PaperlayerView exact revision on arXiv

Source-linked AI summary

Self-Emergence Agent Architecture:Behavior-Inertia HMM, Reflexive Metacognition,and Social-Contrastive Self-Modeling

Xiaoyang Liu

arXiv:2609.17331v1cs.AI

TL;DR

LLM agents lack stable behavioral inertia and reflection that changes their dispositions. SEAA combines an HMM, parameter-updating reflection, and social interaction; its prototype reports distinct, stable personalities emerging from identical starts while controls remain homogeneous. The paper studies observable behavioral emergence, not subjective experience.

  • Problem

    Current LLM agents lack persistent behavioral inertia, and mainstream text-only reflection does not modify underlying dispositions.

    Source: §1 · p. 2

  • Method

    SEAA links social interaction and verbal reflection to runtime updates of an HMM transition matrix that represents behavioral inertia.

    Source: §4.1 · p. 5, §4.2 · p. 6

  • Results

    Pairwise transition-matrix Frobenius distance rose from 0.03 to 1.91 ± 0.29 while the control stayed at 0, and identical agents consolidated distinct, stable personalities.

    Source: §5.7.2 · p. 10, §7 · p. 17

  • Takeaways & Limitations

    SEAA operationalizes self as observable behavioral stabilization, reflection-driven evolution, and differentiation through social comparison.

    Source: §7 · pp. 17–18

    LLM experiments use one hosted model family and one deliberation topic, so leader/outlier topology may reflect that model's stylistic biases.

    Source: §6.3 · p. 17

Abstract

from arXiv · show

Large language model (LLM) agents exhibit strong language-generation and problem-solving capabilities, yet suffer from three structural limitations: personality drift, non-evolutionary reflection, and the absence of a self-other boundary. Existing generative-agent simulations rely on static memory and fixed prompts, maintaining neither behavioral inertia nor endogenous self-evolution. We propose the Self-Emergence Agent Architecture (SEAA), which integrates three components: (i) a Hidden Markov Model (HMM) that encodes long-term behavioral and cognitive inertia as an editable state-transition matrix; (ii) a Reflexion-style verbal metacognition loop whose output updates the HMM parameters themselves, rather than merely being stored as text; and (iii) a multi-agent social environment in which initially identical agents continuously compare their behavior with others'. The three components form a closed loop: social action $\to$ feedback $\to$ self-reflection $\to$ inertia update $\to$ differentiated action. We state three falsifiable hypotheses and provide a reproducible experimental protocol with operational metrics. A language-model-free prototype shows the loop spontaneously breaks symmetry: initially identical agents consolidate distinct, stable personalities whereas matched controls do not. Experiments with a hosted LLM surface these differences as distinct first-person self-narratives, and a five-agent deliberation spontaneously develops social structure---a consensus hub and a unanimously rejected outlier---absent in the control. Following an epistemologically agnostic stance inspired by Zhuangzi, SEAA studies only observable behavioral emergence and makes no claim about subjective qualia. This work contributes a unified framework, a concrete architecture with pseudocode, mechanistic evidence, and a microscope-style sandbox for studying artificial-self emergence.

1 Introduction

Current LLM agents lack stable behavioral inertia, reflection that changes underlying dispositions, and an endogenous self–other boundary. SEAA combines editable inertia, parameter-updating reflection, and social interaction to study observable self-emergence without claims about qualia.

