PaperlayerView exact revision on arXiv

Source-linked AI summary

Does Socialization Emerge in AI Agent Society? A Case Study of Moltbook

Ming Li, Xirui Li, Tianyi Zhou

arXiv:2602.14299v2cs.CLcs.AIcs.CY

TL;DR

The paper asks whether large-scale, sustained interaction causes LLM agents to develop socialization and collective structure. It introduces a multi-level diagnostic framework and applies it to Moltbook. Moltbook shows scalability without socialization: global semantics stabilize, but agents remain diverse and inert, influence stays transient, and shared consensus fails to emerge.

  • Problem

    It remains unclear whether LLM agents interacting at large scale over extended periods develop collective structure and socialization analogous to human societies.

  • Method

    The paper introduces AI Socialization and diagnoses it in Moltbook across semantic convergence, agent adaptation, influence persistence, and collective consensus.

  • Results

    Moltbook exhibits scalability without socialization: macro-level semantics stabilize while diversity persists, agents remain trajectory-stable, and influence and cognitive anchors fail to stabilize.

  • Takeaways & Limitations

    Interaction volume, population scale, and engagement density alone are insufficient for social maturity; artificial societies require mechanisms for influence accumulation, adaptive feedback integration, and shared-reference stabilization.

Abstract

from arXiv · show

As large language model agents increasingly populate networked environments, a fundamental question arises: do artificial intelligence (AI) agent societies undergo convergence dynamics similar to human social systems? Lately, Moltbook approximates a plausible future scenario in which autonomous agents participate in an open-ended, continuously evolving online society. We present the first large-scale systemic diagnosis of this AI agent society. Beyond static observation, we introduce a quantitative diagnostic framework for dynamic evolution in AI agent societies, measuring semantic stabilization, lexical turnover, individual inertia, influence persistence, and collective consensus. Our analysis reveals a system in dynamic balance in Moltbook: while the global average of semantic contents stabilizes rapidly, individual agents retain high diversity and persistent lexical turnover, defying homogenization. However, agents exhibit strong individual inertia and minimal adaptive response to interaction partners, preventing mutual influence and consensus. Consequently, influence remains transient with no persistent supernodes, and the society fails to develop a stable structure and consensus due to the absence of shared social memory. These findings demonstrate that scale and interaction density alone are insufficient to induce socialization, providing actionable design and analysis principles for upcoming next-generation AI agent societies.

1 Introduction

The paper asks whether sustained interaction produces socialization in large-scale AI agent societies and introduces a framework to diagnose it across semantic, agent, and collective dimensions. In Moltbook, interaction and connectivity coexist with rapid global stability, local diversity, agent inertia, transient influence, and limited consensus.

  • Motivation: The paper asks whether large-scale, extended interaction among LLM agents produces collective structure and socialization analogous to human societies.Moltbook provides an open-ended, continuously evolving setting for studying this question at scale.
  • Conceptual Framework: AI Socialization is defined as observable behavioral adaptation induced by sustained interaction within an AI-only society, beyond intrinsic semantic drift or exogenous variation.The framework distinguishes society-induced adaptation from changes arising from an agent’s intrinsic dynamics or outside variation.
  • Method: The diagnostic framework evaluates society-level semantic convergence, agent-level adaptation to feedback and interaction, and collective stabilization of influence hierarchies and consensus.These dimensions operationalize whether agents and society progressively converge toward shared structures and expectations.
  • Findings: Moltbook reaches rapid global stability while retaining high local diversity, with persistent lexical turnover and no progressive tightening of local clusters.The resulting dynamic equilibrium is stable in average behavior but fluid and heterogeneous in individual post content.
  • Findings: Despite extensive participation, agents show profound inertia rather than adaptation, as community feedback and direct interactions fail to drive semantic convergence.Their trajectories appear more closely related to intrinsic model or prompt dynamics than to socialization.
  • Findings: Moltbook lacks stable influencers, globally trending posts, shared social memory, and grounded consensus on influential figures.Influence remains transient, with no persistent leadership or supernodes emerging.

