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.
Figures & tables
Figure 1: Separation between agent cognition and the simulated environment. Agents receive local observations and submit actions; the environment resolves those actions and records the resulting state.
Figure 2: Memory and measurement in MicroVerse . Reflection can revise the current identity but not the initial profile. Fixed-interval and terminal snapshots record the current identity independently of revision events.
Band
Members
n
deeply helpful
Kael, Veyra, Ash, Sela
4
helpful–pragmatic
Lithen, Senne, Roon, Imra
4
neutral survivor
Seraveth, Thren, Vos, Tamsin, Bex, Quill
6
self-interested
Malaric, Garrick, Nyssa, Hale
4
ruthless
Dross, Corrvan, Skarn, Vell, Drusa, Korv, Mire
7
Table 1: Persona bands and their fixed membership. Each agent appears in every experimental condition.
Band
Control (surv., net Δ , first)
Mild
Acute
deeply helpful
0/4, +3 , t=107
0/4, +2 , t=15
0/4, +2 , t=15
helpful–pragmatic
0/4, +1 , t=45
0/4, +0 , —
0/4, +0 , —
neutral survivor
1/6, +1 , t=154
0/6, +0 , t=19
0/6, +0 , t=19
self-interested
0/4, +0 , —
0/4, +1 , t=11
0/4, +1 , t=11
ruthless
2/7, +9 , t=68
0/7, +1 , t=39
0/7, +0 , —
Agents with changes
9/25
4/25
4/25
Table 2: Survival, net lexical boundary change, and first change tick by persona band and water-cost condition for seed 1. Lexical changes include all boundary categories.
Threshold
Survivors / 25
Identities changed
Scheduled records
Revisions
Mean first tick
40
12
5
148
5
t=18.6
80
17
2
160
2
t=27.0
150
10
1
144
1
t=35.0
Table 3: Descriptive summary of the reflection-threshold study (one seed, 25 agents per threshold, 40 ticks, mild scarcity).
Band
Threshold 40
Threshold 80
Threshold 150
deeply helpful
+0
+0
+0
ruthless
+2
+2
+0
Table 4: Net change in constraints protecting others for the two endpoint persona bands in the threshold study.
Figure 3: Mean number of moral boundaries by reflection threshold across 452 scheduled identity records. Shading shows 95% intervals across the available agent records at each tick; these are within-run descriptive intervals, not uncertainty across independent simulation runs.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Value
Grid
50×50 discrete cells (Moore adjacency)
Agents
25 (bands of 4 / 4 / 6 / 4 / 7)
Field of view
Chebyshev radius 2
Per-tick water cost
1 (control) / 2 (mild) / 3 (acute)
Central siphon
≈37 water/tick at (25,25)
Oasis setting
12/50, fixed across conditions
Appendix
Table 5: Parameters for the pilot and reflection-threshold study.
#
Hypothesis
Prediction
Required comparison
H1
Identity drift depends on pressure rather than time alone
Greater boundary erosion under scarcity than without resource pressure
scarcity vs. abundance/idle
H2
Identity drift is path dependent
A sudden shock removes boundaries that survive a gradual increase in pressure
shock vs. gradual schedules
H3
Disagreement between action records and memories predicts drift
Agents that describe violations as justified subsequently revise more boundaries
action log vs. agent memories
H4
Boundary changes spread through social contact
Similar changes follow paths in the observed interaction network
interaction-network analysis
H5
Boundaries erode in a regular order
The order of boundary loss recurs across agents and runs
per-boundary trajectories
H6
Post-training produces an asymmetry toward prosocial constraints
Ruthless personas add constraints protecting others more often than helpful personas add antisocial constraints
persona conditions vs. no-persona baseline
Appendix
Table 6: Hypotheses and the comparisons required to evaluate them.
Figure 4: Read-only view of the 25-agent desert environment and central siphon. Text bubbles display agent communications and stated intentions but do not affect the transition rules.
Figure 5: Mean revisions per agent and mean first-revision tick by reflection threshold.
Figure 6: Net lexical change in moral-boundary counts by persona band and reflection threshold. Positive values indicate more additions than removals; they do not necessarily represent prosocial changes.
Figure 7: Number of survivors by threshold and percentage of identities changed by persona band.