1 Introduction

  • Agents drift without persistent internal state, and text-only reflection stores lessons without changing underlying behavioral dispositions.Existing multi-agent simulations also lack a mechanism for agents to distinguish themselves from others through social comparison.Source: §1 · p. 2

1.1 Motivation and philosophical stance

  • SEAA treats selfhood as observable behavioral emergence and makes no claim about subjective qualia.Its Zhuangzian stance limits claims to measurable behavior.Source: §1.1 · p. 2

1.2 Contributions

  • SEAA encodes inertia in an editable HMM transition matrix, uses verbal metacognition to update it, and lets social interaction support self–other differentiation.These are the paper's three stated contributions.Source: §1.2 · p. 2

1.3 Organization

  • The paper reviews related work, states hypotheses, details the architecture and protocol, discusses scope and philosophy, and provides reproducibility materials in appendices.The appendices include prompts, transcripts, and reproducibility details.Source: §1.3 · pp. 2–3

2 Related Work

Prior work addresses reflection, emergent individuality, behavioral models, and evolving personas, but the paper identifies no prior system integrating them into a closed loop that edits inertia through reflection and social interaction.

2.1 Self-reflective and self-evolving LLM agents

  • Reflexion [1] reuses natural-language reflections without gradient updates, while SEAA updates a personality carrier rather than a skill, memory, or policy.The paper describes this as orthogonal to task-performance-oriented self-evolving agents.Source: §2.1 · p. 3

2.2 Generative multi-agent simulation and emergent individuality

  • Emergent-individuality studies report differentiation [4] [5] [6] [7], but their agents lack editable inertia and a reflection-driven self-evolution loop.The paper says this leaves the differentiation mechanism implicit.Source: §2.2 · p. 3

2.3 Markov models of behavior, affect, and personality

  • Unlike MECoT, which uses a fixed personality-weighted Markov chain, SEAA edits its HMM transition matrix at runtime through natural-language metacognition.The paper states that no prior work, to its knowledge, uses LLM verbal self-reflection to update HMM inertia for self-emergence.Source: §2.3 · p. 3

2.4 Evolving-persona and lifelong agents

  • Generative Life Agents and AutoPersonas [11] edit traits or identity documents, but the paper describes them as single-agent systems without endogenous social self–other boundaries.Their edited objects differ from SEAA's stochastic behavioral-inertia model.Source: §2.4 · p. 4

2.5 Theories of consciousness and machine self-models

  • Consciousness theories differ on substrate, and the paper says none integrates inertia, self-reflection, and social contrast into one behavioral-development loop.It mentions Global Workspace Theory, Predictive Processing, Butlin et al. [15], Metzinger, and Seth.Source: §2.5 · p. 4

2.6 Research gap

  • SEAA targets a gap: prior work treats inertia, reflection, persona evolution, and social interaction separately, without the full loop from social behavior to self-emergence.The paper positions this as the specific behavioral-level gap in existing theories and agent systems.Source: §2.6 · p. 4, §2.5 · p. 4

3 Core Theoretical Hypotheses

The paper proposes three testable hypotheses about behavioral differentiation, reflection-driven evolution, and self–other boundaries, while defining self-concept only through observable behavior.

Core definitions

  • Behavioral inertia is dependence of current cognitive state on prior states, represented by an HMM transition operator P_t.A stable P_t corresponds to consistent disposition; differentiation across agents constitutes individuality.Source: §3 · p. 5

Hypothesis 1 (Individual differentiation)

  • Initially identical agents sharing a social environment are hypothesized to diverge in HMM transition matrices as experiences differ, with behavioral similarity decreasing to a floor.This is Hypothesis 1, not an established result.Source: §3 · p. 5

Hypothesis 2 (Reflection-driven evolution)

  • Agents whose reflections update HMM parameters are hypothesized to adapt faster and develop more stable personalities than matched text-only reflection agents.This is Hypothesis 2, not an established result.Source: §3 · p. 5

Hypothesis 3 (Social-contrastive boundary)

  • Socially embedded agents are hypothesized to articulate self–other distinctions over time, unlike isolated controls; self-concept is measured through behavior, not qualia.The paper explicitly frames the hypotheses as theoretical claims to be tested by its protocol.Source: §3 · p. 5

4 System Architecture

SEAA closes a loop from social observation through action and reflection back to editable behavioral inertia and a self-model. Its modules combine HMM state transitions, private memory, multi-agent interaction, and reflection-driven updates.

