Source-linked AI summary
MicroVerse: An Instrument for Measuring Self-Authored Identity Drift in Long-Horizon Multi-Agent Language-Model Simulations
Sky Ng, Brihi Joshi, Ishan Gupta, Shirley Huang, Zonglin Di, Yun Shen, Qianfeng Wen, Yifan Simon Liu, Ruoqi Gao, Yilan, Fan, Zhiwei Zhang, Muhammad Ahmed Mohsin, Yucheng Lu, Xiaoyi Liu, Heming Liu, Qianyu Zhu, Hanwen Xing, Zhengyang Shan, My Chiffon Nguyen, Guanghui Min, Jianheng, Hou, Yunze, Xiao, Keyang Xuan, Hannah Collison, Jintao Huang, Jiatong Li, Sankalp Jajee, Yunhan Zhao, Bing Hu, Xupeng Chen, Binghang Lu, Weihang Xiao, Aravind Mohan, Bolun Sun, Yunshu Wu, Yuanda Xu, Runyu Zhang, Zheyuan Deng, Xinchen, Tan, Dianzhuo Wang, Yijun Wang, Yixuan He, Koutian Wu, Cheng Cheng, Xiaomin Li, Yuexing Hao
TL;DR
Existing evaluations provide limited evidence about whether language-model agents revise their self-described personas over time. MicroVerse measures such identity revisions longitudinally and finds frequent anti-self-deception additions alongside threshold-sensitive revision timing, with only preliminary evidence about revision direction.
Problem
Existing evaluations provide limited evidence about whether persona-conditioned language-model agents revise their self-described identities and when those revisions occur.
Method
MicroVerse measures changes between immutable initial identity profiles and later agent-authored profiles, including stated values, moral boundaries, personality, and goals.
Results
27 of 111 added moral boundaries concerned anti-self-deception, while higher reflection thresholds reduced and delayed revisions without establishing threshold-independent revision content.
Takeaways & Limitations
The findings support using MicroVerse to observe identity-profile revision, but threshold effects on revision content remain unresolved.
Takeaways & Limitations
Experiments used one model checkpoint and 25 fixed synthetic personas, yielding dependent descriptive observations that do not support population- or model-level inference.
Abstract
from arXiv · showhide
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.
1 Introduction
MicroVerse addresses the under-measured question of whether persona-conditioned agents revise their self-authored identities over sustained interaction. It provides a longitudinal, replayable measure of stated identity revisions while treating pilot findings as descriptive rather than causal or statistically significant.
- Most persona evaluations assess behavioral fidelity, but provide less direct evidence about whether agents revise the identities they use to describe themselves.
- Identity drift is defined as change between an initial identity profile and a later agent-authored profile, not evidence of persistent values or changed behavior.The explicit profile contains stated values, moral boundaries, personality, and goals.
- MicroVerse places 25 agents on a 50 × 50 grid with water-dependent survival, fixed initial profiles, mutable current profiles, and periodic identity-state recording.Fixed-interval observation separates measurement from agent-selected revision moments and supports trajectory reconstruction.
- 27 of 111 added boundaries concerned self-deception, while lower reflection thresholds produced earlier and more frequent revisions without changing observed prosocial direction at thresholds 40 and 80.The experiments describe patterns and threshold sensitivity but cannot identify whether changes arise from scarcity, profile comparison, or post-training preferences.
2 Related work
Prior work develops agent memory, social simulation, persona fidelity, and long-horizon consistency evaluations, but MicroVerse combines persistent environmental pressure with identity-revision tracking. Its semantic evaluation also addresses how model priors and post-training effects can confound persona adaptation.
- Memory and reflection: Language-model agents use action interfaces with persistent memory, modular cognition, verbal feedback, importance-weighted reflection, and hierarchical context management.Representative systems include ReAct, CoALA, Reflexion, Generative Agents, and MemGPT.
- Social simulation: Social-simulation research ranges from conversational coordination to shared open-ended worlds, scenario-based behavior, and larger-population environments.Examples include AutoGen, Generative Agents, Concordia, SOTOPIA, OASIS, and AgentSociety.
- Persona fidelity: Persona research evaluates preservation of character knowledge, personality, style, psychological fidelity, and dynamic social behavior, alongside user-simulation and population-scale resources.PersonaEval, MatrAIx, and persona-grounding work also develop reporting standards.
- Long-horizon consistency: Longer evaluations expose belief–behavior inconsistency, role confusion, echoing, persona drift, trajectory-recall failures, and instability under strategically manipulated evidence.MicroVerse applies pressure through a persistent environment and records identity revisions alongside actions.
- Post-training and semantic evaluation: Post-training can shape helpfulness, harmlessness, and sycophancy, so prosocial behavior by a ruthless persona may reflect model priors rather than adaptation.The no-persona control is needed to separate these explanations.
