pith. sign in

arxiv: 2606.10582 · v1 · pith:5JKSPE7Xnew · submitted 2026-06-09 · 💻 cs.LG · cs.AI

Drawing with Strangers: Population Scaling Drives Zero-Shot Mutual Intelligibility in Emergent Sketching

Jooyeon Kim This is my paper

Pith reviewed 2026-06-27 14:13 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords emergent communicationzero-shot mutual intelligibilitysketchingpopulation scalingperceptual groundingmulti-agent systemsgeneralization
0
0 comments X

The pith

Scaling training populations in emergent sketching agents improves communication between independent groups without prior exposure.

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

The paper studies how agents can communicate with strangers from completely separate training groups, a skill it calls zero-shot mutual intelligibility. Using agents that draw strokes to point to images, it shows that training on bigger populations produces better cross-group success. Larger groups develop more varied sketching styles inside each population, which blocks them from locking into identical codes. At the same time, different groups grow more alike in their sketches. The convergence happens because bigger populations tie their drawings more closely to the actual visual appearance of the images.

Core claim

Scaling the training population substantially improves zero-shot mutual intelligibility across independent groups. As population size grows, in-group communicative variation increases, preventing co-adaptation into homogeneity, while cross-group variation decreases, indicating structural convergence toward universality. This universality is achieved through perceptual grounding, as scaled populations increasingly anchor their emergent sketches on the objective visual resemblance of the target images.

What carries the argument

Zero-shot mutual intelligibility (ZMI) between disjoint agent populations, achieved when population scaling drives convergence on perceptually grounded sketches rather than private conventions.

Load-bearing premise

Training occurs in strictly disjoint populations with no prior exposure, and the rise in in-group communicative variation with scale is the causal driver of reduced cross-group variation and improved zero-shot communication.

What would settle it

Running the same sketching experiments with larger populations but finding no increase in in-group variation or no gain in cross-group communication success would falsify the claim.

Figures

Figures reproduced from arXiv: 2606.10582 by Jooyeon Kim.

Figure 1
Figure 1. Figure 1: Zero-shot mutual intelligibility (ZMI) evaluated on two datasets. Each dot represents group-level [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Population scaling increases training cost approximately linearly. For each communication group [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Convergence of in-group and cross-group communicative variation across population scales. Sender [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Impact of centralized, one-to-many communication topologies on ZMI. ZMI is evaluated across [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect of population scaling on the perceptual grounding of emergent sketches. The visual similarity [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Relationship between perceptual grounding and ZMI [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Zero-shot mutual intelligibility (ZMI) evaluated on CelebA dataset. Each dot represents group [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: ZMI evaluated on the MNIST dataset under varying reference in-group validation communication [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: ZMI evaluated on the CIFAR-10 dataset under varying reference in-group validation communication [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: ZMI evaluated on the MNIST dataset when varying the number of strokes used to draw sketches [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: ZMI evaluated on the CIFAR-10 dataset when varying the number of strokes used to draw [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: ZMI evaluated on the MNIST and CIFAR-10 datasets after decreasing the number of candidate [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Model-scaling ablation on MNIST and CIFAR-10. In all settings, the communication population [PITH_FULL_IMAGE:figures/full_fig_p027_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Population scaling increases training cost approximately linearly, even for one-to-many commu [PITH_FULL_IMAGE:figures/full_fig_p028_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Examples of perceptually grounded sketches. [PITH_FULL_IMAGE:figures/full_fig_p034_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Example MNIST images [PITH_FULL_IMAGE:figures/full_fig_p035_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Create sketches from the MNIST data (Figure 16) by agent in group size [PITH_FULL_IMAGE:figures/full_fig_p035_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Create sketches from the MNIST data (Figure 16) by agent in group size [PITH_FULL_IMAGE:figures/full_fig_p035_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Create sketches from the MNIST data (Figure 16) by agent in group size [PITH_FULL_IMAGE:figures/full_fig_p036_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Create sketches from the MNIST data (Figure 16) by agent in group size [PITH_FULL_IMAGE:figures/full_fig_p036_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Create sketches from the MNIST data (Figure 16) by agent in group size [PITH_FULL_IMAGE:figures/full_fig_p036_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Create sketches from the MNIST data (Figure 16) by agent in group size [PITH_FULL_IMAGE:figures/full_fig_p037_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Create sketches from the MNIST data (Figure 16) by agent in group size [PITH_FULL_IMAGE:figures/full_fig_p037_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Create sketches from the MNIST data (Figure 16) by agent in group size [PITH_FULL_IMAGE:figures/full_fig_p037_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Create sketches from the MNIST data (Figure 16) by agent in group size [PITH_FULL_IMAGE:figures/full_fig_p038_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Example CIFAR10 images [PITH_FULL_IMAGE:figures/full_fig_p039_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: Create sketches from the CIFAR10 data (Figure 26) by agent in group size [PITH_FULL_IMAGE:figures/full_fig_p039_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: Create sketches from the CIFAR10 data (Figure 26) by agent in group size [PITH_FULL_IMAGE:figures/full_fig_p039_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: Create sketches from the CIFAR10 data (Figure 26) by agent in group size [PITH_FULL_IMAGE:figures/full_fig_p040_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: Create sketches from the CIFAR10 data (Figure 26) by agent in group size [PITH_FULL_IMAGE:figures/full_fig_p040_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: Create sketches from the CIFAR10 data (Figure 26) by agent in group size [PITH_FULL_IMAGE:figures/full_fig_p040_31.png] view at source ↗
Figure 32
Figure 32. Figure 32: Create sketches from the CIFAR10 data (Figure 26) by agent in group size [PITH_FULL_IMAGE:figures/full_fig_p041_32.png] view at source ↗
Figure 33
Figure 33. Figure 33: Create sketches from the CIFAR10 data (Figure 26) by agent in group size [PITH_FULL_IMAGE:figures/full_fig_p041_33.png] view at source ↗
read the original abstract

