. [8]

    Contucci, J

    P. Contucci, J. Kert \'e sz, and G. Osabutey. Human-ai ecosystem with abrupt changes as a function of the composition. PloS one, 17 0 (5): 0 e0267310, 2022

    work page 2022

Show all 52 references
  1. [9]

    Cornish, K

    H. Cornish, K. Smith, and S. Kirby. Systems of symbolic communication evolve in an iterated learning paradigm. Current Biology, 23 0 (8): 0 715--719, 2013

    2013

  2. [10]

    T. M. Cover and J. A. Thomas. Elements of Information Theory. Wiley-Interscience, 2nd edition, 2006

    2006

  3. [11]

    Cui and T

    H. Cui and T. Yasseri. Ai-enhanced collective intelligence. volume 5. Elsevier, 2024

    2024

  4. [12]

    J. N. Foerster, Y. M. Assael, N. de Freitas, and S. Whiteson. Learning to communicate with deep multi-agent reinforcement learning. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 29, pages ...

    2016

  5. [13]

    K. Friston. The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11 0 (2): 0 127--138, 2010. doi:10.1038/nrn2787

    2010 doi

  6. [14]

    Friston, R

    K. Friston, R. J. Moran, Y. Nagai, T. Taniguchi, H. Gomi, and J. Tenenbaum. World model learning and inference. Neural Networks, 144: 0 573--590, 2021. doi:10.1016/j.neunet.2021.09.011

    2021 doi

  7. [15]

    K. J. Friston, T. Parr, C. Heins, A. Constant, D. Friedman, T. Isomura, C. Fields, T. Verbelen, M. Ramstead, J. Clippinger, and C. D. Frith. Federated inference and belief sharing. Neuroscience & Biobehavioral Reviews, 156: 0 105500, 2024. ISSN 0149-7634. doi:10.1016/j.neubior...

    2024

  8. [16]

    Galantucci

    B. Galantucci. An experimental study of the emergence of human communication systems. Cognitive Science, 29 0 (5): 0 737--767, 2005. doi:10.1207/s15516709cog0000_26

    2005 doi

  9. [17]

    Galantucci

    B. Galantucci. Experimental semiotics: A new approach for studying communication as a form of joint action. Topics in Cognitive Science, 1 0 (2): 0 393--410, 2009. doi:10.1111/j.1756-8765.2009.01027.x

    2009

  10. [18]

    Galantucci and S

    B. Galantucci and S. Garrod. Experimental semiotics: A review. Frontiers in Human Neuroscience, 5, 2011. doi:10.3389/fnhum.2011.00011

    2011 arXiv

  11. [19]

    Geng and R

    M. Geng and R. Trotta. Is chatgpt transforming academics’ writing style? In Proceedings of the 41st International Conference on Machine Learning (ICML), Workshop on the Next Generation of AI Safety, Vienna, Austria, 2024. Workshop Paper

    2024

  12. [20]

    Ha and J

    D. Ha and J. Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018

    2018 arXiv

  13. [21]

    Hagiwara, H

    Y. Hagiwara, H. Kobayashi, A. Taniguchi, and T. Taniguchi. Symbol emergence as an interpersonal multimodal categorization. Frontiers in Robotics and AI, 6, 2019. doi:10.3389/frobt.2019.00134

    2019

  14. [22]

    J. Hohwy. The Predictive Mind. Oxford University Press, Oxford, 2013. ISBN 9780199682737. doi:10.1093/acprof:oso/9780199682737.001.0001

    2013

  15. [23]

    Hubert and P

    L. Hubert and P. Arabie. Comparing partitions. Journal of Classification, 2 0 (1): 0 193--218, 1985. doi:10.1007/BF01908075

    1985 doi

  16. [24]

    Inukai, T

    J. Inukai, T. Taniguchi, A. Taniguchi, and Y. Hagiwara. Recursive metropolis-hastings naming game: Symbol emergence in a multi-agent system based on probabilistic generative models. Frontiers in Artificial Intelligence, 6, 2023. doi:10.3389/frai.2023.1103663

    2023

  17. [25]

    D. P. Kingma and M. Welling. Auto-encoding variational bayes. In Proceedings of the 2nd International Conference on Learning Representations (ICLR), 2014

    2014

  18. [26]

    Kirby, M

    S. Kirby, M. Dowman, and S. Griffiths. The emergence of linguistic structure: An overview of the iterated learning model. Artificial Life, 9 0 (4): 0 371--386, 2002

    2002

  19. [27]

    Kirby, H

    S. Kirby, H. Cornish, and K. Smith. Cumulative cultural evolution in the laboratory: An experimental approach to the origins of structure in human language. Proceedings of the National Academy of Sciences of the United States of America, 105 0 (31): 0 10681--10686, 2008. doi:1...