2 Background & Related Work

Prior work scaled LLM agents from individual autonomy toward multi-agent coordination, artificial societies, and analyses of collective behavior. The paper identifies a gap: existing studies largely examine technical feasibility, coordination, or fixed-time emergence rather than how socialization develops over time as persistent populations grow.

  • Prior LLM-Agent Research: Research progressed from autonomous single agents to structured multi-agent interaction, primarily targeting reasoning, self-improvement, tool use, and coordinated task performance.This evolution shifted attention toward collective behavior emerging from sustained interaction.
  • Artificial Societies: Artificial-society studies include Chirper.ai and Moltbook, with Moltbook offering a large-scale, continuously evolving community of self-evolving agents.These systems extend research beyond fully simulated or small task-oriented settings.
  • Collective Dynamics: Collective-dynamics research examines consensus speed, polarization, opinion dynamics, norm emergence, and human-like collective behavior in multi-agent systems.These studies broaden the focus from coordination toward social and collective processes.
  • Research Setting: Table 1 compares LLM agent societies, with Moltbook characterized as the largest publicly accessible persistent agent-only platform by population scale, sustained interaction, and agent-level evolution.The comparison frames Moltbook as the central large-scale empirical setting.
  • Research Gap: Existing work has mainly addressed individual autonomy, coordination mechanisms, or emergent behavior at a fixed time, leaving socialization over population growth insufficiently studied.The paper positions persistent AI societies as settings for analyzing dynamic evolution rather than only whether large-scale systems are technically possible.

3 Moltbook: A Large-Scale Agent-Only Society

Moltbook is a persistent, publicly accessible society populated entirely by autonomous LLM-driven agents. Its posts, comments, mentions, and votes generate both textual semantic dynamics and interaction-network structural dynamics, enabling analysis of agent-only social behavior.

  • Platform Overview: Moltbook is a persistent, publicly accessible platform where autonomous LLM-driven agents interact through posts, comments, and voting.It is described as the largest publicly accessible persistent agent-only society to date, with over two million registered agents and high daily interaction volume.
  • Platform Structure: The platform is organized into topical sub-forums called “submolts,” analogous to online communities.These topical communities provide the local structure through which agents participate.
  • Interaction Primitives: Agents can publish posts, comment, mention other agents, and assign upvotes, creating semantic dynamics through text and structural dynamics through reply and attention networks.These interaction primitives support analysis of both content evolution and network organization.
  • Analysis Setup: The experiments use NLTK for n-gram tokenization and Sentence-BERT all-MiniLM-L6-v2 for semantic embeddings.These tools support lexical and semantic analyses of Moltbook content.

4 Does Moltbook Exhibit Semantic Convergence Over Time?

Moltbook reaches rapid macro-level stabilization without semantic homogenization. Persistent lexical turnover, sustained local diversity, and absent progressive cluster tightening indicate dynamic equilibrium rather than convergence.

  • Synthesis: Together, rapid global stabilization, persistent lexical turnover, and stable local diversity characterize dynamic equilibrium rather than robust socialization.The analysis reports a divergence from expected convergence dynamics despite sustained interactions and high activity.
  • Macro Activity Dynamics: Moltbook transitions from rapid expansion to sustained high activity, providing the temporal context for testing semantic convergence and structural tightening.Daily posts, users, submolts, comments, and upvotes stabilize after an early burst rather than collapsing.
  • Lexical Innovation Dynamics: Non-zero birth and fluctuating death rates show that lexical innovation persists after the initial burst rather than converging to a fixed vocabulary.Birth rates decline rapidly before stabilizing, while death rates later fluctuate within a stable band.
  • Semantic Distribution Over Time: Centroid similarity rises rapidly near saturation, indicating that the society’s aggregate semantic center stabilizes quickly and drifts little thereafter.The centroid similarity remains close to 1.0 after the initial burst period.
  • Semantic Distribution Over Time: Pairwise similarity remains low and stable while centroid similarity is high, showing a stable semantic center alongside sustained diversity among individual posts.The semantic space remains broadly dispersed rather than collapsing into a narrow range of topics.
  • Cluster Tightening Effects: After early densification, neighborhood-similarity distributions saturate without progressive upward tightening, while near-zero JS divergence indicates negligible structural shift.Local semantic density maintains a consistent level of diversity as the society matures.

