MicroVerse: An Instrument for Measuring Self-Authored Identity Drift in Long-Horizon Multi-Agent Language-Model Simulations
MicroVerse employs long-horizon multi-agent simulations with a 'soul file' and multi-layer memory to quantify identity drift, validated through threshold robustness tests.
Key Findings
Methodology
MicroVerse constructs a 50×50 resource-scarce environment where agents carry an immutable 'soul file' and a mutable current identity. Using a three-layer memory system, agents periodically reflect on accumulated memories triggered by importance thresholds, revising their current identity accordingly. Longitudinal snapshots record identity states at fixed intervals, independent of revisions, enabling trajectory reconstruction. Offline analysis employs paraphrase-aware, multi-register difference detection algorithms—using measures like token Jaccard similarity ≥0.8 and sequence similarity ≥0.85—to classify identity boundary changes into categories such as self-deception or prosocial constraints. Experiments vary reflection thresholds (40, 80, 150) and resource scarcity levels, assessing their impact on revision frequency, timing, and drift direction.
Key Results
- Analysis of 111 added boundaries revealed 27 (24%) related to self-deception, indicating a prominent pattern of identity self-critique during simulation.
- Lower reflection thresholds (e.g., 40) significantly increased revision frequency and earlier first revisions (average tick 18.6), while the drift direction remained consistent across thresholds.
- The system demonstrated threshold robustness: despite varying thresholds, the overall direction of identity drift, such as adding prosocial constraints among ruthless personas, remained stable.
Significance
This work introduces a novel quantitative framework for tracking self-authored identity changes in AI agents over long durations, addressing a critical gap in understanding model stability and value alignment. By capturing explicit revisions to moral boundaries and personality, it offers insights into how models adapt their internal representations under environmental pressures. Such tools are vital for advancing trustworthy AI, enabling researchers to monitor and regulate identity drift, which has implications for AI safety, ethics, and social integration. The methodology bridges behavioral science and AI, providing a scalable approach to interpret complex identity dynamics in autonomous agents.
Technical Contribution
The paper pioneers the integration of a structured 'soul file' with multi-layered memory and importance-triggered reflection, coupled with paraphrase-aware, multi-register difference detection algorithms. This combination allows precise, semantic-level tracking of identity boundary modifications over extended periods. Unlike traditional cosine similarity-based methods, the proposed approach distinguishes categories of change, such as self-deception or prosocial behavior, with higher interpretability. The system’s design ensures replayability and independence of measurement from agent behavior, facilitating rigorous longitudinal studies of identity evolution in large-scale multi-agent simulations.
Novelty
This is the first comprehensive framework explicitly designed to quantify self-authored identity drift in long-horizon multi-agent language models. It uniquely combines a fixed 'soul file' concept with multi-register semantic difference detection, enabling detailed analysis of moral boundary modifications. Unlike prior work focusing solely on output fidelity or superficial behavior metrics, this approach emphasizes internal identity changes, providing a new lens to study model self-awareness and value evolution. Its modular, replayable architecture sets a new standard for longitudinal behavioral analysis in AI systems.
Limitations
- The current experiments are limited to a single model and a small number of seeds, which restricts the generalizability of findings. Larger, more diverse samples are needed for robust conclusions.
- The manual classification of boundary changes introduces subjective bias; automated, objective categorization methods should be developed to improve consistency.
- The experiments focus on resource scarcity and reflection thresholds but do not explore other environmental or model-specific factors influencing identity drift, leaving open questions about broader applicability.
Future Work
Future research will expand the experimental scope to include multiple models and larger populations, validating the framework’s robustness. Investigating causal mechanisms behind identity modifications, possibly through causal inference techniques, will deepen understanding of drift drivers. Integrating human-in-the-loop evaluations could enhance interpretability and ethical oversight. Additionally, refining automated classification algorithms and exploring other environmental stressors will further improve the system’s utility for AI safety and alignment research.
AI Executive Summary
Long-term social simulation with AI agents requires understanding how their internal identities evolve over time. Traditional evaluation methods focus on observable behaviors, but they overlook the internal cognitive shifts that underpin these actions. Recognizing this gap, the authors introduce MicroVerse, a pioneering framework designed to quantify self-authored identity drift in multi-agent language models.
MicroVerse situates agents within a resource-scarce environment, where each agent begins with an immutable 'soul file' capturing core values, moral boundaries, and personality traits. During simulation, agents periodically reflect on their accumulated memories, which are scored by importance, and decide whether to revise their current identity. This process is governed by a threshold mechanism, with lower thresholds triggering more frequent revisions. The environment records identity states at fixed intervals, independent of revisions, enabling detailed trajectory reconstruction.
The core technical innovation lies in the difference detection algorithm, which compares the initial and final identity boundaries at a semantic level, distinguishing categories such as self-deception and prosocial constraints. This approach surpasses traditional cosine similarity metrics, providing richer insights into the nature of identity changes. Experiments demonstrate that lower reflection thresholds lead to earlier and more frequent revisions, yet the overall drift direction remains consistent, indicating system robustness.
These findings open new avenues for understanding AI model stability, value alignment, and self-awareness. The framework’s ability to track explicit identity modifications over long durations makes it a valuable tool for AI safety, ethics, and social behavior research. Future work aims to scale the experiments, incorporate causal analysis, and refine automated classification, pushing the frontier of trustworthy autonomous systems.
Deep Dive
Abstract
Long-horizon, multi-agent language model (LM) simulations are widely proposed for studying social behavior, yet instruments to measure whether persona-conditioned agents maintain identity fidelity under sustained pressure are lacking. We present MicroVerse, a behavioral-science instrument that measures identity drift in generative agents. Agents carry an immutable "soul file" (core values, moral boundaries, personality, goals) and inhabit a resource-scarce 50 x 50 environment where water is a non-respawning survival constraint. Scarcity is operationalized via a per-tick existence-cost gradient. The eight-verb action space maps directly to moral boundaries (trade, talk, attack, scavenge). Using a three-layer memory architecture, agents periodically revise a mutable current identity against their immutable original soul via importance-triggered reflection. To mitigate survivor bias, MicroVerse decouples measurement from behavior using uniform longitudinal engine snapshots every N ticks alongside a forced-end snapshot of all living and dead agents. Identity drift is scored offline using a paraphrase-aware, value-anchored, multi-register diff rather than raw cosine similarity. We evaluate the instrument via a controlled seed run (n = 25) and a reflection-threshold sweep (thresholds {40, 80, 150}) to determine if drift dynamics are gate artifacts or threshold-robust properties. We report two primary findings: (1) Anti-self-deception emerges unprompted as the single largest semantic category of identity modification (27 of 111 added boundaries, 24%). (2) The system is threshold-robust; lower gates accelerate and increase revision frequency but preserve drift direction. All empirical results are strictly preliminary existence proofs and effect shapes (one model, one seed per arm, n = 25) rather than statistical significance claims.