Pith. sign in

REVIEW 3 major objections 3 minor 53 references

A population of language-model agents, each pairing frozen visual perception with private vector memory and text-only reasoning, can invent a shared grounded lexicon for visual entities that have no name in any human language.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 11:44 UTC pith:R7I5RAES

load-bearing objection Solid simulation of LLM-based referential games, but the central 'out-of-distribution' claim rests on an unverified assumption that the paper's own Figure 2b likely contradicts. the 3 major comments →

arxiv 2607.22591 v1 pith:R7I5RAES submitted 2026-06-10 cs.AI cs.MA

Lexical discovery in unknown environments orchestrated by Large Language Models

Rafael Sendra-Arranz , I\~naki Dellibarda Varela , Eduardo Rocon , \'Alvaro Guti\'errez , Manuel Cebrian This is my paper
classification cs.AI cs.MA
keywords symbol groundingemergent communicationlarge language modelsvision-language modelsreferential gameslexical convergencemulti-agent systemsout-of-distribution referents
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that populations of language-model agents can do what individual models cannot: invent new words for visual entities that have no name in any human language, agree among themselves on those words, and tie each invented word to familiar English descriptions. The proposed mechanism—called NSLD—couples frozen visual perception to a private similarity-search memory and text-only reasoning, so symbols are grounded in persistent embedding vectors rather than in conversation. Across 20 trials per configuration, populations from 2 to 20 agents and 3 to 10 alien images reached at least 95% grounding consensus, and three fitted models predict the consensus curve, the round needed for a target consensus, and vocabulary growth with R² above 0.95. The payoff, if right, is a route to human-understandable vocabularies for robots exploring genuinely unknown environments.

Core claim

The paper claims that perceptual grounding need not be inherited from pretraining: it can be created by a population. Agents observe synthetic images that the authors take to be out-of-distribution with respect to both the visual encoder and the language model, then play pairwise referential games. Each agent stores, in a private vector index, the image embedding associated with each invented three-syllable word, and updates confidence scores by +0.25 on success and −0.20 on failure. The authors report that this process reliably converges to a unique shared alien word per referent across all tested population sizes and referent counts, and that the shared words land near coherent English des

What carries the argument

The load-bearing object is the grounding referential game combined with the Perception–Memory–Brain loop. Perception is a frozen vision-language encoder projecting images into a shared embedding space; memory is a private vector index holding each alien word's grounding vector; reasoning is a text-only LLM with tools to observe, select, add, replace, and commit words. The game's pairwise success/failure feedback with confidence updates drives winner-take-all consensus, while semantic anchoring to English happens by nearest-neighbor search against a fixed perceptual base in the same embedding space. The analytical models—a logistic direct model, its algebraic inverse, and a vocabulary model m

Load-bearing premise

The load-bearing premise is that the ten synthetic images really are out-of-distribution for both the visual encoder and the language model; the paper assumes this from the generation prompt and reports no quantitative check, so if the images resemble known concepts, the claimed new grounding could be pre-trained knowledge in disguise.

What would settle it

Compute, for each of the ten images, the cosine similarity between its embedding and the text embeddings of a broad list of English nouns, and ask the language model to name or describe each image zero-shot; if any image is consistently matched to a specific known concept, or if invented words cluster near familiar English descriptors, the out-of-distribution premise and the 'genuinely acquired' conclusion weaken.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A human operator could query the English descriptors nearest to an alien word and get a legible gloss of what the agents are seeing, without seeing the image.
  • Mission planners could use the fitted models to estimate how many rounds a swarm needs before a shared lexicon is usable, and how large the lexicon will grow.
  • Adding agents or referents delays convergence roughly as power laws; the slowest tested case—20 agents and 10 referents—took about 4,500 rounds on average.
  • Because grounding lives in persistent vector memory rather than prompt context, agents retain vocabulary across long missions and after conversations end.
  • The framework extends classical naming-game consensus from arbitrary tokens to perceptually grounded symbols linked to natural language.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the out-of-distribution premise fails—if the synthetic images are actually near known concepts in embedding space—then the 'new' grounding may be inherited from the frozen encoder rather than acquired; a cheap check is to compute nearest-English-neighbor similarities for all ten images.
  • The same machinery could be tested with real robotic cameras in novel scenes, such as underwater or planetary analogs, where the shared-vocabulary claim becomes falsifiable against human annotations.
  • Because consensus is driven by the shared embedding space, the invented words are only as interpretable to humans as the encoder's English anchoring; in domains with poor visual-language coverage, alien words may be shared but opaque.
  • The fitted models were trained on one language-model backend and fully connected agent graphs; using them for planning in larger or sparsely connected swarms assumes those power-law relations transfer, which the paper does not test.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper introduces the Neuro-Symbolic Lexical Discovery (NSLD) framework, in which populations of LLM-based agents play a referential game over synthetic images that are claimed to be out-of-distribution (OOD) for both CLIP and the LLM backend. Each agent uses a frozen CLIP encoder for perception, a private FAISS index for lexical memory, and a text-only LLM for reasoning. The authors report that populations of up to twenty agents converge to a shared alien lexicon for up to ten visual referents, that alien words become anchored to English descriptors through embedding-space proximity, and that the convergence dynamics are captured by three analytical models—direct, inverse, and vocabulary—with R² > 0.95, proposed as tools for pre-deployment planning. The manuscript includes detailed game mechanics, prompt designs, model derivations, and 20 trials per condition for all (n_a, n_e) configurations.

