Pith. sign in

REVIEW 4 major objections 6 minor 71 references

Searching for Structure: Investigating Emergent Communication with Large Language Models

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read LLM agents shape unstructured artificial languages into structured ones through communication.

desk verdict Novel LLM referential-game study with a hard-coded alignment rule that muddies the attribution of emergent structure to LLM biases. read the letter →

arxiv 2412.07646 v3 pith:7GAIDSUC submitted 2024-12-10 cs.CL

classification cs.CL
keywords languageevolutionemergentcommunicationlargemodelsreferentialgameiteratedlearningcompositionalstructuretopographicsimilarityartificial
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks whether large language models, when they must learn and use an artificial language in a referential game, impose the same kind of structural biases that human learners do. The authors show that two LLM agents communicating about colored shapes gradually reshape an initially random holistic vocabulary so that similar meanings get similar signals, a property known as topographic structure. This structure improves the agents' ability to generalize to unseen meanings, and passing the language down to new generations of LLM learners makes it easier to learn. However, the same cultural-transmission process can drive the vocabulary toward degenerate, underspecified forms that look non-humanlike, such as longer signals and many-to-one mappings. The central claim is that LLMs are viable model learners for language-evolution experiments, not just for evaluating pre-existing structure.

What carries the argument

The Lewis referential game played by two LLM agents, in which the speaker produces a signal for a stimulus, the listener picks the most probable prefilled distractor, and after each interaction the current stimulus is associated with the produced signal in the next prompt. The game is embedded in a four-block pipeline (guessing, labelling, communication over four rounds, testing), and the iterated-learning version transmits the vocabulary of the highest-TopSim agent to the next generation. Topographic similarity, the Mantel-test z-score between signal edit distances and semantic distances, is the primary structure metric, paired with n-gram diversity and a generalisation score for unseen stimuli.

What would settle it

Run the same four-block simulation with the communication block removed (agents only label and re-label the same stimuli) or with the stimulus–signal update rule replaced by a random association; if TopSim still rises significantly and n-gram diversity still falls, the claimed causal role of communicative interaction and LLM bias would be falsified. A more targeted test would be to hold the language fixed and swap in a model with a much smaller context window; if the degenerate long-signal behaviour disappears, memory constraints rather than bias explain it.

Watch

Extended reading notes

Core claim

The paper's central discovery is that initially unstructured holistic languages are indeed shaped to have structural properties when optimised for the implicit biases of LLM agents. In the authors' simulation, two instruction-tuned LLMs play a Lewis referential game: one produces a signal for a target stimulus, the other guesses the target among distractors, and the interaction updates each agent's vocabulary by associating the produced signal with the current stimulus. Over four rounds, topographic similarity (TopSim) rises significantly and n-gram diversity falls, which the authors interpret as the emergence of compositional structure. The structured vocabularies then support better generalisation to unseen stimuli, measured by a generalisation score. Iterated learning across eight generations reduces the edit distance between ground-truth and produced signals, meaning the languages become easier for LLMs to learn, but it also yields degenerate vocabularies (repeated signals for different meanings) and longer signals, behaviours that diverge from human experimental findings. The authors take this as evidence that LLMs can serve as flexible tools in simulations of language evolution, and that the set-up exposes the models' own structural preferences.

Load-bearing premise

The central claim depends on the assumption that the observed structure emerges from the LLMs' own inductive biases, yet the setup includes a hard-coded vocabulary-alignment step that updates the current stimulus to be associated with the produced signal, and no control condition omits communication or this update rule.

Editorial extensions

If this is right

  • LLM-based agents can substitute for human participants in at least some language-evolution experiments, lowering cost and enabling large-scale systematic manipulations.
  • The finding that structured languages support better generalization in LLMs suggests that compositional pressures benefit machine learners, not just human learners.
  • The observed non-humanlike drift (longer signals, degenerate vocabularies) implies that LLM-driven iterated learning must be interpreted with caution; the models' memory and decoding biases are not human-like.
  • Because communicative success rose only modestly and varied across simulations, the setup is sensitive to initial conditions and may need explicit expressivity pressures to maintain human-like structure.
  • The increase in learnability across generations demonstrates that weak inductive biases in LLMs can be amplified by cultural transmission, a process previously documented in humans.

