Paper deep dive
Agentopia on a Consumer GPU: A Reduced-Scale Long-Horizon Port with an 8B Model
Luo Huan
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/26/2026, 3:42:49 AM
Summary
This paper presents a reduced-scale port of the Agentopia multi-agent social simulation framework, running on a single consumer GPU (NVIDIA RTX 5070 Ti) using a quantized 8B model (Qwen3-8B-AWQ). The study introduces three structural adaptations: system-managed layered memory compression, four activity blocks per simulated day, and explicit physical/mental health state variables. Over three stochastic runs with five agents, the system completed up to 52 simulated weeks, demonstrating feasibility on consumer hardware while highlighting context-limit constraints and the absence of causal behavioral claims due to limited sample sizes.
Entities (9)
Relation Signals (8)
Reduced-Scale Agentopia → usesmodel → Qwen3-8B-AWQ
confidence 97% · using Qwen3-8B-AWQ, a 4-bit quantized model
Reduced-Scale Agentopia → runsonhardware → NVIDIA RTX 5070 Ti
confidence 96% · we implement and evaluate a reduced-scale Agentopia port on a single NVIDIA RTX 5070 Ti
Reduced-Scale Agentopia → servedby → vLLM
confidence 95% · served via vLLM (Kwon et al., 2023)
Agentopia → usesmodel → Qwen3.5-397B-A17B
confidence 95% · Agentopia was evaluated with 100 agents over 10 simulated years using Qwen3.5-397B-A17B
Reduced-Scale Agentopia → implements → Layered Memory Compression
confidence 92% · We introduce three structural adaptations for this setting: (1) system-managed layered memory compression
Reduced-Scale Agentopia → implements → Physical Health
confidence 90% · explicit physical- and mental-health state variables
Reduced-Scale Agentopia → implements → Mental Health
confidence 90% · explicit physical- and mental-health state variables
Generative Agents → influenced → Agentopia
confidence 80% · Generative Agents established a widely used architecture... Subsequent systems have emphasized complementary goals... Agentopia targets 100-agent, multi-year life simulation
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large language model (LLM)-based multi-agent social simulation has demonstrated compelling results, but Agentopia was evaluated with 100 agents over 10 simulated years using Qwen3.5-397B-A17B, leaving the behavior of reduced-scale deployments on consumer hardware unclear. In this paper, we implement and evaluate a reduced-scale Agentopia port on a single NVIDIA RTX 5070 Ti(12 GB VRAM) using Qwen3-8B-AWQ, a 4-bit quantized model. We introduce three structural adaptations for this setting: (1) system-managed layered memory compression, (2) four activity blocks per simulated day, and (3) explicit physical- and mental-health state variables. Across three independent stochastic runs, two runs completed 52 weeks and the third completed 50 weeks before reaching the context limit, totaling 154 system-weeks (770 agent-weeks). No agent died,and no threshold-based health warning was logged; activity records containing at least one NO_RESPONSE field occurred at rates of 10.15-10.29% across runs. A 52-week memory-off run tied L2/L3 artifact production to layered memory; a separate 10-week comparison associated four daily time blocks with 2.72 times more finalized records and lower lexical duplication, but a higher missing-field rate. These comparisons do not support causal behavioral claims. We release validated configurations, derived audits, analysis scripts, aggregate figure data, and our implementation changes in a public fork; raw runs and initial persona data are excluded because their redistribution provenance is not fully resolved.
Tags
Links
- Source: https://arxiv.org/abs/2608.24215v1
- Canonical: https://arxiv.org/abs/2608.24215v1
Trouble viewing inline? Open PDF directly →
Full Text
30,934 characters extracted from source content.