Generalization in emergent communication has largely focused on novel inputs or linguistic structures, yet the capacity for agents to communicate with strangers from strictly disjoint communities remains relatively unexplored. In this work, we formalize this capability as \textit{zero-shot mutual intelligibility (ZMI)}: successful communication between independently trained populations without prior exposure. Leveraging emergent sketching -- in which agents communicate through sets of drawn strokes -- as a visually grounded modality, we find that scaling the training population substantially improves ZMI across independent groups. Crucially, as we scale the population size, in-group communicative variation increases, preventing co-adaptation into homogeneity. Simultaneously, cross-group variation decreases, indicating a structural convergence toward a certain type of universality. Further analysis reveals that this universality is achieved through perceptual grounding: scaled populations increasingly anchor their emergent sketches on the objective visual resemblance of the target images. Together, these results position ZMI as a distinct axis of generalization in emergent communication and suggest a route toward socially interoperable artificial agents.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript formalizes zero-shot mutual intelligibility (ZMI) as successful communication between independently trained populations of sketching agents with no prior exposure. It reports that scaling the size of the training population improves ZMI, accompanied by increased in-group communicative variation (preventing homogeneity), decreased cross-group variation (indicating structural convergence), and a shift toward perceptual grounding on objective visual resemblance of target images.

Significance. If the empirical patterns and proposed mechanism hold after appropriate controls, the work would usefully extend emergent-communication research by identifying population scale as a driver of inter-group interoperability and by distinguishing ZMI from other generalization axes. The emphasis on perceptual grounding supplies a concrete, testable route toward socially interoperable agents.

major comments (2)
  1. [Abstract / §4] Abstract and §4 (results on variation): the manuscript states that population scaling increases in-group variation, which in turn prevents homogeneity and produces the observed drop in cross-group variation plus ZMI gains. No mediation analysis, ablation that holds scale fixed while varying communicative diversity, or controlled comparison isolating the variation driver is described; the causal sequence therefore remains an untested assumption rather than a demonstrated mechanism.
  2. [Abstract] Abstract: the claim that scaled populations 'increasingly anchor their emergent sketches on the objective visual resemblance' is presented as the explanation for universality, yet the text supplies no quantitative measure (e.g., correlation with image-feature similarity, human perceptual judgments, or ablation removing visual grounding) that would distinguish this account from alternative explanations such as richer gradients or implicit regularization.
minor comments (2)
  1. [Abstract] The abstract contains no numerical results, error bars, or statistical tests; the full manuscript should include these in the main text or a dedicated results table so readers can evaluate effect sizes.
  2. [§3 / §4] Notation for 'in-group communicative variation' and 'cross-group variation' should be defined explicitly (e.g., via an equation or distance metric) at first use to avoid ambiguity across figures.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the strength of evidence for the proposed mechanisms. We respond point-by-point below.

read point-by-point responses

  1. Referee: [Abstract / §4] Abstract and §4 (results on variation): the manuscript states that population scaling increases in-group variation, which in turn prevents homogeneity and produces the observed drop in cross-group variation plus ZMI gains. No mediation analysis, ablation that holds scale fixed while varying communicative diversity, or controlled comparison isolating the variation driver is described; the causal sequence therefore remains an untested assumption rather than a demonstrated mechanism.