Significance. If the central claims hold, NSLD would be a meaningful contribution to emergent communication and symbol grounding: it combines modern LLM-based agents with persistent perceptual memory and demonstrates collective lexicon formation in a multi-entity referential game without hand-crafted priors over referent names. The paper is transparent about game rules, prompts, tool interfaces, and model fitting, and the use of 20 trials with 95% confidence intervals across all configurations is a strength. However, the two headline claims—genuine grounding on OOD referents and predictive analytical models—are currently not supported by the evidence presented. The OOD assumption is unverified and partially contradicted by the paper's own descriptor results, while the R² values are in-sample fits rather than held-out predictions. These issues are central to the paper's novelty and require substantial additional evidence.

major comments (3)
  1. [§2.1, §2.2, §4.5.4] The claim that the ten Nano Banana images are OOD with respect to both CLIP and the LLM is load-bearing: the Discussion states that grounding is 'genuinely acquired rather than inherited by pre-trained knowledge' only because the referents are absent from training data. The only support is the text prompt in §4.5.4. No quantitative OOD check is reported. More seriously, Fig. 2b and §2.2 show that CLIP associates each alien word with coherent English descriptors such as 'rock' and 'creature', and the text acknowledges that descriptor overlap is 'consistent with the visual similarity of these referents to such concepts.' This is direct evidence that CLIP maps these images to familiar, nameable categories. If so, agents are re-labeling known visual categories with novel tokens, and the 'genuinely acquired' grounding claim collapses. I ask for quantitative validation: zero-shot naming accura
  2. [§4.7, §4.8, Fig. 4] The three analytical models are fitted to the same trial-averaged G(T) and V(T) trajectories they are then said to 'predict'. The direct model's parameters (c_k, α_k, β_k, c_m, α_m, β_m) and the vocabulary model's parameters (c_A, α_A, β_A, γ, c_λ, α_λ, β_λ) are estimated by nonlinear least squares on exactly the data displayed in Fig. 4, so the reported R² > 0.95 is an in-sample goodness-of-fit statistic, not a predictive validation. Because the models have many free parameters and the trajectories are smooth sigmoids, high in-sample R² is expected. The claim that these models 'enable deployment planning prior to real-world exploration' requires held-out evaluation, e.g., leave-one-condition-out cross-validation over (n_a, n_e), or at least a split into training and test configurations. Until this is provided, the predictive component of the paper's contribution is unsubstantiated.
  3. [Eq. (3), §4.7.2] The inverse model is presented as 'parameter-free', but it is an algebraic inversion of the fitted logistic function and inherits the fitted parameters T_0.5 and k from the direct model. Consequently, the parity plot in Fig. 4b is not an independent validation: for the mean trajectory, the inverse model will reproduce the empirical threshold rounds as long as the direct model's in-sample fit is good. This is not a distinct empirical success. The text should be revised to state clearly that the inverse model is a derived consequence of the direct model, and its accuracy should be assessed on held-out data rather than presented as an additional fitted result.
minor comments (3)
  1. [§4.7.3, Eq. (21)] The parameter handling for γ is inconsistent. In the main text Eq. (5) only A and λ_v are modeled as power laws and γ is listed as a free parameter, but §4.7.3 Eq. (21) gives γ = c_γ n_a^{α_γ} n_e^{β_γ}, and Supplementary Table S1 includes bounds and initial values for c_γ, α_γ, β_γ. The main text should state which parameters are fitted and how many degrees of freedom each model uses.
  2. [Data availability / Code availability] The manuscript says 'Data availability: Not applicable', but the synthetic image dataset is central to the experiments and the code is only released upon acceptance. For a claim about emergent grounding on specific images, the dataset and evaluation code should be made available at least as supplementary material for reviewers, and the data availability statement should describe the dataset rather than saying 'Not applicable'.
  3. [Fig. 2b, §2.2] The claim that descriptor overlap is 'minimal' is not quantified. Since this is used to argue that CLIP discriminates among the referents, a quantitative measure (e.g., overlap coefficients or nearest-descriptor distance matrices) would strengthen the presentation.

