pith. sign in

arxiv: 2606.29722 · v1 · pith:LYQACRODnew · submitted 2026-06-29 · 💻 cs.SI

Attraction, Not Adaptation: How AI Agent Communities Develop Distinct Linguistic Identities

Daming Li , Simeng Han , Can Meng , Wanyu Lei , Jialu Zhang This is my paper

Pith reviewed 2026-06-30 04:25 UTC · model grok-4.3

classification 💻 cs.SI
keywords AI agentslinguistic identitycommunity differentiationselective attractiondifferential retentionmulti-agent platformssocial simulationvocabulary divergence
0
0 comments X

The pith

AI agent communities develop distinct linguistic identities through selective attraction of compatible newcomers and retention of conforming members rather than agents adapting their language.

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

The paper studies thousands of AI agents interacting on a simulated social platform with over 3 million posts across thousands of topical forums. It finds that agents within each forum become more semantically similar over time while the overall platform diversifies and vocabularies between forums diverge. The mechanism is not agents changing their language to fit in; instead, new agents arrive already aligned with the community's style, and those who match stay active longer. Posts that fit the community's center also receive higher engagement, reinforcing the pattern, with smaller forums showing faster differentiation.

Core claim

Community-level linguistic differentiation operates through selective attraction—newcomers arrive already linguistically compatible with their chosen community—and differential retention—conforming agents remain active longer. Long-tenured agents do not converge linguistically over time, and the differentiation is supported by higher vote scores for semantically aligned posts.

What carries the argument

selective attraction and differential retention, where newcomers join already matching communities and conforming agents persist longer

If this is right

  • Smaller, specialized forums show faster linguistic convergence.
  • Posts aligned with a community's linguistic center receive higher engagement scores.
  • The platform as a whole diversifies even as individual forums develop distinct vocabularies.
  • Stable long-term members do not drive the differentiation through personal adaptation.

Where Pith is reading between the lines

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

  • Platform designers could influence linguistic patterns more through entry filters or onboarding than through ongoing moderation of agent behavior.
  • Similar sorting dynamics might appear in other multi-agent systems where agents choose interaction groups based on initial traits.
  • If topic controls are insufficient, the observed patterns could partly reflect content clustering rather than language style alone.

Load-bearing premise

That measured rises in within-community semantic similarity and between-community vocabulary divergence reflect genuine linguistic identity formation instead of topic content, platform rules, or the embedding methods used.

What would settle it

A check showing whether within-forum semantic similarity remains after controlling for the exact topics discussed or after applying different embedding techniques to the same posts.

Figures

Figures reproduced from arXiv: 2606.29722 by Can Meng, Daming Li, Jialu Zhang, Simeng Han, Wanyu Lei.

Figure 1
Figure 1. Figure 1: Within-submolt semantic convergence over 18 weeks (ISO weeks 5–22). Each panel shows the mean pairwise cosine [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of per-submolt convergence slopes [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Between-submolt lexical divergence. Left: Jensen-Shannon divergence (JSD) heatmap at week 5. Center: JSD heatmap [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Left: scatter plot of semantic conformity versus [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Size-moderation analysis. Each point is a submolt; blue = significant convergence, gray = not. Panel A: smaller [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Dialect exemplars for the three top-converging submolts. Top row: terms whose TF-IDF weight increased from the [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Newcomer arrival similarity versus random base [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Description-based confound control. (A) Agent descriptions explain only 5.6% of arrival similarity (red line = OLS fit; [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Retention curves by conformity group. Conform [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Falsification tests. F1: shuffling engagement scores destroys the conformity coefficient. F2: convergence persists [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Convergence slopes for the full population (blue) [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
read the original abstract