Reading between the lines

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

  • The built-in vocabulary-update rule, which associates each produced signal with the current stimulus after every interaction, may be driving most of the convergence rather than the LLMs' communicative reasoning; without a control condition that removes communication or randomizes the association, the causal attribution to LLM biases remains open.
  • The absence of a genuine memory bottleneck in LLMs, which have large context windows, could explain why iterated learning produces degenerate underspecified vocabularies instead of the compressed compositional ones seen in human experiments.
  • The paper's metric suite could be ported to other instruction-tuned models or to decoding strategies with higher temperature; if structure emergence disappears under those variations, the effect is tied to the specific model's biases rather than a general property of LLMs.
  • A direct human-LLM comparison in the same referential game would be the natural next test of whether LLMs truly serve as model learners in the sense claimed; the authors note such experiments are future work.
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 / 6 minor

Summary. The paper reports a series of simulations in which two Llama 3 70B agents play a referential game over an initially holistic artificial language, with iterated-learning chains in which one agent's vocabulary is transmitted to the next generation. The authors measure communicative success (PercCom), topographic similarity (TopSim), Ngram diversity, and generalization (GenScore), and report that communication increases structure, structured languages generalize better, and generational transmission increases learnability while producing degenerate vocabularies. The central claim is that LLM implicit biases shape unstructured languages toward structure, making LLMs useful model learners for language evolution research.

Significance. If the central attribution is established, the paper would be a valuable demonstration that instruction-tuned LLMs exhibit inductive biases in artificial-language formation, extending prior work by Galke et al. (2023) and Kirby et al. (2015) and providing a tractable computational substrate for iterated-learning experiments. The study uses standard metrics, transparent statistical models with conditional and marginal R2, and reports quantitative results with uncertainties; it also explicitly acknowledges prompt sensitivity, single-model, and small-N limitations. However, the current design does not separate the hard-coded vocabulary-update mechanism from LLM-generated structure, so the headline claim is not yet supported. The specific chance-level error in Section 5.2 is also directly relevant to the conclusion that communication is successful.

