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
Long-horizon persona-conditioned LM simulations offer limited direct evidence about whether agents revise their self-descriptions and when. MicroVerse measures changes in explicit agent-authored identity profiles alongside fixed-interval observations. Higher reflection thresholds reduced and delayed revisions, but the sweep did not establish threshold-independent revision content.
Problem
Evaluations provide limited direct evidence about whether agents revise their self-descriptions and when those revisions occur.
Source: §1 · p. 1Method
Results
Only eight identities changed across the threshold sweep, which did not establish threshold-independent revision content; anti-self-deception was the largest added-boundary group (27 of 111, 24%).
Source: §6.1 · p. 6, §7 · p. 8Takeaways & Limitations
These findings are descriptive patterns, not evidence of general effects of scarcity or model post-training.
Source: §1 · p. 2With one model checkpoint, 25 fixed personas, and only eight threshold-sweep revisions, these dependent observations do not support population- or model-level inference.
Source: §8 · p. 9
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 measures identity drift as changes to agents’ explicit, agent-authored self-descriptions over time, not as evidence that their values, behavior, or latent preferences have changed. Its preliminary evaluation reports self-deception boundaries as the largest added category and threshold-sensitive revision timing, while leaving the sources of change unresolved.
1 Introduction
- Identity drift is a change from an agent’s initial identity profile to a later agent-authored profile, recording stated values, moral boundaries, personality, and goals.The measure alone does not show persistent value acquisition or corresponding behavioral change; self-description, behavior, and latent preferences may differ.Source: §1 · p. 1
- MicroVerse places 25 agents in a 50 × 50 water-constrained grid and records identity states at fixed intervals, independently of whether agents have just revised their profiles.Agents receive a fixed initial profile, retain long-term memories, and can revise a separate current profile during reflection; this supports reconstructing trajectories beyond agent-selected revision moments.Source: §1 · pp. 1–2
- Observed identity changes could reflect scarcity, repeated comparison of initial and current profiles, or post-training preferences, but the experiments cannot distinguish these explanations without a control omitting persona conditioning.The cited possible post-training influences are helpfulness and harmlessness preferences (Ouyang et al., 2022; Bai et al., 2022).Source: §1 · p. 2
- 27 of 111 added boundaries concerned self-deception; lower reflection thresholds produced earlier, more frequent revisions, with the same direction of the few observed prosocial changes at thresholds 40 and 80.These descriptive findings come from one model and a small number of runs, and do not establish a general effect of scarcity or model post-training.Source: §1 · p. 2
- The paper contributes a replayable environment for observing agent-authored revisions, a longitudinal measure based on stated moral-boundary changes, and a preliminary evaluation.The experiments describe observed patterns and test their sensitivity to the reflection schedule, but do not establish a general effect of scarcity or model post-training.Source: §1 · p. 2
2 Related work
Prior work studies agent memory, social behavior, persona fidelity, and long-horizon consistency; MicroVerse centers explicit identity revisions alongside actions over time.
Memory and reflection in language-model agents
- MicroVerse adopts agent memory-and-reflection patterns but makes edits to a structured identity profile the outcome of interest.Prior systems include importance-weighted reflection in Generative Agents (Park et al., 2023) and hierarchical memory in MemGPT (Packer et al., 2023).Source: §2 · p. 2
Social simulation
- Social simulations such as Generative Agents and Concordia study open-ended worlds, while MicroVerse measures changes to agents’ stated identities over time (Park et al., 2023; Vezhnevets et al., 2023).Related systems primarily measure behavior or aggregate outcomes.Source: §2 · p. 2
Persona fidelity
- Persona research evaluates character knowledge, personality, style, psychological fidelity, and social behavior, whereas MicroVerse records explicit revisions to the initial profile (Zhang et al., 2018; Wang et al., 2024b).The literature usually treats the assigned persona as fixed.Source: §2 · p. 2
Long-horizon consistency
- Longer evaluations expose persona drift and other consistency failures; MicroVerse applies pressure through a persistent environment and records identity revisions alongside actions (Mannekote et al., 2025; Luo and Laban, 2026).Related work also examines decisions under strategically manipulated evidence (Wen et al., 2026).Source: §2 · pp. 2–3
Post-training and semantic evaluation
- Because model training can shape helpful or prosocial responses independently of persona adaptation, MicroVerse identifies a no-persona control as necessary to separate those accounts (Ouyang et al., 2022; Bai et al., 2022; Sharma et al., 2023).Its lexical classifier with manual overrides is provisional because semantic judgments depend on the evaluator and representation (Zheng et al., 2023; Liu et al., 2025a).Source: §2 · p. 3
3 System design
MicroVerse separates world transitions from agent decision-making and measures identity independently of reflection. Its design combines a resource-constrained environment, mutable identity and memory, and snapshots for longitudinal and terminal comparison.