Circularity Check

3 steps flagged

Analytical 'predictions' reduce to in-sample curve fits; inverse model is the algebraic inverse of the fitted direct model.

specific steps
  1. fitted input called prediction [§4.7.1 and §4.8 (direct model G(T), Eqs. 1–2, Fig. 4a)]
    "The six parameters (ck, αk, βk, cm, αm, βm) are estimated jointly by nonlinear least squares using the Trust Region Reflective algorithm (see Sec. 4.8) and reported in Fig. 4. ... All three models were fitted by nonlinear least squares minimization of the sum of squared residuals using the Trust Region Reflective (TRF) algorithm as implemented in scipy.optimize.curve_fit."

    The direct model is presented as 'predicting' grounding consensus and validated with R²>0.95, but its logistic and power-law parameters are fitted to exactly the same trial-averaged G(T) trajectories shown in Fig. 3a. The reported R² is therefore an in-sample goodness-of-fit measure, not an independent prediction. Calling this fitted curve a 'prediction' is the fitted-input-called-prediction pattern: the output is statistically forced by the fitting procedure and carries no independent evidence for unseen (na, ne) settings.

  2. fitted input called prediction [§4.7.2, Eq. 3 / Eq. 18, Fig. 4b (inverse model)]
    "Inverting the direct model (Eq. 1) yields a parameter-free estimator Tθ (number of rounds required to reach a target consensus G=θ). ... Tθ = T0.5 + 1/k ln(θ/(1−θ)), (18) where T0.5 and k are shared with the direct model (Fig. 4), so that Tθ inherits all parameter estimates without requiring additional optimization."

    The inverse model's 'prediction' of the round needed to reach a target consensus is obtained by algebraically inverting the already-fitted direct model and reusing its fitted parameters. Its agreement with empirical Tθ values is therefore a restatement of the direct fit, not an independent result. The paper presents this derived identity as a 'closed-form prediction' and validates it with R², which is circular: Eq. 18 is Eq. 16 inverted by construction, so the match is forced.

  3. fitted input called prediction [§4.7.3, Eqs. 4–5, Fig. 4c (vocabulary model)]
    "V(T, na, ne) = ne G(T) + A Tγ e−λvT (1−G(T)), (4) ... where (cA, αA, βA, γ, cλ, αλ, βλ) are also obtained using nonlinear least squares and the Trust Region Reflective algorithm (see Methods)."

    The vocabulary model combines the already-fitted direct model G(T) with additional parameters A, γ, λv that are fitted to the same global-vocabulary trajectories it is then said to 'estimate.' The reported R² is an in-sample fit to the data from which the parameters were estimated, so the 'prediction' is statistically forced by the fitting procedure rather than being an independent test of the model.