major comments (4)
  1. [Section 3.2] The communication block hard-codes a vocabulary-alignment step: "updating the current stimulus to be associated with the produced signal" after interactions. The paper reports no control condition that omits this update, omits communication, or replaces the LLM-produced signals with a random or static generator. Because this update overwrites both agents' stored mappings for a stimulus with the speaker's signal, it externally forces convergence and alters the prompt vocabulary for all later blocks. The observed TopSim increase, Ngram decrease, and roughly 75% PercCom in Sections 5.2-5.3 are therefore compatible with an explanation in which the update rule plus the structured prompt format, rather than any implicit structural bias of the LLMs, drives the results; the iterated-learning chains inherit the same confound. Please add control simulations to isolate the LLMs' contribution.
  2. [Section 5.2] The task is to discriminate "a target stimulus from four distractor stimuli", which is naturally read as five alternatives, so chance-level communicative success is 20%, not 25% as stated. The 70% and 75% PercCom figures would still be above chance, but the size of the effect is smaller and the statistical comparison to chance should be recomputed and reported with the correct baseline. If the design instead used four total alternatives, the wording should be corrected to avoid ambiguity.
  3. [Section 6] Only the vocabulary of the agent with the highest TopSim is transmitted to the next generation. This selection rule can itself produce the observed increase in learnability and decrease in Ngram diversity, because high-TopSim vocabularies are already more regular and easier to reproduce. A control that transmits a randomly chosen agent, or that averages over agents, is needed to support the claim that generational transmission per se amplifies structural properties.
  4. [Sections 5.2-5.3] The claim that increases in structure "benefit communicative success positively" (beta = 0.035) is in apparent tension with the earlier statement that PercCom does not increase significantly across rounds. The regression model is not described fully enough to know whether this coefficient reflects within-round time dynamics, between-simulation variation, or a combination; please clarify the specification and report the relationship between PercCom and TopSim conditional on round.
minor comments (6)
  1. [Section 3.2] Specify whether the vocabulary update is applied after every interaction or after every round, and whether it applies to both agents' stored mappings; the current sentence is ambiguous and the magnitude of the confound depends on this timing.
  2. [Section 3.3] The statement that "initial explorations with GPT-3.5 and Llama 2 7B showed similar behaviours" is not accompanied by data; if this claim is retained, a supplementary table or figure should be provided.
  3. [Section 7] The explanation that LLMs produce longer signals because their context windows are "considerably larger" is asserted without a manipulation or analysis; present it as a hypothesis rather than as a conclusion.
  4. [Section 6] The paper uses "learnability" to refer to the edit distance between ground-truth and reproduced signals in the labelling block; this operationalization should be stated more explicitly to avoid conflating reproduction accuracy with acquisition in the human sense.
  5. [Section 3] There is a typo: the text refers to "Figure 11" when describing the four simulation blocks, but the corresponding figure in the manuscript is Figure 1.
  6. [General] The paper does not provide a code or data availability statement; for a computational study of this type, a reproducibility statement would be helpful.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all reported structure and learnability metrics are computed from LLM outputs; the §3.2 update rule is a design choice, not a fitted parameter.

full rationale

The paper's central result—that initially unstructured holistic languages become more structured and learnable when used by LLM agents—is an empirical measurement over the models' outputs, not a derivation from fitted parameters. TopSim, Ngram diversity, PercCom, and GenScore are computed directly from produced signals and task success; no quantity that is later called a prediction is used as an input to fit another quantity. The vocabulary-alignment update in §3.2 ('updating the current stimulus to be associated with the produced signal') is a standard interaction rule from the naming-game literature and does not itself assign structured signal inventories: the structure it can propagate must be generated by the LLM's own in-context completions. The absence of a no-communication/no-update control is a validity concern about attributing causality to LLM biases, not a circularity of the derivation chain, since the update rule is not equivalent to the TopSim or learnability gains being claimed and no fitted value is renamed as a prediction. Self-citations (e.g., Kouwenhoven et al. 2022b, Peeperkorn et al. 2024) appear only as contextual background or speculation about future temperature effects, not as load-bearing justifications. The pass therefore finds no significant circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No free parameters are fitted to data; the listed parameters are hand-chosen experimental design constants. The central claim rests on domain assumptions about LLM in-context learning and on the hard-coded vocabulary update rule. No new entities are postulated.

free parameters (4)
  • Number of training stimuli = 15 of 27
    Hand-chosen to match Kirby et al. (2015); the set is balanced across attribute values, and the remaining 12 stimuli are used only in testing.
  • Simulation count = 15 simulations; 6 transmission chains
    Hand-chosen; small sample affects statistical power and chain-level variability.
  • Communication rounds and tasks = 4 rounds x 30 tasks
    Hand-chosen; TopSim and Ngram trends are measured across these rounds.
  • Generations per chain = 8
    Hand-chosen; learnability and degeneracy effects are assessed over this horizon.
assumptions (5)
  • domain assumption LLM in-context learning with a JSON-like prompt is a sufficient model of artificial language learning.
    Section 3.3 states the authors 'assume that a vocabulary of signal-meaning mappings in the context of a prompt provides enough (distributional) information for a LLM to learn an appropriate mapping.'
  • ad hoc to paper The hard-coded update rule ('updating the current stimulus to be associated with the produced signal') implements the intended communicative alignment and does not itself produce the observed structure.
    Section 3.2 introduces the update rule without a control condition that would separate its effect from the LLMs' own biases.
  • domain assumption TopSim (Mantel z-score) and Ngram diversity are valid measures of linguistic structure for this comparison.
    Section 3.4 defines these metrics and uses them as the main evidence for structure emergence; their validity for LLM-generated signals is assumed.
  • domain assumption Results from human iterated-learning experiments (e.g., Kirby et al. 2015) are comparable baselines for 'human-like' behavior.
    Section 7 says the authors are 'not able to directly compare our results to human data' yet repeatedly contrasts LLM behavior with human experiments.
  • standard math Linear mixed effects models with random intercepts by simulation adequately handle the nested data.
    Section 4 describes the LME approach and conditional/marginal R2; this is standard but assumes correct random effect structure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Searching for Structure: Investigating Emergent Communication with Large Language Models." pith.science (2026). https://pith.science/paper/7GAIDSUC