    2008 doi

  20. [28]

    Kobak, R

    D. Kobak, R. Gonz^^c3^^a1lez-M^^c3^^a1rquez, E.-c. Horv^^c3^^a1t, and J. Lause. Delving into chatgpt usage in academic writing through excess vocabulary, 2025. URL https://arxiv.org/abs/2406.07016. Preprint

    2025 arXiv

  21. [29]

    Lazaridou and M

    A. Lazaridou and M. Baroni. Emergent multi-agent communication in the deep learning era. Transactions of the Association for Computational Linguistics, 10: 0 157--173, 2022. doi:10.1162/tacl_a_00457

    2022 doi

  22. [30]

    Lazaridou, A

    A. Lazaridou, A. Peysakhovich, and M. Baroni. Multi-agent cooperation and the emergence of (natural) language. In International Conference on Learning Representations (ICLR) Workshop Track, 2017. URL https://openreview.net/forum?id=Hk8N3Sclg. ICLR 2017 Workshop

    2017

  23. [31]

    Okumura, T

    R. Okumura, T. Taniguchi, Y. Hagiwara, and A. Taniguchi. Metropolis-hastings algorithm in joint-attention naming game: experimental semiotics study. Frontiers in Artificial Intelligence, 6, 2023. doi:10.3389/frai.2023.1235231

    2023

  24. [32]

    Pedreschi, L

    D. Pedreschi, L. Pappalardo, P. Ferragina, R. Baeza-Yates, A.-L. Barab^^c3^^a1si, F. Dignum, V. Dignum, T. Eliassi-Rad, F. Giannotti, J. Kert^^c3^^a9sz, A. Knott, Y. Ioannidis, P. Lukowicz, A. Passarella, A. S. Pentland, J. Shawe-Taylor, and A. Vespignani. Human-ai coevolution...

    2025

  25. [33]

    M. M. Peeters, J. Van Diggelen, K. Van Den Bosch, A. Bronkhorst, M. A. Neerincx, J. M. Schraagen, and S. Raaijmakers. Hybrid collective intelligence in a human--ai society. AI & society, 36: 0 217--238, 2021

    2021

  26. [34]

    Peters, C

    J. Peters, C. Waubert de Puiseau, H. Tercan, A. Gopikrishnan, G. A. Lucas de Carvalho, C. Bitter, and T. Meisen. Emergent language: a survey and taxonomy. Autonomous Agents and Multi-Agent Systems, 39 0 (1): 0 1--73, 2025

    2025

  27. [35]

    A. N. Sanborn and T. L. Griffiths. Markov chain monte carlo with people. In J. C. Platt, D. Koller, Y. Singer, and S. T. Roweis, editors, Advances in Neural Information Processing Systems, volume 20, pages 1249--1256, 2007

    2007

  28. [36]

    E. B. Sandoval, R. Sosa, M. Cappuccio, and T. Bednarz. Human-robot creative interactions (hrci): Exploring creativity in artificial agents using a story-telling game. arXiv preprint, 2022. URL https://arxiv.org/abs/2204.11598

    2022 arXiv

  29. [37]

    Schmidhuber

    J. Schmidhuber. Making the World Differentiable: On Using Self-Supervised Fully Recurrent Neural Networks for Dynamic Reinforcement Learning and Planning in Non-Stationary Environments. Inst. f\" u r Informatik, 1990

    1990

  30. [38]

    H. Shen, T. Knearem, R. Ghosh, K. Alkiek, K. Krishna, Y. Liu, Z. Ma, S. Petridis, Y.-H. Peng, L. Qiwei, et al. Towards bidirectional human-ai alignment: A systematic review for clarifications, framework, and future directions. arXiv preprint arXiv:2406.09264, 2024