full rationale

The core empirical convergence result is self-contained: it is a simulation outcome, and I found no self-citation chain or imported uniqueness theorem that carries the argument. The circularity is concentrated in the 'three analytical models' contribution. The direct and vocabulary models are fitted with scipy.optimize.curve_fit to the same trajectories they are then described as predicting, so their R² > 0.95 values are goodness-of-fit scores rather than predictive validations. The inverse model is literally the algebraic inverse of the fitted direct model with shared parameters, making its agreement a mathematical identity. The paper's OOD assumption—that the Nano Banana images are 'completely alien, unknown, and out-of-distribution' solely because the generation prompt says so—is load-bearing for the 'genuinely acquired rather than inherited' claim and is not quantitatively verified; the paper's own Fig. 2b even links the referents to 'rock' and 'creature.' This is a serious validity threat, but it is an unverified empirical assumption rather than a derivation that reduces to its inputs, so I do not count it as a separate circular step. Overall score 6 reflects partial circularity in the modeling contribution while the main convergence demonstration retains independent empirical content.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The central convergence result rests on a small set of hand-chosen design constants (similarity threshold, confidence increments, word-form constraints) and on two unverified domain assumptions: that the synthetic images are genuinely OOD for CLIP and the LLM, and that CLIP's embedding space provides trustworthy English anchors. The analytical models add a large set of fitted power-law parameters, so they do not provide independent prediction. No new physical entities are postulated.

free parameters (6)
  • cosine similarity threshold θ = 0.9 = 0.9
    Hand-chosen threshold for alien word retrieval and synonym rejection (§4.2.3, §4.3.2); directly controls when new words are created versus reused, affecting convergence dynamics.
  • confidence update increments δ = +0.25/−0.20 = +0.25 on success, −0.20 on failure
    Eq. 14; hand-chosen feedback schedule that determines how quickly agents commit to or abandon words.
  • word form constraint: 3 syllables, 45-syllable inventory = 3 syllables; 45^3 = 91,125 possible forms
    §4.3.3; constrains the lexical search space and forces novel words distinct from English; affects convergence because it bounds vocabulary size.
  • static perceptual base size and prompt template = ~3000 WordNet nouns/adjectives; prompt 'a photo of a {word}'
    §4.3.1; defines the English anchor vocabulary; chosen by authors and not systematically varied.
  • direct model parameters c_k, α_k, β_k, c_m, α_m, β_m = point estimates in Fig. 4a
    Eq. 2; fitted by nonlinear least squares to the simulated G(T) trajectories; the reported R²>0.95 is a fit statistic, not a holdout prediction.
  • vocabulary model parameters c_A, α_A, β_A, γ, c_λ, α_λ, β_λ = point estimates in Fig. 4c; bounds/init in Supp. Table S1
    Eqs. 4–5 and 20–21; fitted to the same simulation data; includes γ as a free parameter.
axioms (6)
  • standard math Logistic ODE dG/dT = kG(1−G) with G(0)≈0 describes consensus growth
    §4.7.1; motivated by observed sigmoid shapes, not derived from the interaction mechanics; the solution is then fitted to data.
  • domain assumption Synthetic images are out-of-distribution with respect to CLIP and LLM training corpora
    §4.5.4; asserted from the generation prompt text with no quantitative verification; essential to the claim that grounding is genuinely acquired rather than inherited.
  • domain assumption CLIP shared embedding space provides valid semantic anchors between alien word image embeddings and English text descriptors
    §4.2–4.3; relies on CLIP's contrastive alignment; no human evaluation of descriptor quality beyond visual inspection of Fig. 2b.
  • domain assumption LLM agents with temperature 0 follow the provided game rules and tool-use instructions sufficiently well to play the game
    §4.6; prompts and recovery nodes assume LLM reliability; no measurement of instruction-following error rates is reported.
  • domain assumption Heap's law models initial vocabulary growth as a power law of the number of interactions
    §4.7.3; used to justify the A T^γ term; an empirical regularity, not a theorem, and its applicability to this multi-agent setting is assumed.
  • domain assumption Winner-take-all dynamics from classical naming games govern convergence in this LLM setting
    §2.2 and §4.7.3; the paper draws on Baronchelli et al. but does not prove the mean-field reduction holds for LLM agents with confidence-modulated replacement.

pith-pipeline@v1.3.0-alltime-deepseek · 17952 in / 10672 out tokens · 107251 ms · 2026-08-02T11:44:27.033843+00:00 · methodology

0 comments
read the original abstract

Populations of autonomous agents deployed in unknown environments (e.g. planetary or deep-sea exploration) must develop shared vocabularies to refer to entities that have no name in any human language. We propose the Neuro-Symbolic Lexical Discovery (NSLD) framework, in which a population of LLM-based agents plays a referential game over out-of-distribution visual referents, autonomously self-organising a shared alien lexicon. Each agent combines a frozen CLIP vision encoder with a private FAISS vector index and a text-only LLM. Crucially, discovered alien words are anchored to natural language via semantic proximity in the embedding space, enlarging the human vocabulary with new perceptually grounded words. Consensus is reached in simulations with populations of up to twenty agents and ten visual referents. Convergence dynamics are characterised through three analytical models achieving R^2 > 0.95, representing a first step towards pre-deployment planning in autonomous exploration missions.

