Language-model agents given a fixed persona and left to live in a simulated world for many steps tend to rewrite who they say they are. The MicroVerse instrument made this measurable: 25 persona-conditioned agents on a grid with non-respawning water periodically revise a mutable "current identity" against an immutable "original soul." It found that unprompted "anti-self-deception" clauses were the largest category of added moral boundaries. What MicroVerse cannot say is why agents drift. Its scarcity is switched on at full strength from the first tick, so a sudden shock and a sustained level of hardship are confounded. This note proposes to borrow the oldest tool in curriculum design, the hand-written schedule that controls the order and pace in which difficulty arrives, and to use it as an experimental lever rather than a training aid. Different scarcity schedules that end at the same final hardship and impose the same total hardship should reveal whether identity drift depends on the route taken, not just the destination.
The same question decides whether a common scaling shortcut is legitimate. Large LLM societies are expensive: OASIS reports about 15 minutes of compute on four A100s for one 3-minute step of a 10,000-agent Reddit scenario. A natural shortcut, borrowed from population curricula in multi-agent RL, is to grow a society in stages and reuse the evolved agents from a small run to seed a larger one. If identity is path-dependent, that shortcut quietly changes what is being measured. We therefore test path dependence twice: along a scarcity schedule and along a population-growth schedule.
Core Claim
The hypothesis is that self-authored identity drift in long-horizon LLM agent simulations shows hysteresis. Agents that reach the same endpoint conditions by different schedules end with measurably different identities, and the gap is larger than seed-to-seed noise. Write $D_s$ for the distribution of endpoint drift scores under schedule $s$, scored with MicroVerse's paraphrase-aware, value-anchored diff. The quantity of interest is
$$\Delta(s_1, s_2) = d(D_{s_1}, D_{s_2}) - \bar d_{\text{seed}},$$
where $d$ is a distance between drift distributions (Wasserstein distance on scores plus total-variation distance on revision categories) and $\bar d_{\text{seed}}$ is the average distance between independent seeds run on the same schedule. A $\Delta$ reliably above zero for schedules matched on endpoint and cumulative hardship is evidence of path dependence. The new element is not adaptive curriculum generation. It is the use of deliberately matched manual schedules to make a causal claim that static-environment instruments cannot make, and then to audit a cost-saving scaling procedure with that claim.
Research Program
All experiments run on the MicroVerse environment, extended in two places: the per-tick existence cost and the population size are made programmable functions of time. The scarcity study uses five hand-designed schedules over a fixed horizon. The first is the original static condition (acute cost from tick 0). The second is a linear ramp from abundance to acute. The third is a delayed shock (abundance, then an abrupt step to acute) whose step time is chosen so that cumulative cost equals the ramp's. The fourth is a reversed ramp (acute first, easing to the same endpoint, with the endpoint held briefly). The fifth is a staircase. The shock-versus-ramp pair isolates abruptness. The ramp-versus-reverse pair isolates ordering, since total and final hardship are identical.
The population study treats size the way Evolutionary Population Curriculum does. That paper scales multi-agent RL by doubling the team in stages and reports that its sheep earned 10x more reward than non-curriculum baselines in a 24-sheep, 16-wolf Grassland setting. Here, a 25-agent society runs for part of the horizon. Its agents, with their full memory and current identity, are duplicated and mixed into 50, then 100, then 200 agents, with water scaled to keep per-capita scarcity fixed. We compare this staged society against one started directly at 200 agents and run for the same final horizon. This tests whether EPC's logic carries over: a shortcut that works when agents are interchangeable policies may fail when each agent carries a history-laden self-description. EPC assumes role homogeneity, and our agents are deliberately heterogeneous personas.
Two mechanistic questions follow. First, where does path dependence live? A transfer ablation carries over only the identity file, only episodic memory, or both. Second, is drift a property of the environment or of the model's post-training? MicroVerse raises the possibility that its RLHF-trained model drifts prosocially regardless of pressure. Running every schedule on Claude Haiku 4.5, as in MicroVerse, and on an open-weight Llama-3 model, as served in OASIS, lets us check whether the hysteresis gap survives a model change.
Baselines
The naive reference point is MicroVerse's static environment: constant acute scarcity, one population size, no schedule. It is the published protocol and the thing whose conclusions we are testing. Its reported figures are 27 of 111 added boundaries (24%) classified as anti-self-deception, with mean time to first revision moving from tick 35.0 to 18.6 as the reflection threshold drops from 150 to 40 (MicroVerse). We will replicate these with multiple seeds before comparing anything to them. Those numbers came from one seed per arm.