4.1 Overview

  • The per-step architecture links social observation, HMM state activation, memory- and self-model-conditioned action, interaction, reflection, and an inertia/self-model update.The highlighted feedback from step 6 to step 1 distinguishes SEAA from text-only reflection.Source: §4.1 · p. 5, Figure 1 · p. 6

4.2 HMM behavioral-inertia module

  • SEAA makes the HMM transition matrix time-varying by updating it from reflection, with a direction matrix and learning rate followed by row normalization.The update is presented as the mathematical locus of reflection changing the agent's disposition.Source: §4.2 · p. 6

4.3 Reflexive metacognition module

  • The metacognition module reviews a trajectory to assess habitual fit, identify driving tendencies and bias, propose adjustments, and structure an edit to P_t.The language-to-parameter mapping is identified as the key engineering novelty and main technical risk.Source: §4.3 · p. 6

4.4 Private autobiographical memory

  • Each agent keeps isolated timestamped records of its own experiences, actions, and reflections, retrieving relevant records for decisions.Source: §4.4 · p. 6

4.5 Multi-agent social-contrastive environment

  • The text-based society begins with identical agents and no assigned roles or traits; agents observe others' behavior as material for social contrast.Source: §4.5 · p. 6

4.6 Self-model updater

  • A periodically rewritten natural-language self-model summarizes who an agent is using recent memory and reflection, and its divergence is an observable for Hypothesis 3.Source: §4.6 · p. 7

4.7 Pseudocode

  • The pseudocode initializes agents homogeneously, samples HMM states, acts using memory and self-model context, reflects on feedback, and updates the transition matrix.Reflection is parsed into a transition delta and the updated matrix is row-normalized.Source: §4.7 · p. 7

5 Experimental Design and Verification Protocol

The protocol tests whether editable behavioral inertia and social interaction produce differentiated, stable agents, first in a language-model-free prototype and then through self-interviews and group deliberation with a hosted LLM. Across these experiments, SEAA shows internal differentiation, adaptation after a contingency shock, distinct self-narratives, and social structure not observed in matched controls.

5 Experimental Design and Verification Protocol

  • The protocol specifies five-agent interaction and isolated controls, tests reflection edits and social setting, and measures behavior, matrix divergence, consolidation, self-model contrast, and stability.It proposes a 2 × 2 design for reflection edits and society, repeated self-interviews, process logging for replay, and explicit falsification criteria; the numerical prototype isolates whether experience → reflection → inertia update breaks symmetry.Source: §5.2 · p. 8, §5.3 · p. 8, §5.4 · p. 8

5.7.1 Formal specification

  • The prototype uses four latent states, agent-specific experience-need walks, state preferences, an HMM transition matrix, and self- and group-model averages.Reflection reinforces states whose behavior fits an agent’s experience need, updating the matrix; controls compute and record the same reflection but leave the matrix unchanged.Source: §5.7.1 · p. 9

5.7.2 Population-level results

  • Across 30 seeds, SEAA alone showed all four signatures: rising inertia divergence, long-term personality distance, determinism, and a self–other gap.End-of-run SEAA vs control values were matrix divergence 1.91 ± 0.29 vs 0; personality distance 0.51 ± 0.13 vs 0.13 ± 0.02; determinism 0.84 ± 0.10 vs 0.25; and self–other gap 0.35 ± 0.09 vs 0.08 ± 0.01. All four effects were significant at p < 10−10, with Cohen’s d > 4.Source: §5.7.2 · p. 10, Figure 2 · p. 11

5.7.3 Individual-level “microscope” view

  • In a representative run, five agents began with identical state preferences but consolidated onto different dominant states; Agent 4 changed disposition twice before stabilizing at step 508.The trajectories show the individual-level process behind the population-level differentiation.Source: §5.7.3 · p. 11, Figure 3 · p. 12