@misc{pith2026241207646,
  author       = {Pith},
  title        = {Pith review of: Searching for Structure: Investigating Emergent Communication with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7GAIDSUC}},
  note         = {Machine review of arXiv:2412.07646}
}
read the original abstract

Human languages have evolved to be structured through repeated language learning and use. These processes introduce biases that operate during language acquisition and shape linguistic systems toward communicative efficiency. In this paper, we investigate whether the same happens if artificial languages are optimised for implicit biases of Large Language Models (LLMs). To this end, we simulate a classical referential game in which LLMs learn and use artificial languages. Our results show that initially unstructured holistic languages are indeed shaped to have some structural properties that allow two LLM agents to communicate successfully. Similar to observations in human experiments, generational transmission increases the learnability of languages, but can at the same time result in non-humanlike degenerate vocabularies. Taken together, this work extends experimental findings, shows that LLMs can be used as tools in simulations of language evolution, and opens possibilities for future human-machine experiments in this field.

Figures

Figures reproduced from arXiv: 2412.07646 by the authors.

Figure 1
Figure 1. A graphical representation of the experimental blocks. The agents first go through a guessing block before [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Communication clearly increases the structure of the vocabularies, as seen by the increasing [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Languages that have evolved to be more struc [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The normalised Levenshtein distance between [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 7
Figure 7. Figure 7: The evolution of TopSim on the words pro￾duced in the testing block. Each line indicates a chain, and the dashed blue line is the average with bars indi￾cating the 95% confidence interval. See [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 6
Figure 6. Figure 6: The average communicative success across [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

71 extracted references · 23 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Stubbersfield

    Alberto Acerbi and Joseph M. Stubbersfield. 2023. https://doi.org/10.1073/pnas.2313790120 Large language models show human-like content biases in transmission chain experiments . Proceedings of the National Academy of Sciences, 120(44):e2313790120

  4. [4]

    Arriaga, and Adam Tauman Kalai

    Gati V Aher, Rosa I. Arriaga, and Adam Tauman Kalai. 2023. https://proceedings.mlr.press/v202/aher23a.html Using large language models to simulate multiple humans and replicate human subject studies . In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 337--371. PMLR

  5. [5]

    Inbal Arnon and Simon Kirby. 2024. https://doi.org/10.1038/s41598-024-56152-9 Cultural evolution creates the statistical structure of language . Scientific Reports, 14(1):5255

  6. [6]

    Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell

    Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. https://doi.org/10.1145/3442188.3445922 On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, FAccT '21, page 610–623, New York, NY, USA. Association for Computing...