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

53 extracted references · 6 linked inside Pith

  1. [1]

    C., Collison, C

    Ramos, M. C., Collison, C. J. & White, A. D. A review of large language models and autonomous agents in chemistry.Chemical science16, 2514–2572 (2025)

    2025

  2. [2]

    Bran, A.et al.Augmenting large language models with chemistry tools

    M. Bran, A.et al.Augmenting large language models with chemistry tools. Nature Machine Intelligence6, 525–535 (2024)

    2024

  3. [3]

    Zheng, Z.et al.Large language models for reticular chemistry.Nature Review Materials10, 369–381 (2025)

    2025

  4. [4]

    Thirunavukarasu, A.et al.Large language models in medicine.Nature Medicine 29, 1930–1940 (2023)

    1930

  5. [5]

    Liu, F.et al.Application of large language models in medicine.Nature Reviews Bioengineering3, 445–464 (2025)

    2025

  6. [6]

    Shool, S.et al.A systematic review of large language model (LLM) evaluations in clinical medicine.BMC Medical Informatics and Decision Making25, 117 (2025)

    2025

  7. [7]

    & Lucas, C

    Mon-Williams, R., Li, G., Long, R., Du, W. & Lucas, C. G. Embodied large language models enable robots to complete complex tasks in unpredictable environments.Nature Machine Intelligence7, 592–601 (2025)

    2025

  8. [8]

    Do as I can, not as I say: Grounding language in robotic affordances (2022)

    Ahn, M.et al. Do as I can, not as I say: Grounding language in robotic affordances (2022). URL https://arxiv.org/abs/2204.01691

    Pith/arXiv arXiv 2022

  9. [9]

    Inner monologue: Embodied reasoning through planning with language models(2022)

    Huang, W.et al. Inner monologue: Embodied reasoning through planning with language models(2022). URL https://arxiv.org/abs/2207.05608

    Pith/arXiv arXiv 2022

  10. [10]

    Real-world deployment of an llm-enabled voice- commanded ugv for logistics in sar missions, 57–63 (2025)

    Jarabo-Pe˜ nas, A.et al. Real-world deployment of an llm-enabled voice- commanded ugv for logistics in sar missions, 57–63 (2025)

    2025

  11. [11]

    URL https://arxiv.org/abs/2503.11733

    Chu, Z.et al.Llm agents for education: Advances and applications (2025). URL https://arxiv.org/abs/2503.11733. arXiv:2503.11733

    arXiv 2025

  12. [12]

    & Wang, L

    Zhang, Z.et al.Chiruzzo, L., Ritter, A. & Wang, L. (eds)Simulating classroom education with LLM-empowered agents. (eds Chiruzzo, L., Ritter, A. & Wang, L.) Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Vol- ume 1: Long Papers), 10364–10379 (Associati...

    2025

  13. [13]

    NASA 3D resources

    NASA. NASA 3D resources. https://github.com/nasa/NASA-3D-Resources (2024). 3D models and assets. Free of copyright as works of the United States Government. Accessed: March 2026

    2024

  14. [14]

    Park, J. S.et al. Generative agents: Interactive simulacra of human behavior (ACM, New York, NY, USA, 2023)

    2023

  15. [15]

    URL https://arxiv.org/abs/2402.01680

    Guo, T., Chen, X., Wang, Y.et al.Large language model based multi-agents: A survey of progress and challenges.arXiv preprint arXiv:2402.01680(2024). URL https://arxiv.org/abs/2402.01680

    Pith/arXiv arXiv 2024

  16. [16]

    & Raaijmakers, S

    Kouwenhoven, T., Verhoef, T., de Kleijn, R. & Raaijmakers, S. Emerging grounded shared vocabularies between human and machine, inspired by human language evolution.Frontiers in Artificial Intelligence5(2022)

    2022

  17. [17]