5.7.4 Contingency shock: does reflection-driven inertia re-adapt? (H2)

  • After a shock at step 300, all 150 SEAA agents switched from the maladaptive dominant state, with median switching time 152 steps; controls could not revise their matrix.Shocked-state occupancy fell from 0.94 to 0.06, and determinism recovered to 0.80 after a transient dip.Source: §5.7.4 · p. 12, Figure 4 · p. 13

5.7.5 Ablation: dose-response of the reflection edit

  • The reflection edit shows a dose response: matrix divergence and determinism rise with η, while η=0 reproduces the text-only control.The reported divergence values for η = 0, 0.01, 0.025, 0.05, 0.1 were 0, 1.07, 1.91, 1.99, 2.04; determinism was 0.25, 0.37, 0.76, 0.83, 0.84. Effects saturate around η ≈ 0.05–0.1.Source: §5.7.5 · p. 12, Figure 5 · p. 13

5.9 Real-LLM validation: self-interviews and group deliberation

  • With a hosted LLM, self-interview overlap was lower for SEAA than controls across 10 seeds: 0.267 ± 0.017 versus 0.391 ± 0.020.The difference was significant (one-sided Mann–Whitney U, p = 9.0 × 10−5; Cohen’s d = 6.76); SEAA voices were described as distinct and profile-consistent, while controls were interchangeable.Source: §5.9 · p. 15, Figure 7 · p. 15

5.9.1 A five-agent round-table deliberation

  • Across six deliberation seeds, SEAA produced an outlier share of 0.77 versus 0.33 for controls, and controls never formed a consensus hub or outlier.SEAA groups also had lower speech-variety index (0.28 ± 0.03 vs 0.34 ± 0.03) and fewer “none” responses (0.00 vs 0.53); the outlier-share difference had p = 0.011. Replication remains confined to one deliberation topic and one model family.Source: §5.9.1 · p. 16, Figure 8 · p. 16

6 Discussion

SEAA defines selfhood through observable behavioral emergence while leaving subjective consciousness unresolved. The discussion identifies modeling, experimental-setting, and ethical boundaries for the framework.

6.1 Behavioral emergence versus subjective consciousness

  • SEAA studies behavioral consistency, differentiation, and an expressed self–other boundary, but does not determine whether these imply subjective experience.Functionalist and biological accounts differ on whether such behavior would entail qualia.Source: §6.1 · pp. 16–17

6.2 Resonance with Zhuangzian epistemology

  • The Zhuangzian stance treats selfhood as an operational account grounded in behavior that can be measured, not a claim about inner experience.The discussion invokes limits on knowing other minds and compares observation to watching paramecia under a microscope.Source: §6.2 · p. 17

6.3 Limitations

  • The model simplifies personality into a small discrete state set and personal experience into a random walk, despite their greater real-world complexity.The authors propose latent-variable and learned-experience extensions; LLM studies also use one hosted model family and one deliberation topic, while the society lacks embodiment and requires careful tuning of the language-to-parameter mapping.Source: §6.3 · p. 17

6.4 Ethical considerations

  • If future runs produce highly complex self-directed behavior, questions about agent moral status and termination may become pressing.The authors recommend complete records, avoiding gratuitous anthropomorphism, and distinguishing behavioral emulation from subjective experience in reporting.Source: §6.4 · p. 17

7 Conclusion

SEAA combines behavioral inertia, reflection-driven updates, and social comparison into a developmental loop for studying observable selfhood. The prototype and hosted-LLM experiments report differentiated behavior and narratives, while the authors frame artificial-self research as empirical inquiry that does not settle inner experience.