  7. [7]

    Robert Boyd, Peter J Richerson, et al. 1996. Why culture is common, but cultural evolution is rare. In Proceedings-british academy, volume 88, pages 77--94. Oxford University Press Inc

  8. [8]

    Henry Brighton and Simon Kirby. 2006. https://doi.org/10.1162/artl.2006.12.2.229 Understanding linguistic evolution by visualizing the emergence of topographic mappings . Artificial Life, 12(2):229--242

Show all 71 references
  1. [9]

    Levin Brinkmann, Fabian Baumann, Jean-Fran c ois Bonnefon, Maxime Derex, Thomas F M \"u ller, Anne-Marie Nussberger, Agnieszka Czaplicka, Alberto Acerbi, Thomas L Griffiths, Joseph Henrich, et al. 2023. https://doi.org/10.1038/s41562-023-01742-2 Machine culture . Nature Human ...

  2. [10]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey W...

  3. [11]

    Carr, Kenny Smith, Hannah Cornish, and Simon Kirby

    Jon W. Carr, Kenny Smith, Hannah Cornish, and Simon Kirby. 2017. https://doi.org/10.1111/cogs.12371 The cultural evolution of structured languages in an open-ended, continuous world . Cognitive Science, 41(4):892--923

  4. [12]

    Angelica Chen, Ravid Shwartz-Ziv, Kyunghyun Cho, Matthew L Leavitt, and Naomi Saphra. 2024. https://openreview.net/forum?id=MO5PiKHELW Sudden drops in the loss: Syntax acquisition, phase transitions, and simplicity bias in MLM s . In The Twelfth International Conference on Lea...

  5. [13]

    Christiansen and Nick Chater

    Morten H. Christiansen and Nick Chater. 2008. https://doi.org/10.1017/S0140525X08004998 Language as shaped by the brain . Behavioral and Brain Sciences, 31(5):489–509

  6. [14]

    Jennifer Culbertson and Paul Smolensky. 2012. https://doi.org/10.1111/j.1551-6709.2012.01264.x A bayesian model of biases in artificial language learning: The case of a word-order universal . Cognitive Science, 36(8):1468--1498

  7. [15]

    Bart de Boer . 2000. https://doi.org/10.1006/jpho.2000.0125 Self-organization in vowel systems . Journal of Phonetics, 28(4):441--465

  8. [16]

    Bart de Boer. 2006. https://doi.org/10.1007/1-4020-3395-8_17 Computer modelling as a tool for understanding language evolution , pages 381--406. Springer Netherlands, Dordrecht

  9. [17]

    Danica Dillion, Niket Tandon, Yuling Gu, and Kurt Gray. 2023. https://doi.org/10.1016/j.tics.2023.04.008 Can ai language models replace human participants? Trends in Cognitive Sciences, 27(7):597--600

  10. [18]

    Florian Jaeger, and Elissa L

    Maryia Fedzechkina, T. Florian Jaeger, and Elissa L. Newport. 2012. https://doi.org/10.1073/pnas.1215776109 Language learners restructure their input to facilitate efficient communication . Proceedings of the National Academy of Sciences, 109(44):17897--17902

  11. [19]

    Bruno Galantucci. 2005. https://doi.org/10.1207/s15516709cog0000\_34 An experimental study of the emergence of human communication systems . Cognitive Science, 29(5):737--767

  12. [20]

    Lukas Galke, Yoav Ram, and Limor Raviv. 2022. https://openreview.net/forum?id=rqUGZQ-0XZ5 Emergent communication for understanding human language evolution: What's missing? In Emergent Communication Workshop at ICLR 2022

  13. [21]

    Lukas Galke, Yoav Ram, and Limor Raviv. 2023. What makes a language easy to deep-learn? arXiv preprint arXiv:2302.12239

  14. [22]

    Lukas Galke and Limor Raviv. 2024. https://doi.org/10.34842/3vr5-5r49 Learning and communication pressures in neural networks: Lessons from emergent communication . Language Development Research, 5(1):116--143

  15. [23]

    Griffiths and Michael L

    Thomas L. Griffiths and Michael L. Kalish. 2007. https://doi.org/10.1080/15326900701326576 Language evolution by iterated learning with bayesian agents . Cognitive Science, 31(3):441--480