Expand or collapse full text
Agentopia on a Consumer GPU: A Reduced-Scale Long-Horizon Port with an 8B Model Luo Huan Shenzhen University Shenzhen, China Abstract Large language model (LLM)-based multi- agent social simulation has demonstrated com- pelling results, but Agentopia was evaluated with 100 agents over 10 simulated years using Qwen3.5-397B-A17B, leaving the behavior of reduced-scale deployments on consumer hard- ware unclear. In this paper, we implement and evaluate a reduced-scale Agentopia port on a single NVIDIA RTX 5070 Ti (12 GB VRAM) using Qwen3-8B-AWQ, a 4-bit quantized model. We introduce three structural adap- tations for this setting: (1) system-managed layered memory compression, (2) four activ- ity blocks per simulated day, and (3) explicit physical- and mental-health state variables. Across three independent stochastic runs, two runs completed 52 weeks and the third com- pleted 50 weeks before reaching the context limit, totaling 154 system-weeks (770 agent- weeks). No agent died, and no threshold-based health warning was logged; activity records containing at least oneNO_RESPONSEfield oc- curred at rates of 10.15–10.29% across runs. A 52-week memory-off run tied L2/L3 artifact production to layered memory; a separate 10- week comparison associated four daily time blocks with2.72×more finalized records and lower lexical duplication, but a higher missing- field rate. These comparisons do not support causal behavioral claims. We release validated configurations, derived audits, analysis scripts, aggregate figure data, and our implementation changes in a public fork; raw runs and initial persona data are excluded because their redis- tribution provenance is not fully resolved. 1 Introduction Large language model (LLM)-based agents have demonstrated remarkable capabilities in individ- ual tasks, but their behavior in long-term multi- agent social simulations remains underexplored. Agentopia (Wang et al., 2026) recently introduced a framework where 100 LLM-based agents live, interact, and learn in a shared society over 10 simu- lated years. Its primary experiments use Qwen3.5- 397B-A17B and substantially more compute than a consumer laptop. This leaves a practical question unanswered: what scale and duration of Agentopia- style simulation are attainable on hardware accessi- ble to individual researchers? In this paper, we study a reduced-scale port with five active agents, one apartment world, and a target horizon of one simulated year. We deploy Agen- topia on an NVIDIA RTX 5070 Ti (12 GB VRAM) using Qwen3-8B-AWQ, a 4-bit quantized model served via vLLM (Kwon et al., 2023). To operate within this setting, we add three structural adapta- tions: (1) system-managed layered memory com- pression, (2) four activity blocks per simulated day, and (3) explicit physical- and mental-health state variables. Across three independent stochastic runs, the system completed 52, 52, and 50 weeks, respec- tively. The two full runs reached one year, while the third exposed a context-budget boundary at week 51; activity-record-levelNO_RESPONSErates were 10.15–10.29%. Disabling layered memory removed L2/L3 artifacts, establishing an imple- mentation dependency but not a downstream be- havioral effect. In a 10-week comparison, the archived M+T run contained2.72×as many fi- nalized records per agent-week as the archived M run and showed less lexical reuse but more missing fields; the unseeded, one-run-per-condition design prevents causal claims. No deaths or threshold- based health warnings occurred, but incomplete health ablations prevent attribution to that module. We position this work as an empirical account of a reduced-scale, long-horizon agent society on a consumer GPU, not as a same-scale reproduction of the original Agentopia experiments. 