7 Conclusion

  • Across 30 seeds, reflection-driven inertia updates produced distinct, stable personalities among initially identical agents, while controls remained homogeneous.A contingency shock prompted revision of an entrenched disposition, and a learning-rate ablation found the effect causal and dose-dependent.Source: §7 · p. 17
  • Hosted-LLM experiments produced distinct first-person self-narratives and a deliberation topology of leaders, hubs, and outliers absent from undifferentiated controls.The LLM experiments were replicated across seeds.Source: §7 · p. 17
  • SEAA couples an HMM of behavioral inertia with reflection that edits inertia and social-contrastive interaction in one developmental closed loop.The framework operationalizes self through inertia stabilization, reflection-driven evolution, and differentiation through social comparison, with falsifiable hypotheses, pseudocode, and a replayable sandbox.Source: §7 · p. 17
  • The authors argue that artificial-self questions can be pursued empirically while acknowledging limits on external knowledge of inner experience.They identify cross-model and cross-topic replication, higher-dimensional latent states, embodiment, and larger societies as next steps.Source: §7 · pp. 17–18

A Prompts Used in the Real-LLM Experiments

The appendix records the prompts used for real-LLM self-interviews and five-agent deliberations. Profiles ground the agents’ speech and sociometry in their simulated dispositions and group interaction.

A Prompts Used in the Real-LLM Experiments

  • All agents share one model, with individual differences supplied at runtime through each agent’s objective profile.The API key is provided through an environment variable and is not included in prompts or artifacts.Source: A Prompts Used in the Real-LLM… · p. 18

A.1 One-on-one self-interview (Section 5.8–5.9)

  • Self-interviews ask agents to describe themselves, explain differences from others, and identify experiences that changed them.The user prompt instructs first-person speech consistent with the profile, in English, with two to four short sentences per question.Source: §A.1 · p. 18
  • Interview inputs include dominant disposition, three self-model traits, group averages, signed gaps, and disposition-change history.The prompt says not to invent traits or numbers that contradict the objective profile.Source: §A.1 · p. 18

A.2 Five-agent round-table deliberation (Section 5.9)

  • Deliberation prompts give each agent a settled disposition and trait profile, with instructions to remain in character and not speak for others.The persona prefix is paired with a shared topic and turn-specific suffix.Source: §A.2 · p. 18
  • The five agents decide whether to travel from an enclosed habitat with dwindling supplies to a distant resource zone with unknown risks.The prompt also asks who goes or stays and how any returned resources should be shared.Source: §A.2 · p. 18
  • Each deliberation turn asks the designated agent to react briefly to the transcript so far, in character and addressing the group.The suffix requests one to three sentences.Source: §A.2 · p. 18
  • After discussion, each agent nominates the most- and least-preferred participant and gives a short reason based only on that discussion.The prompt requires a one-line response in a fixed format.Source: §A.2 · p. 18

B Full Real-LLM Self-Interview Transcripts

The reproduced self-interviews show agents describing distinct dispositions and comparing their trait profiles with the group. Control transcripts instead include agents who describe balanced profiles and no disposition lock-in.

B Full Real-LLM Self-Interview Transcripts

  • One SEAA agent describes an impulsive disposition with risk-taking at 0.92, expressiveness at 0.82, and cooperativeness at 0.30.The agent characterizes itself as acting before deliberation and preferring action over consensus.Source: B Full Real-LLM Self-Interview… · p. 19
  • A calm agent reports cooperativeness +0.36 and risk-taking −0.18 relative to the group, and says reflection updates kept its disposition stable after one shift at step 75.It describes itself as steady and moderately expressive.Source: B Full Real-LLM Self-Interview… · p. 19
  • An alert agent reports cooperativeness +0.13, risk-taking +0.12, and expressiveness −0.06, with no change in its leading disposition.It describes itself as attentive and deliberate rather than reactive.Source: B Full Real-LLM Self-Interview… · p. 19
  • A pessimistic agent reports below-group values for all three traits and says its disposition shifted twice, with the latest shift at step 508.Its account says the update hardened its inertia rather than loosening it.Source: B Full Real-LLM Self-Interview… · p. 19
  • A control agent describes a balanced profile with no dominant disposition and only small trait gaps from the group average.It reports no disposition changes and says it remained consistent with its baseline profile.Source: B Full Real-LLM Self-Interview… · p. 20

C Full Round-Table Deliberation Transcript

The evolving group voices distinct positions on urgency, risk, fairness, and decision rules, while the control group repeatedly favors gathering information before committing. Control-group nominations show some agents aligned with particular peers, but no single peer is most-agreed across the group.