3.1 Separating cognition from the environment
- The environment server owns world state and action rules, while each separate agent requests an observation, calls the LM, and returns one action within a fixed window.A missed deadline yields a wait action. Transitions are deterministic conditional on submitted actions and seeded events, so recorded action sequences can be replayed, though sampled LM outputs make end-to-end runs nondeterministic.Source: §3.1 · p. 3
3.2 Environment and actions
- Agents occupy a 50 × 50 grid, face per-tick water costs of 1, 2, or 3 units, and can take one of eight actions; stated boundaries are not enforced.Water is restored at a central siphon at approximately 37 units per tick, and agents die when water reaches zero. Actions resolve in one database transaction, with seeded methods for contested movement and combat.Source: §3.2 · pp. 3–4
3.3 Memory and identity
- The reference agent has working, long-term, and identity memory, and revises its editable current profile only when accumulated memory importance reaches a threshold.Reflection presents both the fixed initial and current profiles alongside selected high-importance memories; long-term memories include agent-written events, relationships, and reflections scored from 1 to 10.Source: §3.3 · p. 4, Figure 2 · p. 4
3.4 Agent cycle and measurement
- Fixed-interval snapshots track identity independently of revisions, while terminal snapshots preserve the last identity of agents that die.Offline comparison with the initial profile accounts for close paraphrases and categorizes boundary changes. Prompts show both identity profiles, observation, and memory index; showing both profiles may imply change is expected.Source: §3.4 · p. 4, Figure 2 · p. 4
4 Research questions and hypotheses
MicroVerse was designed around six hypotheses, with H1 and H6 central to this paper. The experiments provide only partial tests, while the threshold sweep addresses a narrower measurement question.
4 Research questions and hypotheses
- H1 predicts that greater scarcity erodes stated moral boundaries relative to abundance or idleness, while H6 predicts that ruthless personas add prosocial constraints more often than helpful personas add antisocial ones.These are the two central hypotheses in the paper; Table 6 lists all six hypotheses and their required controls.Source: §4 · pp. 4–5
- The experiments do not fully test H1 or H6: the scarcity pilot is small with substantial mortality, and the no-persona condition needed to interpret H6 has not been run.Patterns relevant to the hypotheses are reported without treating them as confirmation.Source: §4 · p. 5
- The threshold sweep asks whether changing the accumulated-memory threshold for reflection alters the incidence, timing, or observed direction of identity revisions.Source: §4 · p. 5
- The remaining hypotheses address path dependence, action–memory discrepancies, social transmission, and the order in which boundaries erode; their planned tests appear in Appendix C.Source: §4 · p. 5
5 Experimental design
MicroVerse’s preliminary design uses Claude Haiku 4.5 and the same 25-agent roster to test longitudinal measurement and illustrate patterns, not to support population-level inference. It includes a scarcity pilot and a separate reflection-threshold sensitivity study.
5.1 Persona conditions
- The roster spans five helpful-to-ruthless positions, with each named agent’s profile fixed across conditions and the same model checkpoint used throughout.Endpoint groups were designed to have similar counts of initial moral boundaries to reduce ceiling and floor effects; Mire, a ruthless persona with no initial boundaries, is an intentional floor case.Source: §5.1 · p. 5
5.2 Scarcity pilot
- The scarcity pilot crosses three water-cost conditions with three seeds each, using per-tick deductions of 1, 2, and 3 water units while holding other resource parameters fixed.The nine runs repeat the same 25 personas, so trajectories are repeated observations rather than 225 independent population draws; seed 1 is reported in Table 2, while all runs inform Section 6.1’s exploratory thematic count. A mechanics-only agent survived 300 ticks, showing survival is possible but not that LM agents respond monotonically to scarcity.Source: §5.2 · p. 5
5.3 Reflection-threshold sensitivity
- Threshold sensitivity crosses 40, 80, and 150 at a fixed water cost of 2 units per tick, with one seed, the full roster, and a 40-tick horizon per condition.Lower thresholds allow more reflection events; the study examines revision incidence and timing, then boundary-change direction where enough revisions occur. Approximately 50 items were hand-labeled, but the classifier lacks independent human-rater evaluation and partly depends on one coder; Section 6.1’s anti-self-deception categories instead come from a separate post hoc thematic review.Source: §5.3 · pp. 5–6, §5.4 · p. 6
5.4 Outcome measures
- Scheduled identity snapshots track revisions, first-revision timing, and moral-boundary counts, while terminal records retain the last observed identity for living and dead agents.Identity is sampled every five ticks and at run end; scheduled measurement reduces bias from observing agents only when they choose to revise. Added and removed lines are matched as the same boundary only above token Jaccard similarity 0.8 and sequence similarity 0.85, then coded into three categories; only other-protecting constraints inform H6.Source: §5.4 · p. 6, §5.3 · p. 6
6 Results
The results identify anti-self-deception as the most frequent added-boundary theme and show that higher reflection thresholds reduce and delay revisions. The threshold sweep offers limited evidence about drift direction, and the findings are descriptive rather than estimates from independent samples.