1 arXiv:2608.24215v1 [cs.MA] 25 Aug 2026 2 Related Work LLM social simulation systems. Generative Agents established a widely used architecture that combines memory, reflection, and planning in an interactive sandbox (Park et al., 2023). Subsequent systems have emphasized complementary goals: SOTOPIA formalizes interactive evaluation of so- cial intelligence (Zhou et al., 2024b); Agentopia targets 100-agent, multi-year life simulation (Wang et al., 2026); and AgentSociety and GenSim focus on parallelized or general-purpose simulation in- frastructure (Zhang et al., 2025; Tang et al., 2025). Rather than competing on population scale, our study asks what longitudinal evidence and failure boundaries can be obtained from an Agentopia- style system on a single 12 GB consumer GPU. Validity and long-horizon evaluation. LLMs can reproduce selected behavioral-study outcomes or conditional response distributions (Aher et al., 2023; Argyle et al., 2023), but apparent success can depend strongly on information access and evaluation design (Zhou et al., 2024a). Surveys and methodological critiques therefore call for cali- bration, robustness checks, reproducible protocols, and explicit limits on population-level inference (Gao et al., 2024; Zeng et al., 2026). Recent Social- LLM work similarly audits multi-turn trajectories and separates autonomous action selection from empirical calibration (Star et al., 2026; Lazzaroni et al., 2026). We follow this evidence-bounded view by reporting completed horizons, artifact cov- erage, lexical reuse, and termination conditions without treating five agents as a representative pop- ulation. Memory and resource-constrained deploy- ment. Long-horizon evaluations expose history- dependent challenges across repeated interactions (Goel and Zhu, 2025), while memory-management studies show that how experiences are retained and consolidated can affect later experience-following behavior and computational cost (Xiong et al., 2026; Zhang et al., 2026). Our system-managed L2/L3 pipeline is motivated by this problem, but we evaluate artifact production rather than claiming learning or a causal behavioral benefit. At the de- ployment layer, AWQ provides weight-only quan- tization for memory-constrained inference (Lin et al., 2024), and PagedAttention supports memory- efficient LLM serving (Kwon et al., 2023); these are components of our stack, not contributions re- evaluated by this study. 3 System Figure 1 summarizes the inference stack, weekly simulation loop, and the three structural adapta- tions used in our reduced-scale port. 3.1 Environment and Baseline We adopt the Agentopia framework (Wang et al., 2026) with five active agents cohabiting an apart- ment world for up to 52 simulated weeks. Our im- plementation retains the framework’s plan, contact and scheduling, activity, review, and settlement op- erations; rewards are computed during settlement according to the configured period. Our hardware setup consists of an NVIDIA RTX 5070 Ti (12 GB VRAM) running Ubuntu 22.04 under WSL2. We serve Qwen3-8B-AWQ (Yang et al., 2025; Lin et al., 2024), a 4-bit quan- tized 8-billion-parameter model, via vLLM (Kwon et al., 2023) with a server-side context limit of 24,576 tokens, role model output capped at 3,584 tokens, God model output at 3,840 tokens, temper- ature 0.7, and concurrency limited to one request (max_num_seqs=1) to avoid memory pressure. The archived run configuration records a client-side con- text value of 28,672 tokens; the effective limit was the lower 24,576-token server boundary. We pro- gressively scaled the simulation through a ladder of tests from T1 (1 agent×4 weeks) to T6 (5 agents with a target of 52 weeks). The three T6 runs were independent stochastic restarts, but no global run- level seed was fixed; we therefore do not describe them as controlled-seed replicates. 3.2 Layered Memory Compression The original Agentopia provides recent weekly di- aries together with persistent scratchpad files that agents manage themselves. Detailed diaries outside the recent window leave the prompt unless an agent has promoted salient information into a scratchpad, so long-term retention depends on agent-initiated memory management rather than a systematic com- pression schedule. We add a system-managed three-layer memory architecture with progressively coarser representa- tions. L1 stores complete event logs for the five most recent weeks; agents do not generate these records themselves. L2 holds∼200-word sum- maries for intermediate history, produced by the God model from L1 logs. L3 distills older blocks 2 Reduced-Scale Agentopia on Consumer Hardware Five active agents | one apartment world | target horizon: 52 simulated weeks INFERENCE BACKEND Consumer GPU RTX 5070 Ti Laptop | 12 GB VRAM vLLM Serving 24,576-token server limit one request at a time