C Full Round-Table Deliberation Transcript

  • In the evolving group, Agent 1 pushes to act, Agents 2–3 press for agreed thresholds, and Agents 4–5 remain pessimistic and plan to stay.By Round 3, the group accepts scouts, an equal split, and a written turnaround point while Agent 1 still wants to go if the numbers are favorable.Source: C Full Round-Table Deliberation… · pp. 21–22
  • In the control group's first round, all five agents favor obtaining numbers before committing, with no agent pressing a distinct position.Their second-round discussion adds specific information needs but continues the data-first approach.Source: C Full Round-Table Deliberation… · p. 22
  • Control-group agents align on gathering concrete information before deciding, then add break-even, sharing, and contingency rules to the checklist.Their proposals include distance, travel time, hazards, supplies, a pre-agreed resource split, and plans for a failed return.Source: C Full Round-Table Deliberation… · pp. 22–23
  • Control-group nominations vary: Agents 1 and 3 name no standout peer, while Agents 2, 4, and 5 nominate different most-agreed peers.No single peer is named most-agreed by all five agents.Source: C Full Round-Table Deliberation… · p. 23

D.1 Numerical prototype: full configuration

The numerical prototype uses four latent states, three behavior dimensions, five agents, and 600 steps per run, with 30 independent seeds for aggregate results.

D.1 Numerical prototype: full configuration

  • The prototype models four latent states—calm, alert, impulsive, and pessimistic—and three behavior dimensions: cooperate, risk, and express.Each society has five agents and runs for 600 steps.Source: §D.1 · p. 23
  • The reflection learning rate η is 0.05, preference-to-transition strength β is 4.0, and experience random-walk step σwalk is 0.03.Enacted-behavior noise is N(0, 0.04), and reward exploration luck is N(0, 0.10).Source: §D.1 · p. 23
  • Initial inertia uses a row-normalized matrix with 0.10 fill and 0.45 diagonal, yielding approximately 0.647 diagonal and 0.118 off-diagonal values.Preference values are clipped to [−2, 2], and self/others exponential-moving-average weights are 0.97/0.03.Source: §D.1 · p. 23

D.2 Software, hardware, and run settings

The numerical experiments run on a single CPU without a GPU or network, while the language layer uses DeepSeek-Chat API calls with specified temperatures and multi-seed replications.

D.2 Software, hardware, and run settings

  • The numerical layer uses Python 3, numpy 1.26.4, scipy, and matplotlib; hmmlearn 0.3.3 is optional for the base-HMM interface.The 30-seed aggregate, contingency-shock experiment, η-ablation, and figures run in seconds on a single CPU without a GPU or network.Source: §D.2 · p. 23
  • Table 5 presents the prototype behavior vector bs for each latent state.Source: Table 5 · p. 24
  • The contingency shock begins at t = 300, while the η-ablation sweeps six learning rates with 10 seeds per point.The tested η values are 0, 0.01, 0.025, 0.05, 0.1, and 0.2.Source: §D.2 · p. 24
  • DeepSeek-Chat calls use temperature 0.8 for self-interviews and deliberation turns and 0.2 for sociometry nominations.Replication uses 10 society seeds per condition for self-interviews and six per condition for deliberations, with three rounds in both deliberation conditions.Source: §D.2 · p. 24
  • The reproduction sequence runs the simulation, interview experiments, online interviews, group chat, LLM replication, and plotting scripts; raw snapshots, transcripts, and scripts are released.All randomness is seeded, and the API key is read from an environment variable rather than written to code, transcripts, or the manuscript.Source: §D.2 · p. 24
Loading 2609.17331v1…