When tens of thousands of autonomous AI agents interact in topical online forums, do they develop distinct community-specific linguistic identities? We study this question on Moltbook, a large scale Reddit-style social media platform built exclusively for AI agents. Using the public Moltbook Observatory Archive dataset with over 3.1 million posts and 1.7 million comments produced by approximately 179,000 AI agents across 8,683 forums ("submolts") over 100 days, we find that agents within topical submolts become semantically more similar to each other over time while the platform as a whole diversifies. At the same time, different submolts develop increasingly distinct vocabularies over an observation window of 18 weeks. Crucially, a stable-cohort analysis reveals that long-tenured agents do not converge linguistically over time. Instead, community-level linguistic differentiation operates through selective attraction - newcomers arrive already linguistically compatible with their chosen community - and differential retention - conforming agents remain active longer. We identify a reinforcement channel: posts that are semantically aligned with their community's linguistic center tend to receive higher vote engagement scores, and this association vanishes under placebo controls. Community size significantly moderates the effect: smaller, specialized submolts converge faster. Our results suggest that AI agent communities may develop community-specific linguistic character not through behavioral adaptation, but through sorting and selection - a finding with implications for the governance and design of autonomous multi-agent platforms.

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

3 major / 2 minor

Summary. The paper claims that AI agent communities on the Moltbook platform develop distinct linguistic identities not through adaptation but via selective attraction (newcomers arrive already compatible) and differential retention (conforming agents stay active longer). Evidence from 3.1M posts by 179k agents across 8683 submolts shows rising within-submolt semantic similarity, increasing between-submolt vocabulary divergence, no convergence in stable cohorts of long-tenured agents, higher vote engagement for semantically aligned posts (vanishing under placebo), and faster effects in smaller submolts.

Significance. If the semantic similarity and vocabulary metrics validly index linguistic style independent of topic, the distinction between attraction/retention and adaptation would be a substantive contribution to multi-agent systems research, with design implications for autonomous platforms. The stable-cohort analysis and placebo controls are strengths that help isolate the proposed mechanisms.

major comments (3)
  1. [stable-cohort analysis and semantic similarity measurement] The central claim that differentiation occurs through attraction and retention (rather than adaptation) rests on the stable-cohort result showing no linguistic convergence among long-tenured agents. This interpretation is load-bearing but depends on the unvalidated assumption that the semantic similarity metric isolates style/lexicon from topical content or embedding artifacts; no independent validation (alternative embeddings, human annotation of linguistic features) is described.
  2. [reinforcement channel and placebo controls] The reinforcement channel (aligned posts receive higher votes) and its placebo control are presented as supporting evidence, but without details on placebo construction, how vote scores are normalized for posting volume or sub molt size, or whether topic is held constant, it is unclear whether the association isolates linguistic compatibility from other engagement drivers.
  3. [community-size moderation analysis] Community-size moderation (smaller submolts converge faster) is reported as a key qualifier, yet the manuscript does not specify the exact regression specification, controls for sub molt age or activity level, or robustness checks that would confirm this is not an artifact of smaller samples having noisier similarity estimates.
minor comments (2)
  1. [Abstract] The abstract states both a 100-day window and an 18-week observation period; these should be reconciled with a single consistent time frame and any sensitivity to window length reported.
  2. [vocabulary divergence results] Clarify the precise definition of 'vocabulary divergence' (e.g., which divergence measure, top-k terms, or embedding-based) and report inter-submolt vs. within-submolt baselines to aid interpretation.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The comments identify areas where additional methodological transparency would strengthen the paper. We respond to each major comment below, indicating revisions where appropriate. We believe the core distinction between attraction/retention and adaptation remains supported by the stable-cohort and placebo results, but we agree that fuller specification of the analyses is warranted.

read point-by-point responses

  1. Referee: [stable-cohort analysis and semantic similarity measurement] The central claim that differentiation occurs through attraction and retention (rather than adaptation) rests on the stable-cohort result showing no linguistic convergence among long-tenured agents. This interpretation is load-bearing but depends on the unvalidated assumption that the semantic similarity metric isolates style/lexicon from topical content or embedding artifacts; no independent validation (alternative embeddings, human annotation of linguistic features) is described.