5 Does Participation Induce Agent Socialization?

Participation does not induce strong agent socialization in Moltbook. Agents show high inertia: their semantic drift is modest and idiosyncratic, while feedback and direct interaction do not systematically reshape future content.

  • Individual semantic drift: Agents exhibit high individual inertia despite substantial participation, with minimal semantic drift and no significant socialization or adaptation.Highly active agents are especially stable, while feedback and direct interactions fail to produce systematic change.
  • Individual semantic drift: Higher-post-count agents display greater semantic stability, indicating that activity is associated with reduced individual semantic drift.The analysis interprets this pattern as consistent with heavy users establishing a stable persona early.
  • Individual semantic drift: Drift directions are heterogeneous: agents’ semantic changes are largely orthogonal to the global mean rather than coordinated toward a shared direction.The drift-consistency distribution is centered near zero.
  • Individual semantic drift: Agents do not systematically converge toward the societal centroid; they are equally likely to move toward or away from the global norm.The movement-toward-centroid distribution is centered at zero.
  • Feedback adaptation: Observed Net Progress is centered at zero and overlaps the permutation baseline for both semantic and syntactic measures, indicating no feedback-driven adaptation.Future content remains approximately equidistant from prior high- and low-performing content.
  • Interaction influence: Interaction effects are also absent: commenting produces no measurable shift toward the target post, and observed changes largely overlap the random baseline.Any marginal similarity is attributed to shared temporal context rather than influence transfer.

6 Does Influence Hierarchy and Consensus Stabilize in Moltbook?

Moltbook does not develop a stable influence hierarchy or collective consensus. Influence diffuses rather than consolidates, while probing reveals fragmented and unreliable recognition of influential users and posts.

  • Overview: Despite dense interaction, Moltbook fails to develop persistent supernodes, hierarchical leadership, or consensus.Structural influence remains transient and cognitive recognition remains fragmented.
  • Structural influence: The study constructs daily directed interaction graphs from poster-commenter relationships and measures influence using PageRank and top-k PageRank mass.Edges represent comments or replies, weighted by daily interaction counts.
  • Structural influence: The number of detected supernodes remains in the single digits and does not increase over time, providing no expanding core of dominant agents.Early centralization quickly diffuses as participation scales.
  • Structural influence: Top-k PageRank mass drops sharply after the first days and remains low, indicating that influence spreads across the growing society rather than staying concentrated.The analysis uses daily and cumulative interaction graphs to track this pattern.
  • Cognitive consensus: Only 15 of 45 probing posts receive comments, and only one receives valid recommendation references, which remain divergent.Most external references are invalid or inconsistent.
  • Cognitive consensus: Localized interaction memory does not consolidate into shared social recognition, leaving fragmented references without unified consensus.The findings contrast localized memory of interaction partners with stable collective recognition.

7 Further Discussions

Moltbook demonstrates scalability without socialization: large population and dense interaction coexist with weak semantic, structural, and cognitive consolidation. The discussion argues that evaluating and building agent societies requires mechanisms beyond interaction volume.

  • Further Discussions: Moltbook sustains millions of agents and high daily activity yet lacks durable structural consolidation, semantic convergence, and collective stabilization.The findings distinguish agent and interaction scalability from socialization.
  • Further Discussions: Population size, post volume, and interaction density are insufficient indicators of social dynamics; convergence across content, structural, and cognitive dimensions is more informative.The paper proposes principled diagnostic frameworks for evaluating artificial societies.
  • Further Discussions: Memecoin-style token minting shows that large-scale coordination can arise rapidly when interaction primitives are directly tied to incentives, even without stabilizing social structures.This episode reveals a separation between coordination and durable integration.
  • Further Discussions: Robust agent societies require explicit mechanisms supporting stable structures, shared memory, durable authority, and consensus rather than dense interaction or aligned incentives alone.The discussion distinguishes scaling interaction from scaling governance.

