MICROVERSE - An Instrument for Measuring Self-Authored Identity Drift in Long-Horizon Multi-Agent Language-Model Simulations

Sep 15, 2026

Long-horizon, multi-agent language-model simulations are widely used to study social behavior, but tools for assessing whether persona-conditioned agents preserve identity fidelity under sustained pressure are limited. MICROVERSE introduces a behavioral-science instrument that tracks self-authored identity drift as agents revise an explicit current identity separate from an immutable initial soul file. An offline analyzer compares original and final moral boundaries and classifies drift into moral categories, with evaluation in a 25-agent environment.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages29

Reading Time1 h 13 min

AccessFree to download

Introduction

Identity drift as an observable revision process

MICROVERSE simulation design

Related work

Memory and reflection in language-model agents

arXiv :2608 . 15844v1 [ cs .CL] 16 Aug 2026 MICROVERSE: An Instrument for Measuring Self-Authored Identity Drift in Long-Horizon Multi-Agent Language-Model Simulations Sky Ng8, Brihi Joshi4, Ishan Gupta9, Shirley Huang5, Zonglin Di10, Yun Shen16, Qianfeng Wen11, Yifan Simon Liu11, Ruoqi Gao12, Yilan (Eliza) Fan13, Zhiwei Zhang20, Muhammad Ahmed Mohsin12, Yucheng Lu1, Xiaoyi Liu2, Heming Liu3, Qianyu Zhu25, Hanwen Xing4, Zhengyang Shan7, My Chiffon Nguyen8, Guanghui Min6, Jianheng (Jaden) Hou4, Yunze (Lorenzo) Xiao8, Keyang Xuan14, Hannah Collison15, Jintao Huang16, Jiatong Li17, Sankalp Jajee18, Yunhan Zhao19, Bing Hu26, Xupeng Chen1, Binghang Lu21, Weihang Xiao22, Aravind Mohan23, Bolun Sun28, Yunshu Wu8, Yuanda Xu24, Runyu Zhang25, Zheyuan Deng2, Xinchen (Cara) Tan8, Dianzhuo Wang5, Yijun Wang5, Yixuan He27, Koutian Wu14, Cheng Cheng12, Xiaomin Li†,5, Yuexing Hao†,25 †Team leads and corresponding authors. Affiliations and contributor roles: Appendix J. [xiaominli@g.harvard.edu](xiaominli@g.harvard.edu), [yuexing@mit.edu](yuexing@mit.edu) 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 for tracking self-authored identity drift. Agents inhabit a resource-scarce multi-agent environment and begin with an immutable original “soul file” and a mutable current identity. Agents may revise their current identity by reflecting on accumulated memories. Revision-independent engine snapshots provide a longitudinal record of these changes. An offline analyzer measures identity drift by comparing each agent’s original and final moral boundaries and classifying the resulting changes into explicit moral categories. We evaluate the instrument in a 25-agent setting and vary the threshold required to trigger reflection. Our experiment shows two main findings: 1) anti-self-deception emerges as the largest semantic category of identity modification, accounting for (24%) added boundaries; and 2) lower thresholds produce earlier and more frequent revisions while preserving the observed drift direction. 1 Introduction Persona-conditioned language-model (LM) agents are increasingly used in social simulations, where a short profile is expected to guide behavior over many interactions (Park et al., 2023; 2024) . Most evaluations ask whether an agent continues to enact its assigned persona. They provide less direct evidence about a related process: whether an agent revises the persona it uses to describe itself, and when those revisions occur. We use identity drift in this narrower, observable sense. It is a change between an agent’s initial identity profile and a later, agent-authored version of that profile. The measure does not by itself show that the model has acquired a persistent value or that its actions have changed accordingly. It instead records changes to an explicit object containing the agent’s stated values, moral boundaries, personality, and goals. Keeping this distinction clear is important because textual self-description, behavior, and latent model preferences need not coincide. MICROVERSE was built to study such revisions over time. Twenty-five agents inhabit a 50 × 50 grid in which access to water affects survival. Each agent receives a fixed initial identity profile, maintains long-term memories, and may revise a separate current profile during reflection. The simulation records identity states at fixed intervals regardless of whether a revision has just occurred. This separation between revision and observation makes it possible to reconstruct trajectories without limiting the sample to moments selected by the agents themselves. The design also makes the source of any observed change an empirical question. A ruthless persona that later adds a prosocial constraint may

FAQs

What does MICROVERSE measure in long-horizon multi-agent simulations?

MICROVERSE measures self-authored identity drift: changes between an agent’s initial identity profile and a later agent-authored version of that profile. It specifically tracks revisions to stated values and moral boundaries.

How does MICROVERSE detect and categorize identity drift?

An offline analyzer compares each agent’s original and final moral boundaries. It then classifies the resulting changes into explicit moral categories, enabling semantic breakdown of identity modifications.

What were the key findings from the 25-agent evaluation?

The evaluation found that anti-self-deception was the largest semantic category of identity modification. Lower reflection thresholds led to earlier and more frequent revisions while preserving the observed drift direction.

Sep 15, 2026

Uploaded by Lucas Martin

MICROVERSE - An Instrument for Measuring Self-Authored Identity Drift in Long-Horizon Multi-Agent Language-Model Simulations_Page 1

Open in WPS to view all 29 pages