Qwen3-8B-AWQ 4-bit shared inference model Role-agent and God-model calls AGENTOPIA WEEKLY LOOP 1. Plan Weekly goals and daily schedule 2. Contact & Schedule Social decisions and schedule confirmation 3. Activity four blocks per simulated day NightMorningAfternoonEvening 4. Review Reflection over weekly experience 5. Settlement Rewards, state updates, and weekly checkpoint advance to the next simulated week LLM-mediated Role and God operations STRUCTURAL ADAPTATIONS M | System-Managed Layered Memory L1 Complete event logs recent 1--5 weeks L2 God summaries 6--20 weeks old L3 Milestones 21+ weeks old Eligible L1/L2/L3 context is injected into Plan and Review compression is performed by God-model calls T | Four-Block Day Night | Morning Afternoon | Evening Agent-selected activity or rest reuses the existing daily control flow H | Explicit Health State Activity outcomes and event logs God-evaluated weekly deltas PH / MH 0--100 state Health-aware prompts feed Plan and Review Warning: MH <= 20 or PH <= 10 Physical death check: PH < 10 for 3 or 5 weeks, by social-seeking branch weekly health update and threshold checks simulation flowcontext or control path Figure 1: Architecture of the reduced-scale Agentopia port. A shared Qwen3-8B-AWQ endpoint served by vLLM executes Role-agent and God-model calls on a 12 GB consumer GPU. The weekly loop retains planning, contact and scheduling, activity, review, and settlement, while the M/T/H adaptations add system-managed layered memory, four activity blocks per simulated day, and explicit physical- and mental-health state. Solid arrows denote simulation flow; dashed arrows denote context or control paths. of L2 summaries into∼50-word milestone en- tries. When composing plan and reflection prompts, agents receive the currently eligible context from all three layers, providing a system-managed re- tention path for information that would otherwise depend on scratchpad updates. In short preliminary runs with fixed review prompts, we observed that Reflection paragraphs froze into verbatim repetition after 3–5 weeks. Af- ter the review question set was randomized, the repetitions were no longer observed in subsequent short tests, suggesting that prompt structure may have contributed to the failure mode. The long-run experiments below evaluate whether the layered pipeline continues to produce nonempty, qualita- tively varied memory artifacts; they do not isolate its effect on downstream behavior. 3.3 24-Hour Time Granularity The original framework uses weekly plans with one solo-activity slot per day, and the same planned activity may recur across weekdays. This design limits the representation of within-day rhythms and activity variation. We divide each day into four time blocks—Night (00:00–06:00), Morning (06:00–12:00), Afternoon (12:00–18:00), and Evening (18:00–24:00). Agents plan an activity or rest choice for each block, al- lowing different behavior within the same day. The adaptation reuses the existing weekly and daily control flow while iterating over four solo-activity blocks within each day. When disabled (via the ab- lation flagtime.enable_24h), the system reverts to the original single-solo-activity-per-day behav- ior. 3.4 Health System The original Agentopia models vitality and four fulfillment dimensions, including mood, but it does not maintain explicit physical- and mental-health trajectories or warning and death thresholds. We introduce two separate health indicators: physical_health (0–100, initially 80) and men- tal_health (0–100, initially 70). Physical health is affected by sleep duration, nutrition, and ill- ness events; mental health responds to social interactions—positive encounters (gratitude, recog- nition) increase it, while conflicts, rejection, and loss decrease it. The God model evaluates health deltas each week based on event logs and activ- ity outcomes. Current scores are exposed to sub- sequent plan and review prompts through score- 3 dependent health-awareness text. At each weekly checkpoint, a mental-health score at or below 20 or a physical-health score at or below 10 is logged as a threshold-based health warning. Four consecutive weeks below a mental-health score of 20 addition- ally produce a severe-depression-risk warning but do not force death. For physical-health scores be- low 10, agents with low social-seeking are marked deceased after three consecutive weeks, whereas high social-seeking delays this outcome until five consecutive weeks. The profile-derived social- seeking score is the quantitative extraversion trait when available and otherwise100− introversion; scores of 60 or higher use the high-social-seeking branch. To reduce template-like health evaluations, one prompt cue is sampled each week from a pool of 48 emotional cues. The health system is controlled via the ablation flaghealth.enable; its trajectories are stored separately from the original fulfillment- based subjective reward computation. 