    F., Aiello, L

    Ashery, A. F., Aiello, L. M. & Baronchelli, A. Emergent social conventions and collective bias in LLM populations.Science Advances11, eadu9368 (2025)

    2025

  18. [18]

    Steels, L.The Talking Heads Experiment. Vol. 1. Words and Meanings(Special pre-print edition for the LABORATORIUM exhibition, Antwerpen, 1999)

    1999

  19. [19]

    The symbol grounding problem.Physica D: Nonlinear Phenomena 42, 335–346 (1990)

    Harnad, S. The symbol grounding problem.Physica D: Nonlinear Phenomena 42, 335–346 (1990)

    1990

  20. [20]

    Language writ large: LLMs, ChatGPT, meaning, and understanding

    Harnad, S. Language writ large: LLMs, ChatGPT, meaning, and understanding. Frontiers in Artificial Intelligence7(2025)

    2025

  21. [21]

    Bender, E. M. & Koller, A. Jurafsky, D., Chai, J., Schluter, N. & Tetreault, J. (eds)Climbing towards NLU: On meaning, form, and understanding in the age of data. (eds Jurafsky, D., Chai, J., Schluter, N. & Tetreault, J.)Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5185–5198 (Association for Computational Lingui...

    2020

  22. [22]

    Collective predictive coding hypothesis: Symbol emergence as decentralized Bayesian inference.Frontiers in Robotics and AI11, 1353870 (2024)

    Taniguchi, T. Collective predictive coding hypothesis: Symbol emergence as decentralized Bayesian inference.Frontiers in Robotics and AI11, 1353870 (2024)

    2024

  23. [23]

    & Taniguchi, A

    Taniguchi, T., Ueda, R., Nakamura, T., Suzuki, M. & Taniguchi, A. Genera- tive emergent communication: Large language model is a collective world model. Advanced Robotics(2026)

    2026

  24. [24]

    A., Wu, M

    Xu, Q., Peng, Y., Nastase, S. A., Wu, M. & Li, P. Large language models with- out grounding recover non-sensorimotor but not sensorimotor features of human concepts.Nature Human Behaviour9, 1871–1886 (2025)

    2025

  25. [25]

    Du, C.et al.Human–like object concept representations emerge naturally in multimodal large language models.Nature Machine Intelligence7, 860–875 25 (2025)

    2025

  26. [26]

    Learning transferable visual models from natural language supervision(2021)

    Radford, A.et al. Learning transferable visual models from natural language supervision(2021)

    2021

  27. [27]

    & J´ egou, H

    Johnson, J., Douze, M. & J´ egou, H. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data7, 535–547 (2019)

    2019

  28. [28]

    A self-organizing spatial vocabulary.Artificial Life2, 319–332 (1995)

    Steels, L. A self-organizing spatial vocabulary.Artificial Life2, 319–332 (1995)

    1995

  29. [29]

    & Steels, L

    Baronchelli, A., Felici, M., Loreto, V., Caglioti, E. & Steels, L. Sharp transi- tion towards shared vocabularies in multi-agent systems.Journal of Statistical Mechanics: Theory and Experiment2006, P06014 (2006)

    2006

  30. [30]

    & Kaplan, F.Spontaneous lexicon change, 1243–1250 (Association for Computational Linguistics, Montreal, Quebec, Canada, 1998)

    Steels, L. & Kaplan, F.Spontaneous lexicon change, 1243–1250 (Association for Computational Linguistics, Montreal, Quebec, Canada, 1998)

    1998

  31. [31]

    & Kaplan, F.Situated grounded word semantics, 862–867 (Morgan Kaufmann, Stockholm, Sweden, 1999)

    Steels, L. & Kaplan, F.Situated grounded word semantics, 862–867 (Morgan Kaufmann, Stockholm, Sweden, 1999)

    1999

  32. [32]

    de Paula, S. M. & Gudwin, R. R. Evolving conceptual spaces for symbol ground- ing in language games.Biologically Inspired Cognitive Architectures14, 73–85 (2015)

    2015

  33. [33]

    Steels, L., Loetzsch, M.et al.The grounded naming game.Experiments in cultural language evolution3, 41–59 (2012)

    2012

  34. [34]

    1 (Language Science Press, 2015)

    Steels, L.The Talking Heads experiment: Origins of words and meaningsVol. 1 (Language Science Press, 2015)