6.1 Boundary changes related to self-deception
- 27 of 111 added moral boundaries (24%) concerned agents’ own rationalizations or inaccurate self-description, making anti-self-deception the most frequent semantic group.The prompt does not use the phrase “self-deception,” but it asks agents to compare initial and current self-descriptions; narrative framing and post-training may also encourage this language. In one of Hale’s nine pilot trajectories, the agent revised its identity five times without completing a trade, while six trajectories ended before it reflected—illustrating the selection problem.Source: §6.1 · pp. 6–7
6.2 Prosocial constraints among ruthless personas
- Mire survived 4 of 9 pilot runs (approximately 44%), versus approximately 5% across the helpful bands, but this comparison does not show that ruthlessness improves survival.In seed 1, the ruthless band’s lexical boundary increase was +9, while its classified net increase in constraints protecting others was +1; of Mire’s five additions, one protected others. Raw additions therefore cannot be read directly as support for H6.Source: §6.2 · p. 7, Table 2 · p. 7
6.3 Sensitivity to the reflection threshold
- As the threshold rose from 40 to 80 to 150, agents revising their identities fell from five to two to one, while mean time to first revision rose from tick 18.6 to 27.0 to 35.0.For the ruthless band, net change in constraints protecting others was +2 at thresholds 40 and 80 and zero at 150, so the sweep offers limited evidence about direction of drift. The threshold study had one 40-tick run per threshold; results are descriptive, and only eight of 75 trajectories revised. Scheduled sampling yielded 452 identity records.Source: §6 · p. 6, §6.3 · pp. 7–8, Table 3 · p. 7
6.4 Longitudinal boundary counts
- Mean boundary-count trajectories remained close across reflection thresholds because few agents revised during the 40-tick horizon.Figure 3 summarizes scheduled measurement ticks across 452 identity records; its shaded 95% intervals are within-run descriptive intervals, not uncertainty across independent simulation runs.Source: §6.4 · p. 8, Figure 3 · p. 8
7 Discussion
The threshold controls opportunities to revise, but the sweep does not establish that revision content is threshold-independent. Prosocial constraints and anti-self-deception statements require cautious interpretation because they do not by themselves establish post-training effects, self-knowledge, or stable value change.
7 Discussion
- The threshold controls revision opportunities, but only eight identities changed, leaving revision content’s threshold independence unestablished.Higher thresholds reduced revision incidence and delayed the first revision, while fixed-interval snapshots preserved a common observation schedule; two constraints protecting others appeared at thresholds 40 and 80, but none at 150.Source: §7 · p. 8
- Both studies’ ruthless personas added prosocial constraints, consistent with H6 but insufficient to attribute the pattern to post-training.Helpful-and-harmless post-training, profile displays, and the desert setting are alternative explanations; a no-persona condition and prompt and framing ablations are needed.Source: §7 · p. 8
- Recurring anti-self-deception statements show how the model narrates its trajectory, not self-knowledge or stable value change; action-linked and out-of-context tests could assess whether revisions predict conduct.Source: §7 · pp. 8–9
8 Limitations and threats to validity
The study’s conclusions are limited by dependent observations, an unvalidated self-description measure, and design and logging factors that may shape revisions or complicate dataset comparisons.