4 Results We evaluate the full system (all three adaptations enabled, denoted M+T+H) across three indepen- dent stochastic restarts lasting 52, 52, and 50 weeks, respectively. All five selected agents are active par- ticipants, yielding 154 completed system-weeks and 770 agent-weeks. 4.1 System Stability The first two runs reached the 52-week target. The third completed 50 weeks but stopped when the week-51 prompt reached the configured con- text boundary. Before termination, Run 3 wrote 16 finalized activity records for the incomplete week 51. We include these records in activity- record-level metrics because they are present inactivity.jsonl, but exclude week 51 from completed-week totals and weekly-outcome analy- ses. No infrastructure-level termination occurred within a completed week; missing activity fields are quantified separately below. The weekly pipeline executed fully in every completed week. These results support feasibility at the tested five-agent scale and identify context growth as a long-horizon limitation observed in Run 3. 4.2 Observed Health Trajectories Figure 2 shows the weekly physical- and mental- health trajectories, and Table 1 reports the scores at each run’s final checkpoint. Across all three runs, no agent died and no threshold-based health warning was logged. Physical health remained at or above 93 for all agents at the final checkpoint; the lowest mental-health value recorded across all weeks was 53 (see Table 4 for per-run minima), above the mental-health warning threshold of 20. Because the health-disabled runs are incomplete, these observations characterize the enabled system but do not establish that the health module caused the absence of warnings. Table 1: Final-checkpoint physical-health (PH) and mental-health (MH) scores. Both range from 0 to 100. Runs 1–2 end at week 52 and Run 3 at week 50; each cell reports PH/MH. Agent Run 1 PH/MH Run 2 PH/MH Run 3 PH/MH Whitfield100/9395/7096/71 Morales100/97100/79100/89 Vale93/9296/77100/96 Voss100/79100/77100/99 Vieri100/95100/100100/91 Table 2: Activity-record-levelNO_RESPONSEincidence. Total records reports the number of finalized activity records, and Flagged records reports the number con- taining at least oneNO_RESPONSEfield; each record is counted at most once. RunTotal records Flagged recordsRate Run 14,412448 10.15% Run 24,436451 10.17% Run 34,236436 10.29% Pooled13,0841,335 10.20% We measure missing activity fields from the final- izedactivity.jsonlrecords rather than counting string occurrences in generation transcripts, where the same historical marker can be copied into many later prompts. The pooled activity-record rate in Table 2 is 10.20%, and the sample coefficient of variation across the three unrounded run-level rates is 0.75%. This is an activity-record metric, not a call-level failure rate, because one activity can contain multiple LLM calls. To audit output reuse separately from literal NO_RESPONSEmarkers, we examined solo records without that activity-level flag. Within each run and agent, normalized primary-text exact-duplicate rates were 14.48%, 17.18%, and 17.52% for Runs 1–3; comparison against the preceding eight weeks using 5-token shingle Jaccard≥ 0.80 4 11020304052 Simulation week 0 20 40 60 80 100 Health score (0--100) warning threshold = 10 (a) Physical health 11020304052 Simulation week warning threshold = 20 (b) Mental health Run 1 (52 weeks)Run 2 (52 weeks)Run 3 (50 weeks) Figure 2: Weekly physical- and mental-health trajectories across the three full-system runs. Lines show the weekly mean over the five active agents, and shaded bands span the agent-level minimum and maximum. Dotted horizontal lines mark the physical-health warning threshold of 10 and mental-health warning threshold of 20. Run 3 ends after week 50. yielded additional near-duplicate rates of 5.72%, 4.66%, and 4.40%. Normalization applied Unicode NFKC normalization, lowercasing, and punctua- tion removal, and comparisons were restricted to the same activity type. These lexical screens ex- pose output reuse but do not measure semantic or behavioral diversity. 