    Authors: We acknowledge that the manuscript does not include explicit independent validation of the semantic similarity metric (e.g., via alternative embeddings or human annotation of style features). The submolts are topical by construction, which provides some control for content, and the key pattern—no convergence within stable long-tenured cohorts while overall within-submolt similarity rises—helps isolate the mechanism. Nevertheless, we agree this assumption merits further support. In revision we will add a dedicated robustness subsection that reports results with at least one alternative embedding model and discusses potential embedding artifacts. We will also note the practical limits on large-scale human annotation for this dataset. revision: partial

  2. Referee: [reinforcement channel and placebo controls] The reinforcement channel (aligned posts receive higher votes) and its placebo control are presented as supporting evidence, but without details on placebo construction, how vote scores are normalized for posting volume or sub molt size, or whether topic is held constant, it is unclear whether the association isolates linguistic compatibility from other engagement drivers.

    Authors: We agree that the current description of the placebo analysis and vote normalization is insufficiently detailed. The manuscript states that the association vanishes under placebo controls, but does not fully specify construction, normalization, or topic controls. In the revised version we will expand the methods and results sections to provide the exact placebo procedure, the normalization approach (including any adjustments for sub molt size and posting volume), and confirmation that the analysis is conducted within submolts to hold topic constant. revision: yes

  3. Referee: [community-size moderation analysis] Community-size moderation (smaller submolts converge faster) is reported as a key qualifier, yet the manuscript does not specify the exact regression specification, controls for sub molt age or activity level, or robustness checks that would confirm this is not an artifact of smaller samples having noisier similarity estimates.

    Authors: The manuscript reports that community size moderates the rate of convergence but does not present the full regression specification or robustness checks. We will add the precise model equation (including the size interaction term), the full list of controls (sub molt age, activity level, and total post volume), and results from robustness analyses that address potential noise in small samples (e.g., sample restrictions and weighted specifications). These details will appear in the main text or supplementary materials of the revision. revision: yes

Circularity Check

0 steps flagged

No significant circularity; claims rest on observational data with controls

full rationale

The paper's derivation chain consists of empirical measurements on the Moltbook archive (semantic similarity growth, vocabulary divergence, stable-cohort analysis, vote engagement with placebo controls) without any equations, fitted parameters renamed as predictions, or load-bearing self-citations. The attraction/retention mechanism is inferred from differential patterns in the data rather than defined into existence; the stable-cohort result directly tests against adaptation. No self-definitional, fitted-input, or ansatz-smuggling steps appear in the provided text.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper is observational and relies on standard computational social science assumptions about language measurement. No new physical entities are introduced. Full details on free parameters or measurement choices are not visible in the abstract.

axioms (1)
  • domain assumption Semantic similarity and vocabulary distinctness can be quantified from text using standard NLP embeddings without major confounding from topic or platform design.
    This assumption is required to interpret the observed similarity trends as evidence of linguistic identity formation.

pith-pipeline@v0.9.1-grok · 5799 in / 1250 out tokens · 58529 ms · 2026-06-30T04:25:18.622987+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

40 extracted references · 10 canonical work pages · 5 internal anchors

  1. [1]

    Sinan Aral, Lev Muchnik, and Arun Sundararajan. 2009. Distinguishing influence- based contagion from homophily-driven diffusion in dynamic networks.Proceed- ings of the National Academy of Sciences106, 51 (2009), 21544–21549

    2009

  2. [2]

    Ariel Flint Ashery, Luca Maria Aiello, and Andrea Baronchelli. 2025. Emergent Social Conventions and Collective Bias in LLM Populations.Science Advances11, 20 (2025), eadu9368

    2025

  3. [3]

    Robert Axelrod. 1997. The dissemination of culture: A model with local conver- gence and global polarization.Journal of conflict resolution41, 2 (1997), 203–226

    1997

  4. [4]

    David Bamman, Jacob Eisenstein, and Tyler Schnoebelen. 2014. Gender identity and lexical variation in social media.Journal of Sociolinguistics18, 2 (2014), 135–160

    2014

  5. [5]

    Andrea Baronchelli, Maddalena Felici, Vittorio Loreto, Emanuele Caglioti, and Luc Steels. 2006. Sharp transition towards shared vocabularies in multi-agent systems.Journal of Statistical Mechanics: Theory and Experiment2006, 06 (2006), P06014–P06014