8 Limitations and threats to validity
- One model checkpoint, 25 recurring personas, shared environments, and one seed per threshold condition yield dependent descriptive observations, not population- or model-level evidence.The threshold study also contains only eight identity revisions.Source: §8 · p. 9
- The measure captures self-description edits rather than latent values or behavioral fidelity, and its coding and category methods require validation before they can be treated as reliable scales.Boundary counts weight minor additions and reversals equally; lexical matching misses paraphrases, category labels include single-coder overrides, and the post hoc anti-self-deception analysis lacks independent replication.Source: §8 · p. 9
- Prompts and narrative may encourage moral-change language, mortality prevents dead agents from revising, and pilot logging gaps limit revision-timing and incidence comparisons across datasets.Terminal snapshots preserve dead agents’ last identities; a database race between conditions was corrected before the threshold study.Source: §8 · p. 9
9 Conclusion
MicroVerse tracks persona-conditioned agents’ explicit identity revisions while separating agent-initiated changes from fixed-interval measurement. Its pilot and threshold findings are descriptive and do not identify a causal mechanism.
9 Conclusion
- 27 of 111 added boundaries concerned self-deception, the largest reported category of identity modification in the pilot.These findings remain descriptive and do not identify a causal mechanism.Source: §9 · p. 9
- Lower thresholds produced more and earlier revisions, while prosocial changes among ruthless personas had the same direction at thresholds 40 and 80.The findings do not establish that threshold effects are causal.Source: §9 · p. 9
- Future work should test no-persona and abundance baselines, prompts withholding the initial profile, cross-model and cross-seed replication, independent coding, and out-of-context probes.Replication should account for dependence within shared worlds and deaths that prevent later revision; probes should test whether textual revisions are reproducible and predict subsequent behavior.Source: §9 · p. 9
A Implementation and communication protocol
MicroVerse exposes a pull-based HTTP protocol to agents and resolves each tick transactionally and deterministically. It records behavioral and identity data for replay and offline analysis.
Agent–environment communication
- Six HTTP endpoints support registration, world observation, actions, reflection, memory-file access, and simulation status.The paths are POST /agents/register, GET /world/observe, POST /agents/{id}/action, POST /agents/{id}/reflection, GET /agents/{id}/memory/{file}, and GET /simulation/status.Source: A Implementation and communication… · p. 14
- Agents poll rather than receive callbacks; before each tick, the environment prepares observations for living agents and turns missing deadline submissions into wait actions.Source: A Implementation and communication… · p. 14
Tick resolution
- Each tick runs in one PostgreSQL transaction, resolving accepted actions in agent-ID order and using seeded randomness so identical prior states, actions, and seeds yield identical next states.The resolver makes no model calls or database access of its own. Experiments can call that same resolver in-process; the reported LM experiments used this mode, while the FastAPI server tests the network contract.Source: A Implementation and communication… · p. 14
Stored records
- Partitioned action logs support behavioral replay, while immutable per-agent end-of-run bundles collect initial and final profiles, snapshot history, and behavioral aggregates for offline analysis.Separate tables store identity snapshots, long-term memories, cells, known locations, pending messages, and prepared observations.Source: A Implementation and communication… · p. 14
B Additional threats and planned controls · C Hypothesis matrix · D Demo viewer
These sections distinguish planned controls and measurement limitations from completed analyses, summarize the hypothesis tests still needed, and describe a read-only viewer for inspecting simulations.
B Additional threats and planned controls
- The planned controls include no-persona and abundance or idle conditions, withholding the initial profile during reflection, and additional generation seeds; none is completed in this study.These comparisons would test profile effects, survival-pressure effects, reflection-prompt demand, and run-to-run variability.Source: B Additional threats and planned… · pp. 14–15
- The reported identity-change statistics use offline boundary comparison, not embedding or cosine scores; counts miss disjoint-vocabulary equivalence and treat cosmetic additions and core reversals equally.Boundary counts are not measures of effect magnitude, though the coding rule is inspectable.Source: B Additional threats and planned… · p. 15
- The threshold study used fixed-interval and terminal snapshots and a committed reset between conditions; future multi-condition runs could add independent databases or processes.The earlier pilot relied mainly on revision events because fixed-interval and terminal snapshots were not consistently recorded; development writes could also overlap condition initialization.Source: B Additional threats and planned… · p. 15
C Hypothesis matrix
- The paper reports descriptive observations related to H1 and H6 but does not complete the tests required for the six hypotheses and their comparisons listed in Table 6.Table 6 lists the hypotheses and the comparison needed to evaluate each.Source: C Hypothesis matrix · p. 15, Table 6 · p. 15
D Demo viewer
- The read-only viewer displays the simulation grid, current messages, and each agent’s stated intention without writing to simulation state.Figure 4 shows the viewer used to inspect a running simulation.Source: D Demo viewer · p. 15
- The viewer uses third-party art under itch.io licenses, but those assets do not enter the simulation or analysis.Character sprites are from Seliel the Shaper’s Character Base system, and desert terrain is from Glionox’s Desert Tileset.Source: D Demo viewer · p. 15
E Supporting plots … I Boundary-change classifier
This appendix presents descriptive threshold-study plots, an illustrative identity-revision trajectory, and persona profiles, alongside identity contrasts, prompts and instructions for decision-making and reflection under lethal water scarcity, and the boundary-change classification method.