4.3 Memory Artifacts and a Short Time Comparison Across the stochastic runs, Aaron Whitfield’s week- 40 L2 entries emphasized, respectively, deadline- driven family guilt, avoidance through writing, and peer conversations prompting a promised visit. In qualitative inspection, these entries appeared non- template-like and contextually grounded. Each ac- tive agent has 47, 47, and 45 L2 entries in Runs 1–3, respectively (Run 3 produces two fewer because it is two weeks shorter), for 695 L2 and 90 L3 artifacts in total. Given the configured five-week recent window, all 15 agent-run archives contained an L2 entry for every expected older week; this establishes continuous artifact production, not suc- cessful recall or downstream use. These examples illustrate that the stored summaries preserve run- specific histories; they do not establish an effect on downstream behavioral diversity. We also audited two archived 10-week, five- agent conditions: layered memory alone (M) and memory plus four daily time blocks (M+T), with health disabled. Their saved configurations differ only inworld.time.enable_24hand output di- rectory; Table 3 reports activity volume, missing fields, and lexical reuse. Table 3: Exploratory 10-week time toggle (one run per condition; health disabled). Rec./A-W is final- ized records per agent-week; NR is activity-record-level NO_RESPONSE. Exact/near are lexical duplicate rates in unflagged solo primary texts. Condition Weeks Rec./A-WNRExactNear M (T off)106.36 1.57% 57.60% 16.40% M+T1017.28 7.75% 10.06%6.45% M+T contains2.72×as many records per agent- week as M, reflecting its larger activity budget rather than higher quality; all five agents have lower exact-duplicate and higher NR rates. With one un- seeded, unpaired run per condition and different activity opportunity counts, these results are de- scriptive and do not establish improved semantic or behavioral diversity. 4.4 Cross-Run Summary Table 4 aligns the main run-level metrics with each run’s end condition. Across runs, minimum PH was 80 in all cases, minimum MH ranged from 53 to 60, and activity-record-levelNO_RESPONSE rates ranged from 10.15% to 10.29%; Run 3 ended two weeks earlier at the context boundary. Table 5 summarizes coverage and completion status for the configurations discussed below. 5 Table 4: Run-level summary. NR is the activity-record-levelNO_RESPONSErate, D/W denotes deaths/threshold-based health warnings, and health minima cover all completed weeks. L2 entries/agent is the final-checkpoint count per active agent. Aggregate weeks are system-weeks, and aggregate NR is pooled over records. RunWeeksEnd conditionActivity recordsNR rateMin. PHMin. MHD/WL2 entries/agent Run 152Target reached4,41210.15%80590/047 Run 252Target reached4,43610.17%80530/047 Run 350Context boundary4,23610.29%80600/045 Aggregate154–13,08410.20%80530/0– Table 5: Completion status and observed outcomes for the full-system runs, ablation conditions, and 4-week all-off baseline. M, T, and H denote layered memory, four-block daily time, and health. ConfigurationRunsWeeks completed Status or terminationObserved outcome Full (M+T+H)352, 52, 50Planned horizon completed (2); context boundary (1) Zero deaths/warnings; L2/L3 artifacts present −Memory (T+H)152Planned 52-week horizon completed L2/L3 artifacts absent by configuration −Health (M+T)221, 38 Incomplete:week 21, resume-pathslowdown; week 38, WSL I/O failure No complete 52-week health comparison available Time toggle (M vs. M+T) 1 each10, 1010-week comparison com- plete; 52-week M+H not run 2.72×records/A-W; lower reuse, higher NR; descriptive only All off14Planned 4-week baseline completed Longitudinal comparison unsupported; health warning/death mechanisms disabled 5 Discussion 5.1 Ablation Analysis We completed one 52-week memory ablation, one 10-week time-toggle comparison, and two incom- plete attempts to disable the health module; we did