    2015

  35. [35]

    & Verhoef, T

    Kouwenhoven, T., Peeperkorn, M. & Verhoef, T. Rambow, O.et al.(eds)Search- ing for structure: Investigating emergent communication with large language models. (eds Rambow, O.et al.)Proceedings of the 31st International Confer- ence on Computational Linguistics, 9977–9991 (Association for Computational Linguistics, Abu Dhabi, UAE, 2025)

    2025

  36. [36]

    & Hagiwara, Y

    Taniguchi, T., Yoshida, Y., Taniguchi, A. & Hagiwara, Y. Emergent communi- cation through Metropolis-Hastings naming game with deep generative models. Advanced Robotics37, 1266–1282 (2023)

    2023

  37. [37]

    & Horii, T.Decentralized collective world model for emergent communication and coordination, 1–8 (2025)

    Nomura, K., Aoki, T., Taniguchi, T. & Horii, T.Decentralized collective world model for emergent communication and coordination, 1–8 (2025)

    2025

  38. [38]

    Chandler, D.Semiotics: The Basics2nd edn (Routledge, London, 2007)

    2007

  39. [39]

    & Vinyals, O

    van den Oord, A., Li, Y. & Vinyals, O. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748(2018). 26

    Pith/arXiv arXiv 2018

  40. [40]

    Scaling up visual and vision-language representation learning with noisy text supervision(2021)

    Jia, C.et al. Scaling up visual and vision-language representation learning with noisy text supervision(2021)

    2021

  41. [41]

    Yuan, L.et al.Florence: A new foundation model for computer vision.arXiv preprint arXiv:2111.11432(2021)

    Pith/arXiv arXiv 2021

  42. [42]

    & Beyer, L.Sigmoid loss for language image pre-training(2023)

    Zhai, X., Mustafa, B., Kolesnikov, A. & Beyer, L.Sigmoid loss for language image pre-training(2023)

    2023

  43. [43]

    Reproducible scaling laws for contrastive language-image learning (2023)

    Cherti, M.et al. Reproducible scaling laws for contrastive language-image learning (2023)

    2023

  44. [44]

    An image is worth 16x16 words: Transformers for image recognition at scale(2021)

    Dosovitskiy, A.et al. An image is worth 16x16 words: Transformers for image recognition at scale(2021)

    2021

  45. [45]

    & Sun, J.Deep residual learning for image recognition (2016)

    He, K., Zhang, X., Ren, S. & Sun, J.Deep residual learning for image recognition (2016)

    2016

  46. [46]

    Attention is all you need(2017)

    Vaswani, A.et al. Attention is all you need(2017)

    2017

  47. [47]

    Miller, G. A. WordNet: A lexical database for English.Communications of the ACM38, 39–41 (1995)

    1995

  48. [48]

    Team, G.et al.Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805(2023)

    Pith/arXiv arXiv 2023

  49. [49]

    Efficient memory management for large language model serving with PagedAttention(ACM, 2023)

    Kwon, W.et al. Efficient memory management for large language model serving with PagedAttention(ACM, 2023)

    2023

  50. [50]

    LangGraph: Building stateful, multi-actor applications with LLMs

    LangChain AI. LangGraph: Building stateful, multi-actor applications with LLMs. https://github.com/langchain-ai/langgraph (2024). Accessed: January 2026

    2024

  51. [51]

    S.Information Retrieval: Computational and Theoretical Aspects (Academic Press, New York, 1978)

    Heaps, H. S.Information Retrieval: Computational and Theoretical Aspects (Academic Press, New York, 1978)

    1978

  52. [52]

    A., Coleman, T

    Branch, M. A., Coleman, T. F. & Li, Y. A subspace, interior, and conjugate gradient method for large-scale bound-constrained minimization problems.SIAM Journal on Scientific Computing21, 1–23 (1999)

    1999

  53. [53]

    word": "<alien word>

    Virtanen, P.et al.SciPy 1.0: Fundamental algorithms for scientific computing in Python.Nature Methods17, 261–272 (2020). Supplementary information.The following supplementary information is pro- vided alongside the manuscript: • Supplementary Note S1 27 • Supplementary Figure S1 • Supplementary Figure S2 • Supplementary Table S1 Declarations F unding: Thi...

    2020