    Authors: We agree that the manuscript presents the causal sequence as an inference from observed scaling patterns rather than through formal mediation analysis or an ablation that holds population size fixed while manipulating communicative diversity. The reported experiments show that larger populations reliably produce higher in-group variation, lower cross-group variation, and higher ZMI, but these remain correlational. We will add a mediation analysis and a controlled ablation isolating the variation driver in the revised manuscript. revision: yes

  2. Referee: [Abstract] Abstract: the claim that scaled populations 'increasingly anchor their emergent sketches on the objective visual resemblance' is presented as the explanation for universality, yet the text supplies no quantitative measure (e.g., correlation with image-feature similarity, human perceptual judgments, or ablation removing visual grounding) that would distinguish this account from alternative explanations such as richer gradients or implicit regularization.

    Authors: The manuscript's further analysis section reports quantitative correlations between emergent sketch features and objective image features that strengthen with population scale, together with supporting human judgment data. These results favor perceptual grounding over purely optimization-based alternatives. We acknowledge that an explicit ablation removing visual grounding would provide a sharper contrast and will include such an ablation in the revision. revision: yes

Circularity Check

0 steps flagged

Empirical simulation results exhibit no circularity

full rationale

The paper reports experimental findings on population scaling effects in emergent sketching agents, with ZMI, in-group variation, and cross-group convergence measured directly from simulations. No derivation chain, equations, or self-citations are invoked to derive the central claims; results are presented as observations from disjoint training populations. This matches the default expectation of self-contained empirical work against external benchmarks (simulation runs), warranting score 0 with empty steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no information on free parameters, axioms, or invented entities used in the work.

pith-pipeline@v0.9.1-grok · 5700 in / 1364 out tokens · 29121 ms · 2026-06-27T14:13:01.387154+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

67 extracted references · 6 linked inside Pith

  1. [1]

    ICLR , year=

    Multi-Agent Cooperation and the Emergence of (Natural) Language , author=. ICLR , year=

  2. [2]

    NeurIPS , year=

    Learning to Communicate with Deep Multi-Agent Reinforcement Learning , author=. NeurIPS , year=

  3. [3]

    NeurIPS , year=

    Learning multiagent communication with backpropagation , author=. NeurIPS , year=

  4. [4]

    NeurIPS , year=

    Emergence of Language with Multi-agent Games: Learning to Communicate with Sequences of Symbols , author=. NeurIPS , year=

  5. [5]

    arXiv preprint arXiv:1611.03218 , year=

    Learning to play guess who? and inventing a grounded language as a consequence , author=. arXiv preprint arXiv:1611.03218 , year=

    Pith/arXiv arXiv

  6. [6]

    ICLR , year=

    Revisiting Populations in multi-agent Communication , author=. ICLR , year=

  7. [7]

    NeurIPS , year=

    Emergent Communication under Varying Group Sizes and Connectivities , author=. NeurIPS , year=

  8. [8]

    EMNLP , year=

    Emergent linguistic phenomena in multi-agent communication games , author=. EMNLP , year=

  9. [9]

    ICLR , year=

    Emergent Communication at Scale , author=. ICLR , year=

  10. [10]

    arXiv preprint arXiv:1711.09846 , year=

    Population based training of neural networks , author=. arXiv preprint arXiv:1711.09846 , year=

    Pith/arXiv arXiv

  11. [11]

    NeurIPS , year=

    Collaborating with humans without human data , author=. NeurIPS , year=

  12. [12]

    2020 , url=

    Hu, Hengyuan and Lerer, Adam and Peysakhovich, Alex and Foerster, Jakob , booktitle=. 2020 , url=

    2020

  13. [13]

    NeurIPS , year=

    Learning to draw: Emergent communication through sketching , author=. NeurIPS , year=