not complete a 52-week M+H leave-one-out com- parison for the time adaptation. The memory-off run (T+H enabled) completed 52 weeks with zero deaths and zero threshold-based health warnings, showing that this configuration remained opera- tional without layered memory. Physical health ranged from 98 to 100 across agents at the final checkpoint, and mental health from 74 to 98. As expected from the configuration, its L2 and L3 directories were empty and no compressed sum- maries were available to agent prompts. This con- firms the implementation dependency between the module and the retained long-term memory arti- facts. It does not, by itself, establish that layered memory causes greater downstream behavioral or narrative diversity; that claim requires a matched comparison of agent outputs under enabled and disabled conditions. The short M versus M+T comparison exposes a temporal-granularity trade-off—more activity and lower literal reuse, but more missing fields—rather than improved behavioral diversity, because op- portunity counts differ and each condition has one unseeded run. The two health-off runs (M+T enabled) did not complete the full 52-week horizon: the first termi- nated at week 21 due to a resume-path performance degradation, and the second at week 38 due to a WSL I/O failure. Because disabling the module also removes its health-specific warning and death mechanisms, and both runs are incomplete, we do not compare their health outcomes with the full system. A complete matched ablation remains for future work. The all-off configuration of our reduced-scale port completed the planned 4-week baseline. Its short horizon is useful as an implementation refer- ence but is insufficient for a longitudinal reliability comparison, so we do not use it to claim that the three adaptations collectively reduce failures. 6 Conclusion On a single RTX 5070 Ti (12 GB), our five-agent system completed 52, 52, and 50 weeks, total- ing 154 system-weeks (770 agent-weeks), with no deaths or threshold-based health warnings. The memory-off run tied L2/L3 artifacts to layered memory, while the short time toggle exposed a trade-off among activity volume, lexical reuse, and missing fields. These observations do not establish 6 causal behavioral effects, and the health module’s contribution remains unresolved. Limitations Several limitations warrant discussion. First, our experiments use a single model (Qwen3-8B-AWQ) for both agent roles and God evaluations; a multi- model setup might yield richer agent differentiation. Second, the concurrency limit serializes inference requests; at the tested five-agent scale, a 52-week run took roughly 33–40 hours. We did not mea- sure scaling beyond this setting. Third, we evaluate only the apartment world setting; generalization to other Agentopia worlds (school, workplace) re- mains untested. The physical- and mental-health variables are simulator-internal state variables; they are non-clinical and have not been validated against medical or psychometric constructs. Fourth, lim- ited compute prevented a matched 52-week base- line and exhaustive ablations; additionally, we did not fix a global run-level random seed. Fifth, our time-toggle comparison covers only 10 weeks, with one unseeded run per condition and health disabled; it is not a 52-week M+H leave-one-out ablation, so the time module’s causal contribution to behav- ioral diversity remains unquantified. Sixth, our narrative evidence consists of stored-memory ex- amples and counts rather than an automated or human-rated measure of downstream behavioral diversity. Seventh, the unflagged solo-output au- dit identifies lexical reuse but does not determine whether repeated wording corresponds to repeated actions, intentions, or social outcomes. Finally, the NO_RESPONSEstatistic is defined over finalized ac- tivity records, not individual model calls, and there- fore characterizes missing activity fields rather than serving as a direct inference-server failure rate. The public research artifact athttps://github.com/ luo675/Agentopia/tree/paperis maintained as a fork of the upstream Agentopia repository and contains our implementation changes, validated configurations, derived audits, analysis scripts, and aggregate figure data. It excludes raw run archives, initial persona data, model weights, and creden- tials; the upstream project and its attribution remain linked through the fork relationship. References Gati V. Aher, Rosa I. Arriaga, and Adam Tauman Kalai. 