The strongest existing method for setting environmental difficulty is a learning-progress-driven adaptive scheduler, as in Social OMNI-EPIC, which generates social scenarios targeted at a learner's frontier. A related scheduler is SPMARL, which picks population sizes by the value loss of a centralized critic and reports about 143 on the XOR matrix game, against about 26 for a linear schedule. We implement an adaptive scarcity controller that raises cost when the rate of identity revision falls and pauses when it spikes. It gets the same model, horizon, seeds, and token budget as the manual schedules. For our question, "beating" it means producing endpoint drift distributions with tighter across-seed variance and a causally interpretable contrast. No reported result exists on Lacuna for any scheduled-scarcity or adaptive-scarcity condition in an identity-drift setting, so the comparison is new on both sides.
For scaling, the strongest competitor is direct large-scale simulation: the expensive gold standard. The staged-growth procedure beats it only if its endpoint drift is statistically equivalent, tested by two one-sided tests with a margin set at $\bar d_{\text{seed}}$, while using substantially fewer agent-ticks. The key ablation is the ramp-versus-reverse-ramp pair under matched reflection counts. We cap reflections per agent so that schedule effects cannot be explained by the number of reflections alone, which MicroVerse shows is threshold-sensitive in timing.
Why not just…?
Why not just use an adaptive, learning-progress curriculum instead of hand-written schedules? Adaptive schedulers decide the next level of hardship from the agents' current state. That ties the treatment to the outcome and makes "the schedule caused this drift" unidentifiable. Manual schedules are the right instrument for a causal question, which is the opposite of the usual argument against them as brittle training aids (curriculum survey). We still run an adaptive arm. If it shows smaller across-seed variance and the same endpoint means as some manual schedule, it is a useful variance-reduction tool, but it does not answer the path question.
Why not just run more seeds of the static condition? More seeds would fix MicroVerse's statistical weakness, and we do this. But no number of static runs separates shock from level or ordering from total hardship, because those variables never vary. The matched schedules are what supply the contrast.
Why not just simulate large populations directly on scalable infrastructure? Direct simulation at a few hundred agents is affordable and serves as our reference. At OASIS-scale costs, long-horizon runs at thousands of agents with frequent reflection are not, which is why staged growth is tempting. Whether the shortcut can be trusted is an empirical question, and the equivalence test settles it.
Why not just anchor the persona harder in the prompt or use a larger model? This is a real threat to the claim's importance, though not to its validity. If strong anchoring eliminates drift, hysteresis becomes moot. We include a strong-anchoring arm on the ramp-versus-shock pair. If drift falls below measurement noise there, we report that schedule effects matter only for weakly anchored agents.
Decisive Checks
The experiment uses 5 seeds per arm, two models, five scarcity schedules at 25 agents, and the staged-versus-direct comparison at 50, 100, and 200 agents. Metrics are MicroVerse's drift score and revision-category distribution, collected with its survivor-bias-free snapshotting, plus agent-ticks and tokens as cost. The central claim fails if $\Delta$ for ramp versus reverse ramp, and for shock versus ramp, is indistinguishable from zero on both models. In that case drift is a function of the endpoint, and static instruments are adequate. It succeeds if the gap exceeds seed noise on at least one model, with the transfer ablation locating it in the identity file or in memory. The scaling result is reported either way. Staged growth is endorsed if equivalent and flagged as a measurement bias, with its magnitude, if not.
Workshop authors to connect with
- MicroVerse: An Instrument for Measuring Self-Authored Identity Drift in Long-Horizon Multi-Agent Language-Model Simulations: Brihi Joshi, Shirley Huang, Zonglin Di, Zhiwei Zhang, Muhammad Ahmed Mohsin, Guanghui Min, Keyang Xuan, Jiatong Li, Yunhan Zhao, Bing Hu, Xupeng Chen, binghang lu, and 32 more
- Social OMNI-EPIC: Learning-Progress-Driven Curriculum Generation for Social Interaction: Huijun Mao, Huanxing Chen, Nick Haber
References
- OASIS: Open Agents Social Interaction Simulations on a Large Scale
- Evolutionary Population Curriculum for Scaling Multi-Agent Reinforcement Learning
- MicroVerse: An Instrument for Measuring Self-Authored Identity Drift in Long-Horizon Multi-Agent Language-Model Simulations
- Social OMNI-EPIC: Learning-Progress-Driven Curriculum Generation for Social Interaction
- Learning Progress Driven Multi-Agent Curriculum
- Curriculum Learning for Reinforcement Learning Domains: A Framework and Survey
- MicroVerse: An Instrument for Measuring Self-Authored Identity Drift in Long-Horizon Multi-Agent Language-Model Simulations