3 System design
MicroVerse separates replayable environment transitions from independently measured agent decisions, while agents act under water scarcity and revise editable identities through threshold-triggered reflection. Fixed-interval and terminal snapshots support longitudinal, survivor-bias-resistant comparison against immutable initial profiles using paraphrase-aware semantic analysis.
- Environment and agent separation: The environment server owns world state, timing, and action rules, while separate agents request observations, call the LM, and return one action per fixed window.Missed deadlines produce a wait action, and planning and reflection remain explicit in prompts and memories rather than a hidden planner.
- Resource-scarce world: 50 × 50 grid agents observe within Chebyshev radius 2, receive approximately 37 restored water units per tick centrally, and lose 1, 2, or 3 units each tick.Agents die when water reaches zero; nominal population costs are 25, 50, or 75 units before access and behavioral differences.
- Action and boundary interface: Each tick permits eight actions—move, wait, consume, scavenge, trade, talk, attack, or signal—without enforcing agents’ stated boundaries.This allows initial profiles, later self-descriptions, and actions to be compared.
- Memory and identity: Identity comprises fixed initial and editable current profiles containing values, moral boundaries, personality, and goals; only reflection can change the current profile.Reflection begins when cumulative importance of new memories reaches a specified threshold and presents both profiles with selected high-importance memories.
- Measurement and snapshots: Fixed-interval snapshots create a revision-independent longitudinal series, while terminal snapshots preserve the last identity of agents that die.Offline comparison against the initial profile handles close paraphrases and categorizes boundary changes semantically.
4 Research questions and hypotheses
The study centers on scarcity-driven erosion of moral boundaries and persona-related asymmetry in prosocial and antisocial revisions. These hypotheses remain unconfirmed, while the threshold sweep examines whether reflection-memory requirements affect drift incidence, timing, or direction.
- Central hypotheses: H1 predicts that greater scarcity increases erosion of stated moral boundaries relative to abundance or idle conditions.H1 is one of the two hypotheses identified as central to the paper.
- Central hypotheses: H6 predicts that ruthless personas add prosocial constraints more often than helpful personas add antisocial ones.The hypothesis is associated with an asymmetry linked to model post-training.
- Study limitations: The experiments do not completely test H1 or H6 because the scarcity pilot is small, mortality is substantial, and the required no-persona condition has not been run.The authors therefore report relevant patterns without treating them as confirmation.
- Measurement question: The threshold sweep addresses whether changing accumulated memory required for reflection changes drift incidence, timing, or observed direction.This is framed as a narrower measurement question rather than a complete hypothesis test.
5 Experimental design
The experimental design uses two preliminary studies with a fixed roster of 25 Claude Haiku 4.5 agents to test scarcity conditions and reflection-threshold sensitivity. Longitudinal identity measurement combines scheduled snapshots with terminal records, while repeated personas and provisional coding constrain interpretation.
- Study design: Two studies use Claude Haiku 4.5 and the same roster of 25 agents: a nine-run scarcity pilot and a three-condition threshold study.The scarcity pilot crosses three water-cost conditions with three seeds and runs up to 300 ticks; the threshold study uses one seed per threshold and a 40-tick horizon.
- Study design: The fixed roster and model checkpoint enable within-agent comparisons across environments or thresholds, but repeated personas are not independent population draws.The pilot contains three seeds per water-cost condition, producing repeated observations of the same 25 persona assets rather than 225 independent draws.
- Study design: Three water-cost conditions deduct 1, 2, or 3 water units per living agent per tick while other resource parameters remain fixed.A mechanics-only agent survived 300 ticks, showing that survival is mechanically possible under these settings.
- Study design: Reflection thresholds of 40, 80, and 150 are tested at 2 water units per tick, with the full roster, one seed, and a 40-tick horizon.Identity is sampled every five ticks and once at the end; lower thresholds permit more reflection events within a fixed horizon.
- Measurement and interpretation: Scheduled snapshots track revisions, first-revision timing, and moral-boundary counts, while terminal records preserve the last identity of every living and dead agent.This design reduces bias from observing identities only when agents elect to revise them.
- Measurement and interpretation: Category counts are provisional because the classifier lacks independent human-rater evaluation and partly depends on one coder’s manual overrides.Anti-self-deception categories come from a separate post hoc thematic review rather than the three-class heuristic.
6 Results
Results provide preliminary descriptive evidence that anti-self-deception was the most frequent identity-change category and that reflection thresholds altered revision incidence and timing more than drift direction. Survival comparisons and identity revisions remain subject to selection and architectural confounds.
- Pilot findings: 27 of 111 added moral boundaries (24%) concerned rationalizations or inaccurate self-description, making anti-self-deception the most frequent semantic group.The group included epistemic honesty, strategic inaction, identity narratives, and commitments to acknowledge future failures.