  14. [14]

    arXiv preprint arXiv:2103.16194 , year=

    Differentiable drawing and sketching , author=. arXiv preprint arXiv:2103.16194 , year=

    arXiv

  15. [15]

    Emergent communication: Generalization and overfitting in

    Rita, Mathieu and Tallec, Corentin and Michel, Paul and Grill, Jean-Bastien and Pietquin, Olivier and Dupoux, Emmanuel and Strub, Florian , booktitle=. Emergent communication: Generalization and overfitting in. 2022 , url=

    2022

  16. [16]

    ACL , year=

    Compositionality and Generalization In Emergent Languages , author=. ACL , year=

  17. [17]

    NeurIPS , year=

    Emergent communication of generalizations , author=. NeurIPS , year=

  18. [18]

    ICLR , year=

    Environmental Drivers of Systematicity and Generalization in a Situated Agent , author=. ICLR , year=

  19. [19]

    ACL (Findings) , year=

    Concept-Best-Matching: Evaluating Compositionality In Emergent Communication , author=. ACL (Findings) , year=

  20. [20]

    ICML , year=

    Countering Language Drift with Seeded Iterated Learning , author=. ICML , year=

  21. [21]

    1969 , publisher=

    Convention: A Philosophical Study , author=. 1969 , publisher=

    1969

  22. [22]

    PNAS , volume=

    The Evolution of Language , author=. PNAS , volume=. 1999 , url=

    1999

  23. [23]

    2010 , publisher=

    Origins of Human Communication , author=. 2010 , publisher=

    2010

  24. [24]

    2010 , publisher=

    Signals: Evolution, Learning, and Information , author=. 2010 , publisher=

    2010

  25. [25]

    Collective Dynamics of

    Watts, Duncan and Strogatz, Steven , journal=. Collective Dynamics of. 1998 , url=

    1998

  26. [26]

    Science , volume=

    Emergence of Scaling in Random Networks , author=. Science , volume=. 1999 , url=

    1999

  27. [27]

    TMLR , issn=

    A Review of the Applications of Deep Learning-Based Emergent Communication , author=. TMLR , issn=. 2024 , url=

    2024

  28. [28]

    Pragmatics & Cognition , volume=

    Iconicity: From sign to system in human communication and language , author=. Pragmatics & Cognition , volume=. 2014 , url=

    2014

  29. [29]

    Cognitive science , volume=

    Foundations of representation: where might graphical symbol systems come from? , author=. Cognitive science , volume=. 2007 , url=

    2007

  30. [30]

    ICLR , year=

    Emergent Tool Use From Multi-Agent Autocurricula , author=. ICLR , year=

  31. [31]

    NeurIPS , year=

    On the Utility of Learning About Humans for Zero-Shot Coordination , author=. NeurIPS , year=

  32. [32]

    Natural Language Does Not Emerge '

    Kottur, Satwik and Moura, Jos. Natural Language Does Not Emerge '. EMNLP , year=

  33. [33]

    AAAI , year=

    Emergence of Grounded Compositional Language in Multi-Agent Populations , author=. AAAI , year=

  34. [34]

    ICML , year=

    Off-Belief Learning , author=. ICML , year=

  35. [35]

    ICML , year=

    A New Formalism, Method and Open Issues for Zero-Shot Coordination , author=. ICML , year=

  36. [36]

    NeurIPS , year=

    Emergent Communication in Interactive Sketch Question Answering , author=. NeurIPS , year=

  37. [37]

    ICLR , year=

    A Neural Representation of Sketch Drawings , author=. ICLR , year=

  38. [38]

    ACL , year=

    Multi-Agent Communication Meets Natural Language: Synergies Between Functional and Structural Language Learning , author=. ACL , year=

  39. [39]

    EMNLP , year=

    Countering Language Drift via Visual Grounding , author=. EMNLP , year=

  40. [40]

    IEEE Transactions on Evolutionary Computation , volume=

    Spontaneous evolution of linguistic structure---an iterated learning model of the emergence of regularity and irregularity , author=. IEEE Transactions on Evolutionary Computation , volume=. 2001 , doi=

    2001

  41. [41]

    Current Opinion in Neurobiology , volume=

    Iterated learning and the evolution of language , author=. Current Opinion in Neurobiology , volume=. 2014 , doi=

    2014

  42. [42]

    Artificial Life , volume=

    Iterated learning: a framework for the emergence of language , author=. Artificial Life , volume=. 2003 , doi=