    2024

  31. [39]

    M. Shin, J. Kim, B. van Opheusden, and T. L. Griffiths. Superhuman artificial intelligence can improve human decision-making by increasing novelty. Proceedings of the National Academy of Sciences, 120 0 (12): 0 e2214840120, 2023. doi:10.1073/pnas.2214840120

    2023 doi

  32. [40]

    L. Steels. The spontaneous self-organization of an adaptive language. Machine Intelligence, 15: 0 205--232, 1999

    1999

  33. [41]

    L. Steels. The Talking Heads Experiment: Origins of Words and Meanings. Language Science Press, Berlin, 2015

    2015

  34. [42]

    Steels, T

    L. Steels, T. Belpaeme, et al. Coordinating perceptually grounded categories through language: A case study for colour. Behavioral and Brain Sciences, 28 0 (4): 0 469--488, 2005. doi:10.1017/S0140525X0500009X

    2005 doi

  35. [43]

    Suzuki and Y

    M. Suzuki and Y. Matsuo. A survey of multimodal deep generative models. Advanced Robotics, 36 0 (5-6): 0 261--278, 2022

    2022

  36. [44]

    Taniguchi

    T. Taniguchi. Collective predictive coding hypothesis: Symbol emergence as decentralized bayesian inference. Frontiers in Robotics and AI, 11, 2024. doi:10.3389/frobt.2024.1353870

    2024

  37. [45]

    Taniguchi, T

    T. Taniguchi, T. Nagai, T. Nakamura, N. Iwahashi, T. Inamura, T. Ogata, and Y. Hagiwara. Symbol emergence in robotics: a survey. Advanced Robotics, 30 0 (11-12): 0 706--728, 2016. doi:10.1080/01691864.2016.1170443

    2016

  38. [46]

    Taniguchi, E

    T. Taniguchi, E. Ugur, M. Hoffmann, L. Jamone, T. Nagai, B. Rosman, T. Matsuka, N. Iwahashi, T. Inamura, K. Maruyama, and Y. Hagiwara. Symbol emergence in cognitive developmental systems: a survey. IEEE Transactions on Cognitive and Developmental Systems, 11 0 (4): 0 494--516,...

    2019

  39. [48]

    Taniguchi, Y

    T. Taniguchi, Y. Yoshida, Y. Matsui, N. L. Hoang, A. Taniguchi, and Y. Hagiwara. Emergent communication through metropolis-hastings naming game with deep generative models. Advanced Robotics, 37 0 (19): 0 1266--1282, 2023 b . doi:10.1080/01691864.2023.2241484

    2023

  40. [49]

    Taniguchi, S

    T. Taniguchi, S. Takagi, J. Otsuka, Y. Hayashi, and H. T. Hamada. Collective predictive coding as model of science: Formalizing scientific activities towards generative science. Royal Society Open Science, 12 0 (6): 0 241678, 2025

    2025

  41. [50]

    Ueda and T

    R. Ueda and T. Taniguchi. Lewis's signaling game as - VAE for natural word lengths and segments. In The Twelfth International Conference on Learning Representations (ICLR), 2024. URL https://openreview.net/forum?id=nO5VPdSo3y

    2024

  42. [51]

    E. M. van Zoelen, K. van den Bosch, and M. Neerincx. Becoming team members: Identifying interaction patterns of mutual adaptation for human-robot co-learning. Frontiers in Robotics and AI, 8, 2021. doi:10.3389/frobt.2021.675361

    2021

  43. [52]

    S. P. Veissiere, A. Constant, M. J. Ramstead, K. J. Friston, and L. J. Kirmayer. Thinking through other minds: A variational approach to cognition and culture. Behavioral and Brain Sciences, 43: 0 e90, 2020

    2020

  44. [53]

    Yakura, E

    H. Yakura, E. Lopez-Lopez, L. Brinkmann, I. Serna, P. Gupta, and I. Rahwan. Empirical evidence of large language model’s influence on human spoken communication, 2024. URL https://arxiv.org/abs/2409.01754. Preprint

    2024 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.