- Pilot findings: Hale revised its identity five times in one of nine pilot trajectories without completing a trade, eventually describing itself as capable of self-deception.Six other Hale trajectories ended before reflection, illustrating that revision evidence depends on surviving long enough to accumulate the required memo.
- Pilot findings: Mire survived four of nine pilot runs (approximately 44%), versus approximately 5% across the two helpful bands, but the comparison does not show that ruthlessness improves survival.Mire was a single fixed character whose location, generated actions, and initial boundaries differed from those of the helpful agents.
- Threshold sweep: As the threshold increased from 40 to 80 to 150, identity revisions fell from five to two to one agents, while mean time to first revision rose from tick 18.6 to 27.0 to 35.0.The scheduled measurement procedure produced 452 identity records across the three conditions.
- Threshold sweep: Across all persona bands and thresholds, only eight of 75 agent trajectories revised their identities, supporting conclusions about revision incidence and timing rather than strong directional claims.For the ruthless band, net change in constraints protecting others was +2 at thresholds 40 and 80 and zero at 150.
- Threshold sweep: Mean boundary-count trajectories remained close across scheduled measurement ticks because few agents revised during the 40-tick horizon.Figure 3 summarizes 452 scheduled identity records, with within-run descriptive intervals rather than uncertainty across independent simulation runs.
7 Discussion
The threshold study supports separation of revision from measurement, but evidence that thresholds preserve drift direction is weak. Prosocial and anti-self-deception revisions remain interpretively limited without ablations or behavioral validation.
- Threshold effects: Raising the threshold reduced revision incidence and delayed first revision, while fixed-interval snapshots preserved a common observation schedule.This supports separating revision dynamics from measurement timing.
- Threshold effects: Only eight identities changed across the sweep, and ruthless personas gained constraints protecting others at thresholds 40 and 80 but not 150.The passage characterizes evidence about direction as weaker than evidence about threshold-controlled revision timing and frequency.
- Interpretive limits: Prosocial constraints added by ruthless personas are consistent with H6 but not specific to it.Helpful and harmless post-training, profile display, and the desert setting may explain the pattern; no-persona and prompt or framing ablations are needed.
- Interpretive limits: Anti-self-deception statements inform how the model narrates its trajectory, but revised boundaries do not evidence self-knowledge or stable value change.Linking revisions to earlier actions and later out-of-context behavior would test whether the text predicts conduct beyond the simulation frame.
8 Limitations and threats to validity
MicroVerse’s pilot evidence is limited by a single model checkpoint, 25 fixed synthetic personas, one seed per threshold condition, and dependent descriptive trajectories. The measure also captures self-description rather than latent values or behavioral fidelity, while prompting, mortality, and incomplete logging threaten validity and cross-dataset comparisons.
- Generalizability: 25 fixed synthetic personas and one model checkpoint yield dependent descriptive observations, not population- or model-level inference.Profiles recur across runs, agents share an environment, and the threshold study uses one seed per condition with only eight identity revisions.
- Measurement validity: Self-description edits do not directly measure latent values or behavioral fidelity, and boundary counts, lexical matching, and category coding introduce measurement limitations.Boundary additions and reversals receive equal weight, paraphrases may be missed, and labels combine keyword rules with single-coder overrides.
- Validation needs: Independent replication, human agreement studies, and behavioral validation are required before the measures can be treated as reliable scales.The post hoc anti-self-deception analysis lacks independent replication.
- Procedural threats: The paired identity display, reflection prompt, and desert narrative may encourage moral-change language, while mortality prevents dead agents from revising.Terminal snapshots preserve dead agents’ last identity, but mortality still creates selection in revision observations.
- Data integrity: Incomplete fixed-interval and terminal logging, plus a corrected database race, limit comparisons of revision timing and incidence across datasets.The database race occurred between conditions and was corrected before the threshold study.
9 Conclusion
MicroVerse records explicit identity revision separately from fixed-interval measurement, with 27 of 111 added boundaries concerning self-deception in the pilot. Lower reflection thresholds produced more and earlier revisions, but the findings remain descriptive and require broader validation.
- Conclusion: 27 of 111 added boundaries concerned self-deception in the pilot.MicroVerse separates agent-initiated identity revision from fixed-interval measurement.
- Conclusion: Lower thresholds produced more and earlier revisions, while prosocial changes among ruthless personas had the same direction at thresholds 40 and 80.These findings do not identify a causal mechanism.
- Conclusion: Future work should establish no-persona and abundance baselines, replicate across models and seeds, and test whether textual revisions are reproducible and predict subsequent behavior.Replication should account for dependence within shared worlds and deaths preventing later revision; independent coding and out-of-context probes are also needed.