    2003

  43. [43]

    ICML , year=

    Fully Decentralized Multi-Agent Reinforcement Learning with Networked Agents , author=. ICML , year=

  44. [44]

    ICLR , year=

    Compositional Languages Emerge in a Neural Iterated Learning Model , author=. ICLR , year=

  45. [45]

    PLoS ONE , volume=

    Language structure is partly determined by social structure , author=. PLoS ONE , volume=. 2010 , url=

    2010

  46. [46]

    2011 , url=

    Sociolinguistic Typology: Social Determinants of Linguistic Complexity , author=. 2011 , url=

    2011

  47. [47]

    Lingua , volume=

    The consequences of talking to strangers: Evolutionary corollaries of socio-cultural influences on linguistic form , author=. Lingua , volume=. 2007 , url=

    2007

  48. [48]

    PLoS ONE , volume=

    Speaker input variability does not explain why larger populations have simpler languages , author=. PLoS ONE , volume=. 2015 , url=

    2015

  49. [49]

    Psychological Science , volume=

    Variability and detection of invariant structure , author=. Psychological Science , volume=. 2002 , url=

    2002

  50. [50]

    Psychological Science , volume=

    Learn locally, think globally: Exemplar variability supports higher-order generalization and word learning , author=. Psychological Science , volume=. 2010 , url=

    2010

  51. [51]

    Cognitive Science , volume=

    How the size of our social network influences our semantic skills , author=. Cognitive Science , volume=. 2016 , url=

    2016

  52. [52]

    Proceedings of the Royal Society B: Biological Sciences , volume=

    Larger communities create more systematic languages , author=. Proceedings of the Royal Society B: Biological Sciences , volume=. 2019 , url=

    2019

  53. [53]

    Proceedings of the IEEE , volume =

    Gradient-Based Learning Applied to Document Recognition , author =. Proceedings of the IEEE , volume =. 1998 , url =

    1998

  54. [54]

    NeurIPS , year =

    ImageNet Classification with Deep Convolutional Neural Networks , author =. NeurIPS , year =

  55. [55]

    arXiv preprint arXiv:1412.6980 , year =

    Adam: A Method for Stochastic Optimization , author =. arXiv preprint arXiv:1412.6980 , year =

    Pith/arXiv arXiv

  56. [56]

    ICLR , year=

    Compositional Obverter Communication Learning from Raw Visual Input , author=. ICLR , year=

  57. [57]

    arXiv preprint arXiv:2006.02419 , year=

    Emergent multi-agent communication in the deep learning era , author=. arXiv preprint arXiv:2006.02419 , year=

    arXiv 2006

  58. [58]

    Journal of Multilingual and Multicultural Development , volume=

    The contribution of linguistic factors to the intelligibility of closely related languages , author=. Journal of Multilingual and Multicultural Development , volume=. 2007 , url=

    2007

  59. [59]

    International Journal of Multilingualism , volume=

    Mutual intelligibility between closely related languages in Europe , author=. International Journal of Multilingualism , volume=. 2018 , url=

    2018

  60. [60]

    2009 , note =

    Learning Multiple Layers of Features from Tiny Images , author =. 2009 , note =

    2009

  61. [61]

    Journal of the Royal Statistical Society: Series B (Methodological) , volume =

    Ramsey, James , title =. Journal of the Royal Statistical Society: Series B (Methodological) , volume =. 1969 , doi =

    1969

  62. [62]

    ICCV , year=

    Deep learning face attributes in the wild , author=. ICCV , year=

  63. [63]

    arXiv preprint arXiv:1712.00409 , year=

    Deep Learning Scaling is Predictable, Empirically , author=. arXiv preprint arXiv:1712.00409 , year=

    Pith/arXiv arXiv

  64. [64]

    arXiv preprint arXiv:2001.08361 , year=

    Scaling Laws for Neural Language Models , author=. arXiv preprint arXiv:2001.08361 , year=

    Pith/arXiv arXiv 2001

  65. [65]

    NeurIPS , year=

    Language Models are Few-Shot Learners , author=. NeurIPS , year=

  66. [66]

    arXiv preprint arXiv:2102.01293 , year=

    Scaling Laws for Transfer , author=. arXiv preprint arXiv:2102.01293 , year=

    Pith/arXiv arXiv

  67. [67]

    CVPR , year=

    Scaling Vision Transformers , author=. CVPR , year=