  16. [24]

    Charles F. Hockett. 1960. https://www.jstor.org/stable/24940617 The origin of speech . Scientific American, 203(3):88--97

  17. [25]

    Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. https://openreview.net/forum?id=rygGQyrFvH The curious case of neural text degeneration . In International Conference on Learning Representations

  18. [26]

    Jennifer Hu and Michael Frank. 2024. https://openreview.net/forum?id=U5BUzSn4tD Auxiliary task demands mask the capabilities of smaller language models . In First Conference on Language Modeling

  19. [27]

    Jennifer Hu and Roger Levy. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.306 Prompting is not a substitute for probability measurements in large language models . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5040--5060, ...

  20. [28]

    Julie Kallini, Isabel Papadimitriou, Richard Futrell, Kyle Mahowald, and Christopher Potts. 2024. https://doi.org/10.18653/v1/2024.acl-long.787 Mission: Impossible language models . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Vol...

  21. [29]

    Aisha Khatun and Daniel G Brown. 2024. https://arxiv.org/abs/2401.07955 A study on large language models' limitations in multiple-choice question answering . Preprint, arXiv:2401.07955

  22. [30]

    Simon Kirby, Hannah Cornish, and Kenny Smith. 2008. https://doi.org/10.1073/pnas.0707835105 Cumulative cultural evolution in the laboratory: An experimental approach to the origins of structure in human language . Proceedings of the National Academy of Sciences, 105(31):10681--10686

  23. [31]

    Simon Kirby, Monica Tamariz, Hannah Cornish, and Kenny Smith. 2015. https://doi.org/10.1016/j.cognition.2015.03.016 Compression and communication in the cultural evolution of linguistic structure . Cognition, 141:87--102

  24. [32]

    Tom Kouwenhoven, Roy De Kleijn, Stephan Raaijmakers, and Tessa Verhoef. 2022 a . Need for structure and the emergence of communication. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 44

  25. [33]

    Tom Kouwenhoven, Max Peeperkorn, Bram Van Dijk, and Tessa Verhoef. 2024. https://doi.org/10.18653/v1/2024.cmcl-1.5 The curious case of representational alignment: Unravelling visio-linguistic tasks in emergent communication . In Proceedings of the Workshop on Cognitive Modelin...

  26. [34]

    Tom Kouwenhoven, Tessa Verhoef, Roy De Kleijn, and Stephan Raaijmakers. 2022 b . https://doi.org/10.3389/frai.2022.886349 Emerging grounded shared vocabularies between human and machine, inspired by human language evolution . Frontiers in Artificial Intelligence, 5:886349

  27. [35]

    Angeliki Lazaridou and Marco Baroni. 2020. https://arxiv.org/abs/2006.02419 Emergent multi-agent communication in the deep learning era . Preprint, arXiv:2006.02419

  28. [36]

    Yuchen Lian, Arianna Bisazza, and Tessa Verhoef. 2023. https://doi.org/10.1162/tacl_a_00587 Communication Drives the Emergence of Language Universals in Neural Agents: Evidence from the Word-order/Case-marking Trade-off . Transactions of the Association for Computational Lingu...

  29. [37]

    Yuchen Lian, Tessa Verhoef, and Arianna Bisazza. 2024. https://arxiv.org/abs/2407.13999 Nellcom-x: A comprehensive neural-agent framework to simulate language learning and group communication . Preprint, arXiv:2407.13999

  30. [38]

    Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang

    Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. https://doi.org/10.1162/tacl_a_00638 Lost in the middle: How language models use long contexts . Transactions of the Association for Computational Linguistics, 12...