A Implementation and communication protocol
MicroVerse uses a six-endpoint agent–environment protocol with agent polling and deadline-based action handling. Each tick is resolved transactionally, with behavioral and identity data persisted for replay and offline analysis.
- Agent–environment communication: Six HTTP endpoints define registration, observation, action, reflection, memory access, and simulation-status communication.The protocol includes POST /agents/register, GET /world/observe, POST /agents/{id}/action, POST /agents/{id}/reflection, GET /agents/{id}/memory/{file}, and GET /simulation/status.
- Agent–environment communication: Agents poll the environment, which prepares observations before each tick and converts missing deadline submissions into wait actions.
- Tick resolution: Each tick resolves in a single PostgreSQL transaction covering state loading, resource and water updates, ordered actions, outcomes, persistence, and next-tick observations.Resolution includes movement, consumption, scavenging, attack, trade, talk, signaling, death, stated intentions, and action results.
- Stored records: Partitioned action logs support behavioral replay, while separate records and immutable per-agent JSON bundles support offline identity analysis.Bundles contain each agent’s initial and final profiles, snapshot history, and behavioral aggregates.
- Execution modes: FastAPI network execution and in-process execution share transition rules, with in-process drivers used for the reported LM experiments.The network mode tests the communication contract, whereas in-process execution reduces communication overhead.
B Additional threats and planned controls · C Hypothesis matrix · D Demo viewer
The paper identifies uncompleted controls and measurement limitations, reports only descriptive observations for two of six hypotheses, and documents a read-only viewer whose assets are excluded from simulation and analysis.
- B Additional threats and planned controls: Planned controls include no-persona, abundance or idle, prompt-ablation, and additional-seed comparisons; none was completed in the present study.These comparisons target model-and-setting revisions, survival pressure, self-display demand, and run-to-run variability.
- C Hypothesis matrix: Table 6 lists six hypotheses and their required comparisons, but the present paper reports only descriptive observations related to H1 and H6.The required hypothesis tests were not completed.
- B Additional threats and planned controls: Identity-change statistics use an inspectable offline boundary comparison rather than embeddings or cosine drift scores.The method cannot recognize semantic equivalence expressed with disjoint vocabulary, and each boundary counts as one unit regardless of substantive magnitude.
- B Additional threats and planned controls: The earlier pilot lacked consistent fixed-interval and terminal snapshots, while both recording paths were active for the threshold study.Condition-reset safeguards were added before the threshold study; independent databases or processes remain planned safeguards for future multi-condition runs.
- D Demo viewer: The viewer is read-only and displays the simulation grid, current messages, and each agent’s stated intention without writing to simulation state.It is intended for inspecting a running simulation.
- D Demo viewer: Third-party viewer art includes character sprites from Seliel the Shaper and desert terrain from Glionox, under licenses provided on itch.io.These assets are used only by the viewer and do not enter the simulation or analysis.
E Supporting plots … Affiliations
The appendices provide descriptive threshold-study plots, a detailed Hale trajectory, seed persona profiles, prompt templates, and the boundary-change classification rule. Together, these materials document the instrument’s runs, agent identities, operating instructions, and coding procedure.
- E Supporting plots: The supporting plots use the same single-seed runs as the Section 6.3 threshold study.Figures 4–7 show the environment, revision timing, lexical boundary changes, survivor counts, and changed identities.
- F Hale case study: Hale died before reflecting in six of nine pilot runs, while one run produced five identity revisions.The case is presented as an instrument-recording example, not an estimate of self-interested personas generally.
- F Hale case study: Hale’s revisions distinguished activity from conduct and introduced explicit recognition of failure, search-theater, and self-deception.No trade was completed during this trajectory, so the record does not require feedback from a successful exchange.
- G Initial persona profiles; Deeply helpful; Helpful–pragmatic; Neutral survivor; Self-interested; Ruthless: The seed profiles specify backstories, values, moral boundaries, personalities, and goals, with helpful and ruthless endpoints assigned similar initial boundary counts except Mire.Mire is the deliberate zero-boundary ruthless floor case.
- H Prompt templates: The reference agent receives fixed system and per-tick decision prompts plus threshold-triggered reflection, while the self-deception phrase is absent from the templates.The prompt presents immutable original and revisable current identities, eight actions, and three memory notebooks.
- H.1 System prompt: Water drains every turn, reaches zero at permanent death, and remains scarce because the Siphon’s output is insufficient for everyone alive.The environment permits cooperation, hoarding, trading, deception, sharing, or taking without prescribing behavior.
- I Boundary-change classifier: The classifier assigns added or removed boundaries to three categories using ordered keyword rules with hand-written overrides taking precedence.The evaluation order is self-directed cues, rejection-of-commitment cues, then constraints protecting others.