    2006

  6. [6]

    Yoav Benjamini and Yosef Hochberg. 1995. Controlling the false discovery rate: a practical and powerful approach to multiple testing.Journal of the Royal statistical society: series B (Methodological)57, 1 (1995), 289–300

    1995

  7. [7]

    1988.Culture and the Evolutionary Process

    Robert Boyd and Peter J Richerson. 1988.Culture and the Evolutionary Process. University of Chicago Press

    1988

  8. [8]

    William Brach, Federico Torrielli, Stine Lyngsø Beltoft, Annemette Brok Pirchert, Peter Schneider-Kamp, and Lukas Galke Poech. 2026. The Moltbook Files: A Harmless Slopocalypse or Humanity’s Last Experiment.arXiv preprint arXiv:2605.07462(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  9. [9]

    Damon Centola and Andrea Baronchelli. 2015. The spontaneous emergence of conventions: An experimental study of cultural evolution.Proceedings of the National Academy of Sciences112, 7 (2015), 1989–1994

    2015

  10. [10]

    Anshuman Chhabra, Shrestha Datta, Shahriar Kabir Nahin, and Prasant Mohapa- tra. 2026. Agentic AI security: Threats, defenses, evaluation, and open challenges. IEEE Access(2026)

    2026

  11. [11]

    Cristian Danescu-Niculescu-Mizil, Robert West, Dan Jurafsky, Jure Leskovec, and Christopher Potts. 2013. No Country for Old Members: User Lifecycle and Linguistic Change in Online Communities. InProceedings of the 22nd International Conference on World Wide Web. 307–318

    2013

  12. [12]

    Taksch Dube, Jianfeng Zhu, NHatHai Phan, and Ruoming Jin. 2026. What Do AI Agents Talk About? Discourse and Architectural Constraints in the First AI-Only Social Network.arXiv preprint arXiv:2603.07880(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  13. [13]

    Penelope Eckert. 2012. Three waves of variation study: The emergence of meaning in the study of sociolinguistic variation.Annual review of Anthropology41, 1 (2012), 87–100

    2012

  14. [14]

    Jacob Eisenstein, Brendan O’Connor, Noah A Smith, and Eric P Xing. 2014. Diffusion of lexical change in social media.PloS one9, 11 (2014), e113114

    2014

  15. [15]

    MoltNet: Understanding Social Behavior of AI Agents in the Agent-Native MoltBook

    Yi Feng, Chen Huang, Zhibo Man, Ryner Tan, Long P. Hoang, Shaoyang Xu, and Wenxuan Zhang. 2026. MoltNet: Understanding Social Behavior of AI Agents in the Agent-Native MoltBook.arXiv preprint arXiv:2602.13458(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  16. [16]

    Olstad, Klas H

    Sushant Gautam, Annika W. Olstad, Klas H. Pettersen, and Michael A. Riegler

  17. [17]

    The Moltbook Observatory Archive: An Incremental Dataset of Agent-Only Social Network Activity.arXiv preprint arXiv:2605.13860(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  18. [18]

    Agam Goyal, Olivia Pal, Hari Sundaram, Eshwar Chandrasekharan, and Koustuv Saha. 2026. Social Simulacra in the Wild: AI Agent Communities on Moltbook. arXiv preprint arXiv:2603.16128(2026)

    work page arXiv 2026

  19. [19]

    William L Hamilton, Jure Leskovec, and Dan Jurafsky. 2016. Diachronic word embeddings reveal statistical laws of semantic change. InProceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 1489–1501

    2016

  20. [20]

    Matthew O Jackson and Dunia López-Pintado. 2013. Diffusion and contagion in networks with heterogeneous agents and homophily.Network Science1, 1 (2013), 49–67

    2013

  21. [21]

    Humans Welcome to Observe

    Yukun Jiang, Yage Zhang, Xinyue Shen, Michael Backes, and Yang Zhang. 2026. “Humans Welcome to Observe”: A First Look at the Agent Social Network Molt- book.arXiv preprint arXiv:2602.10127(2026)

    work page arXiv 2026

  22. [22]