  31. [39]

    Meta Llama Team. 2024. https://arxiv.org/abs/2407.21783 The llama 3 herd of models . Preprint, arXiv:2407.21783

  32. [40]

    Nathan Mantel. 1967. The detection of disease clustering and a generalized regression approach. Cancer research, 27(2\_Part\_1):209--220

  33. [41]

    Clara Meister, Tiago Pimentel, Gian Wiher, and Ryan Cotterell. 2023. https://doi.org/10.1162/tacl_a_00536 Locally typical sampling . Transactions of the Association for Computational Linguistics, 11:102--121

  34. [42]

    Clara Meister, Gian Wiher, Tiago Pimentel, and Ryan Cotterell. 2022. https://doi.org/10.18653/v1/2022.acl-short.5 On the probability -- quality paradox in language generation . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2...

  35. [43]

    Swaroop Mishra, Daniel Khashabi, Chitta Baral, Yejin Choi, and Hannaneh Hajishirzi. 2022. https://doi.org/10.18653/v1/2022.findings-acl.50 Reframing instructional prompts to GPT k ' s language . In Findings of the Association for Computational Linguistics: ACL 2022, pages 589-...

  36. [44]

    Shinichi Nakagawa and Holger Schielzeth. 2013. https://doi.org/10.1111/j.2041-210x.2012.00261.x A general and simple method for obtaining r2 from generalized linear mixed-effects models . Methods in Ecology and Evolution, 4(2):133--142

  37. [45]

    Jonas Nölle, Marlene Staib, Riccardo Fusaroli, and Kristian Tylén. 2018. https://doi.org/10.1016/j.cognition.2018.08.014 The emergence of systematicity: How environmental and communicative factors shape a novel communication system . Cognition, 181:93--104

  38. [46]

    Bernstein

    Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. https://doi.org/10.1145/3586183.3606763 Generative agents: Interactive simulacra of human behavior . In Proceedings of the 36th Annual ACM Symposium on User Int...

  39. [47]

    Max Peeperkorn, Tom Kouwenhoven, Dan Brown, and Anna Jordanous. 2024. Is temperature the creativity parameter of large language models? In 15th International Conference on Computational Creativity. Association for Computational Creativity

  40. [48]

    Piantadosi, Harry Tily, and Edward Gibson

    Steven T. Piantadosi, Harry Tily, and Edward Gibson. 2011. https://doi.org/10.1073/pnas.1012551108 Word lengths are optimized for efficient communication . Proceedings of the National Academy of Sciences, 108(9):3526--3529

  41. [49]

    Limor Raviv, Marianne de Heer Kloots , and Antje Meyer. 2021. https://doi.org/10.1016/j.cognition.2021.104620 What makes a language easy to learn? a preregistered study on how systematic structure and community size affect language learnability . Cognition, 210:104620

  42. [50]

    Limor Raviv, Antje Meyer, and Shiri Lev-Ari. 2019 a . https://doi.org/10.1016/j.cognition.2018.09.010 Compositional structure can emerge without generational transmission . Cognition, 182:151--164

  43. [51]

    Limor Raviv, Antje Meyer, and Shiri Lev-Ari. 2019 b . https://doi.org/10.1098/rspb.2019.1262 Larger communities create more systematic languages . Proceedings of the Royal Society B: Biological Sciences, 286(1907):20191262

  44. [52]

    Scott-Phillips, Simon Kirby, and Graham R.S

    Thomas C. Scott-Phillips, Simon Kirby, and Graham R.S. Ritchie. 2009. https://doi.org/10.1016/j.cognition.2009.08.009 Signalling signalhood and the emergence of communication . Cognition, 113(2):226--233

  45. [53]

    Reinhard Selten and Massimo Warglien. 2007. https://doi.org/10.1073/pnas.0702077104 The emergence of simple languages in an experimental coordination game . Proceedings of the National Academy of Sciences, 104(18):7361--7366

  46. [54]

    Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, and Yarin Gal. 2024. https://doi.org/10.1038/s41586-024-07566-y Ai models collapse when trained on recursively generated data . Nature, 631(8022):755--759

  47. [55]