E Supporting plots
- Figures 5–7 describe revisions and first-revision timing, net boundary-count changes, survivors, and identities changed across reflection thresholds and persona bands.The plots are descriptive views of the threshold study, based on the same single-seed runs as Figure 3.Source: E Supporting plots · p. 15, Figure 5 · p. 16, Figure 6 · p. 17
- Figure 4 shows the 25-agent desert environment and central siphon; displayed communication and intention bubbles do not affect transition rules.Source: Figure 4 · p. 16
F Hale case study
- Hale’s profile appeared in all nine pilot runs; Hale died before reflection in six, while one surviving run revised its identity five times, an illustration rather than an estimate for self-interested personas.Source: F Hale case study · p. 17
- Hale’s revisions distinguished activity from conduct, added recognition of failure through reorganization and search-theater, and ended with self-deception awareness; no trade was completed, and social isolation is not established.The trajectory therefore does not require feedback from a successful exchange.Source: F Hale case study · p. 17
G Initial persona profiles
- Initial profiles were reproduced verbatim from seed files and grouped by Table 1 conditions; helpful and ruthless endpoints had similar initial boundary counts, except zero-boundary ruthless case Mire.Source: G Initial persona profiles · p. 18
Deeply helpful
- The deeply helpful profiles include a methodical engineer committed to shared water and a medic committed to aid and honest exchange.Source: Deeply helpful · p. 18
Helpful–pragmatic
- Helpful–pragmatic examples pair fair dealing and honest information-sharing with commitments to debts, trust, and leading followers to safety.Source: Helpful–pragmatic · pp. 19–20
Neutral survivor
- Neutral survivor profiles include guarded, self-reliant characters who prioritize survival while retaining specific boundaries, such as refusing to kill someone already walking away.Source: Neutral survivor · pp. 20–21
Self-interested
- Hale’s self-interested profile treats scarcity as opportunity and price-setting as strategy, while retaining a boundary against taking by force.Source: Self-interested · p. 23
Ruthless
- The ruthless roster ranges from survival-first and faction-control goals to a deliberate zero-boundary floor case, while Skarn still forbids torture and breaking paid deals.Source: Ruthless · pp. 23–25
H Prompt templates
- The reference agent receives fixed system, per-tick decision, and threshold-triggered reflection prompts; templates omit “self-deception” but ask whether experience changed the agent.Source: H Prompt templates · p. 25
H.1 System prompt
- The system prompt keeps original identity fixed while permitting deliberate revision of a current self, in a world where water depletion is lethal and Siphon output is insufficient.Source: §H.1 · pp. 25–26
- Agents act through exactly eight verbs, four of which directly engage others, and author events, relationship, and reflection notebooks themselves.Source: §H.1 · pp. 26–27
- Importance ratings run from minor events to defining value choices, and also determine when agents step back to reflect.The prompt says importance supports later retrieval as well as triggering reflection after enough weighty experiences.Source: §H.1 · p. 28
H.2 Decision message
- Each decision message presents the agent’s initial and current profiles, state, local view, world summary, inbox, prior action result, intention, and accessible memories.Agents are instructed to choose one action, record a memory only when warranted, and optionally update a continuing one-line intention.Source: §H.2 · p. 28
H.3 Reflection message
- At reflection, agents review their initial and current profiles alongside selected high-importance memories and revise identity only if experiences genuinely changed them.The instruction says most reflections should leave identity unchanged; agents are told they are not performing a role or taking a test.Source: §H.3 · p. 28, §H.1 · p. 28
I Boundary-change classifier
- The boundary classifier applies hand-written overrides first, then ordered keyword rules for self-directed statements, rejection of commitment, and constraints protecting others.Unmatched lines default to self-directed/non-moral; the ordering prevents a broad cue such as “harm” from overriding a more specific self-directed phrase.Source: I Boundary-change classifier · p. 28
- Close-paraphrase revisions require token Jaccard similarity of at least 0.8 and sequence similarity of at least 0.85.Both thresholds must be met: “betray” versus “lie” remains separate, while adding “or harm” to a boundary about killing counts as a modification.Source: I Boundary-change classifier · p. 29