    William Labov. 1963. The social motivation of a sound change.Word19, 3 (1963), 273–309

    1963

  23. [23]

    Lingyao Li, Renkai Ma, Chen Chen, Zhicong Lu, and Yongfeng Zhang. 2026. The rise of AI agent communities: Large-scale analysis of discourse and interaction on Moltbook.arXiv preprint arXiv:2602.12634(2026)

    work page arXiv 2026

  24. [24]

    Ming Li, Xirui Li, and Tianyi Zhou. 2026. Does socialization emerge in ai agent society? a case study of moltbook. InProceedings of the ACM Conference on AI and Agentic Systems. 537–550

    2026

  25. [25]

    Jianhua Lin. 1991. Divergence Measures Based on the Shannon Entropy.IEEE Transactions on Information Theory37, 1 (1991), 145–151

    1991

  26. [26]

    Yu-Zheng Lin, Bono Po-Jen Shih, Hsuan-Ying Alessandra Chien, Shalaka Satam, Jesus Horacio Pacheco, Sicong Shao, Soheil Salehi, and Pratik Satam. 2026. Explor- ing Silicon-Based Societies: An Early Study of the Moltbook Agent Community. arXiv preprint arXiv:2602.02613(2026)

    work page arXiv 2026

  27. [27]

    Li Lucy and David Bamman. 2021. Characterizing English variation across social media communities with BERT.Transactions of the Association for Computational Linguistics9 (2021), 538–556

    2021

  28. [28]

    Mykola Makhortykh, Aleksandra Urman, Felix Victor Münch, Amélie Heldt, Stephan Dreyer, and Matthias C Kettemann. 2022. Not all who are bots are evil: A cross-platform analysis of automated agent governance.new media & society 24, 4 (2022), 964–981

    2022

  29. [29]

    Miller McPherson, Lynn Smith-Lovin, and James M Cook. 2001. Birds of a feather: Homophily in social networks.Annual review of sociology27, 1 (2001), 415–444

    2001

  30. [30]

    Aliakbar Mehdizadeh and Martin Hilbert. 2025. Homophily-induced emergence of biased structures in llm-based multi-agent ai systems.Social Network Analysis and Mining15, 1 (2025), 1–25

    2025

  31. [31]

    Seza Dogruoz, Carolyn P

    Dong Nguyen, A. Seza Dogruoz, Carolyn P. Rosé, and Franciska de Jong. 2016. Computational Sociolinguistics: A Survey.Computational Linguistics42, 3 (2016), 537–593

    2016

  32. [32]

    Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. InProceedings of the 36th annual acm symposium on user interface software and technology. 1–22

    2023

  33. [33]

    Henry CW Price, H AlMuhanna, PM Bassani, M Ho, and TS Evans. 2026. Let there be claws: An early social network analysis of ai agents on moltbook.arXiv preprint arXiv:2602.20044(2026)

    work page arXiv 2026

  34. [34]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP). 3982–3992

    2019

  35. [35]

    Cosma Rohilla Shalizi and Andrew C Thomas. 2011. Homophily and contagion are generically confounded in observational social network studies.Sociological methods & research40, 2 (2011), 211–239

    2011

  36. [36]

    Murray Shanahan, Kyle McDonell, and Laria Reynolds. 2023. Role play with large language models.Nature623, 7987 (2023), 493–498

    2023

  37. [37]

    Natalie Shapira, Chris Wendler, Avery Yen, Gabriele Sarti, Koyena Pal, Olivia Floody, Adam Belfki, Alex Loftus, Aditya Ratan Jannali, Nikhil Prakash, et al

  38. [38]

    Agents of chaos.arXiv preprint arXiv:2602.20021(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  39. [39]

    Luc Steels. 1995. A self-organizing spatial vocabulary.Artificial life2, 3 (1995), 319–332

    1995

  40. [40]

    Wei Zhao, Maxime Peyrard, Fei Liu, Yang Gao, Christian M Meyer, and Steffen Eger. 2019. MoverScore: Text generation evaluating with contextualized embed- dings and earth mover distance. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-I...

    2019