    Kenny Smith. 2011. https://doi.org/10.3378/027.083.0207 Learning Bias, Cultural Evolution of Language, and theBiological Evolution of the Language Faculty . Human Biology, 83(2):261 -- 278

  48. [56]

    Kenny Smith. 2022. https://doi.org/10.1177/09637214211068127 How Language Learning and Language Use Create Linguistic Structure . Current Directions in Psychological Science, 31(2):177--186

  49. [57]

    Kenny Smith, Simon Kirby, Shangmin Guo, and Thomas L Griffiths. 2024. https://doi.org/10.1038/d41586-024-03023-y Ai model collapse might be prevented by studying human language transmission . Nature, 633(8030):525

  50. [58]

    Kenny Smith, Monica Tamariz, and Simon Kirby. 2013. Linguistic structure is an evolutionary trade-off between simplicity and expressivity. In Proceedings of the annual meeting of the cognitive science society, volume 35

  51. [59]

    Luc Steels, Martin Loetzsch, et al. 2012. The grounded naming game. Experiments in cultural language evolution, 3:41--59

  52. [60]

    Mónica Tamariz and Simon Kirby. 2015. https://doi.org/10.1111/cogs.12144 Culture: Copying, compression, and conventionality . Cognitive Science, 39(1):171--183

  53. [61]

    Michael Tomasello. 1999. http://www.jstor.org/stable/j.ctvjsf4jc The Cultural Origins of Human Cognition . Harvard University Press

  54. [62]

    Bram van Dijk, Tom Kouwenhoven, Marco Spruit, and Max Johannes van Duijn. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.779 Large language models: The need for nuance in current debates and a pragmatic perspective on understanding . In Proceedings of the 2023 Conference on...

  55. [63]

    Tessa Verhoef. 2012. https://doi.org/10.1515/langcog-2012-0019 The origins of duality of patterning in artificial whistled languages . Language and Cognition, 4(4):357–380

  56. [64]

    Tessa Verhoef, Esther Walker, and Tyler Marghetis. 2016. Cognitive biases and social coordination in the emergence of temporal language. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 38

  57. [65]

    Tessa Verhoef, Esther Walker, and Tyler Marghetis. 2022. https://escholarship.org/uc/item/7vj8b55x Interaction dynamics affect the emergence of compositional structure in cultural transmission of space-time mappings . In Proceedings of the Annual Meeting of the Cognitive Scien...

  58. [66]

    Alex Warstadt and Samuel R Bowman. 2022. What artificial neural networks can tell us about human language acquisition. Algebraic Structures in Natural Language, pages 17--60

  59. [67]

    Lucas Weber, Elia Bruni, and Dieuwke Hupkes. 2023. https://doi.org/10.18653/v1/2023.conll-1.20 Mind the instructions: a holistic evaluation of consistency and interactions in prompt-based learning . In Proceedings of the 27th Conference on Computational Natural Language Learni...

  60. [68]

    Ethan Gotlieb Wilcox, Richard Futrell, and Roger Levy. 2023. https://doi.org/10.1162/ling_a_00491 Using Computational Models to Test Syntactic Learnability . Linguistic Inquiry, pages 1--44

  61. [69]

    Eunice Yiu, Eliza Kosoy, and Alison Gopnik. 2024. https://doi.org/10.1177/17456916231201401 Transmission versus truth, imitation versus innovation: What children can do that large language and language-and-vision models cannot (yet) . Perspectives on Psychological Science, 19(...

  62. [70]

    Yuqing Zhang, Tessa Verhoef, Gertjan van Noord, and Arianna Bisazza. 2024. https://aclanthology.org/2024.lrec-main.516 Endowing neural language learners with human-like biases: A case study on dependency length minimization . In Proceedings of the 2024 Joint International Conf...

  63. [71]

    Chenhao Zheng, Jieyu Zhang, Aniruddha Kembhavi, and Ranjay Krishna. 2024. Iterated learning improves compositionality in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13785--13795

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.