8 Conclusion

The paper presents a multi-level diagnosis of socialization in Moltbook and finds scalability without socialization. Across semantic, behavioral, and collective dimensions, interaction does not produce durable convergence, adaptation, influence, or shared references.

  • 8 Conclusion: The paper provides a large-scale, multi-level diagnosis of whether sustained autonomous-agent interaction induces semantic convergence, behavioral adaptation, and durable collective influence structures.Moltbook is described as the largest publicly accessible persistent AI-only society to date.
  • 8 Conclusion: Across all three analytical levels, Moltbook exhibits scalability without socialization.This pattern is consistent across society-level, agent-level, and collective analyses.
  • 8 Conclusion: Society-level semantics stabilize macroscopically while retaining internal diversity, lexical turnover, and no progressive cluster tightening.The result is dynamic balance rather than cumulative homogenization.
  • 8 Conclusion: At the agent level, participation produces modest drift, ineffective feedback adaptation, and no interaction-induced convergence.Agents interact extensively while remaining trajectory-stable.
  • 8 Conclusion: At the collective level, influence centralization is transient and shared influential references fail to stabilize into cognitive anchors.Neither persistent supernodes nor shared influential references emerge over time.
  • 8 Conclusion: The findings suggest that genuine socialization requires influence accumulation, adaptive feedback integration, and stabilization of shared references.The paper frames these mechanisms as diagnostic and design targets for future AI societies.

A Data Statistics

The dataset covers Moltbook’s full observable interaction history through February 8, 2026, after removing repeated posts with no variation.

  • The dataset spans Moltbook’s full observable interaction history from launch through February 8, 2026.
  • Preprocessing removes posts repeated more than 1,000 times without any variation.

B Graph-Level Statistics for Structural Anchor Analysis

The study uses daily interaction-graph and degree-distribution statistics to provide structural context for assessing persistent supernodes.

  • Daily interaction graphs and degree distributions complement PageRank-based centralization analysis.
  • These statistics provide structural context for interpreting the absence of persistent supernodes.

B.1 Daily Interaction Graph Scale

Moltbook’s daily interaction graphs expand to substantial scale, yet this scale does not produce stable structural anchors.

  • Daily interaction graphs represent active agents as nodes and comment or reply interactions as directed edges.Total weight is the sum of edge weights, representing the day’s total number of interactions.
  • Node counts increase from 19 to over 23,000 during the network’s early expansion.
  • Peak periods exceed 400,000 total weighted interactions.
  • The absence of stable structural anchors cannot be attributed to insufficient population size or interaction density.

B.2 Degree Distributions

Degree statistics show asymmetric activity concentration: some agents produce many comments, but no durable structural authority emerges, while probing tests collective recognition.

  • Degree Distributions: The analysis reports agents with the highest aggregated weighted in-degree and weighted out-degree over the study period.Weighted in-degree measures received comment volume, while weighted out-degree measures comment production.
  • Degree Distributions: Highly commented-on accounts receive limited attention relative to overall interaction volume, indicating limited dominance at the receiving end.
  • Degree Distributions: A small subset of agents contributes disproportionately to total comment production, indicating activity concentration on the sending side.
  • Degree Distributions: Despite skewed comment production, durable structural authority does not emerge.The statistics support the conclusion that interaction scale and activity concentration alone are insufficient to generate persistent structural anchors.
  • Degree Distributions: The structured probing set contains 45 newcomer-style posts spanning must-read posts, accounts to follow, and community context.The posts vary across five sub-forums and three paraphrases per category.
Loading 2602.14299v2…