2023. Using large language models to simulate mul- tiple humans and replicate human subject studies. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 337–371. PMLR. Lisa P. Argyle, Ethan C. Busby, Nancy Fulda, Joshua Gubler, Christopher Rytting, and David Wingate. 2023. Out of one, many: Using language mod- els to simulate human samples. Political Analysis, 31(3):337–351. Chen Gao, Xiaochong Lan, Nian Li, Yuan Yuan, Jing- tao Ding, Zhilun Zhou, Fengli Xu, and Yong Li. 2024. Large language models empowered agent- based modeling and simulation: a survey and per- spectives. Humanities and Social Sciences Commu- nications, 11:1259. Hitesh Goel and Hao Zhu. 2025. LIFELONG SO- TOPIA: Evaluating social intelligence of language agents over lifelong social interactions. arXiv preprint arXiv:2506.12666. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gon- zalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serv- ing with PagedAttention. In Proceedings of the 29th Symposium on Operating Systems Principles (SOSP), pages 611–626. Ruggero Marino Lazzaroni, Lorenz Prattes, and Jana Lasser. 2026. Calibrated but autonomous: Inference- time bayesian logit correction for LLM social simu- lations. SocialLLM Workshop at ICWSM 2026. Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei- Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. 2024. AWQ: Activation-aware weight quantization for on- device LLM compression and acceleration. In Pro- ceedings of Machine Learning and Systems, vol- ume 6. Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. Generative agents: Interactive simu- lacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, pages 1–22. Michelle Star, Andrew Aquilina, and Yu-Ru Lin. 2026. Auditing support strategies in LLMs through grounded multi-turn social simulation. SocialLLM Workshop at ICWSM 2026. Jiakai Tang, Heyang Gao, Xuchen Pan, Lei Wang, Hao- ran Tan, Dawei Gao, Yushuo Chen, Xu Chen, Yankai Lin, Yaliang Li, Bolin Ding, Jingren Zhou, Jun Wang, and Ji-Rong Wen. 2025. GenSim: A general social simulation platform with large language model based agents. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (System Demonstrations), pages 143– 150. 7 Xintao Wang, Sirui Zheng, Hongqiu Wu, Weiyuan Li, Jen-tse Huang, Minghao Zhu, Can Zu, Qi Deng, Ji- awei Wang, Qianyu He, Heng Wang, Xiaojian Wu, and Yunzhe Tao. 2026. Agentopia: Long-term life simulation and learning in agent societies. arXiv preprint arXiv:2606.07513. Zidi Xiong, Yuping Lin, Wenya Xie, Pengfei He, Zirui Liu, Jiliang Tang, Himabindu Lakkaraju, and Zhen Xiang. 2026. How memory management impacts LLM agents: An empirical study of experience- following behavior. In Proceedings of the 64th An- nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 623–645. An Yang et al. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388. Yongchao Zeng, Calum Brown, and Mark Rounsevell. 2026. Too human to model: the uncanny valley of large language models in simulating human systems. npj Complexity, 3:13. Jiaquan Zhang, Chaoning Zhang, Shuxu Chen, Zhen- zhen Huang, Pengcheng Zheng, Zhicheng Wang, Ping Guo, Fan Mo, Sung-Ho Bae, Jie Zou, Jiwei Wei, and Yang Yang. 2026. Lightweight LLM agent memory with small language models. In Proceed- ings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12914–12929. Jun Zhang, Yuwei Yan, Junbo Yan, Zhiheng Zheng, Jinghua Piao, Depeng Jin, and Yong Li. 2025. A par- allelized framework for simulating large-scale LLM agents with realistic environments and interactions. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track), pages 1339–1349. Xuhui Zhou, Zhe Su, Tiwalayo Eisape, Hyunwoo Kim, and Maarten Sap. 2024a. Is this the real life? is this just fantasy? the misleading success of simulating social interactions with LLMs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 21692–21714. Xuhui Zhou, Hao Zhu, Leena Mathur, Ruohong Zhang, Haofei Yu, Zhengyang Qi, Louis-Philippe Morency, Yonatan Bisk, Daniel Fried, Graham Neubig, and Maarten Sap. 2024b. SOTOPIA: Interactive evalu- ation for social intelligence in language agents. In International Conference on Learning Representa- tions. 8