A problem with static snapshots
AI agents given a running record of life events generally produced survey answers closer to human response patterns than static-profile agents and other comparison methods, according to an arXiv preprint. The approach, called LifeMem, combines a structured record of events that can be retrieved with an agent-specific LoRA adapter that carries accumulated experience as persistent parametric memory.
The paper uses “identity essentialism” for a model pattern in which static-profile agents are more tightly separated by socioeconomic group than human respondents—in plain terms, the group label appears to matter too much. In a separate diagnostic of 2,000 randomly sampled WVS wave 7 respondents, a clustering score called the silhouette score was −0.02 for humans and 0.19 for profile-conditioned Llama-8B agents. The authors describe that comparison as a motivating diagnostic, not a stand-alone measure of simulation quality or evidence about a human psychological mechanism.
Testing against longitudinal surveys
For the main experiments, the researchers sampled 100 respondents from each of two longitudinal surveys, Add Health and Understanding Society. The complete-case aligned pools contained 2,048 Add Health respondents across six waves and 14,104 Understanding Society respondents across 15 waves. Three similarly sized LLM backbones generated responses using shared respondent samples, questions and settings.
Outputs were judged with four measures, all better when lower: KL divergence, which captures the difference between response distributions; a within-group pairwise-distance gap; a normalized entropy gap, used here to compare response diversity; and a JS divergence for transition distributions between adjacent waves. The paper reports largely significant gains across settings, but performance was not uniformly best on every individual metric.
The main comparison
Across the model-and-survey combinations, LifeMem generally produced the closest match on response distributions, within-group differences, diversity and, where measured, longitudinal transitions. When respondent sampling was repeated with seeds 42, 43 and 44, LifeMem had the best mean result in 16 of 18 model-dataset-metric settings, with generally small standard deviations.
One exception appeared with Qwen3.5-9B on Add Health. The Multilingual baseline had slightly lower within-group and entropy gaps than LifeMem, but its KL divergence was substantially higher—an example of the trade-off between matching diversity and matching the overall response distribution.
Ablation tests pointed to the same two-part design: removing either the structured event memory or the persistent parametric memory worsened performance on most metrics. The pattern supports the authors’ view that the components work in complementary ways, although their relative contribution varied across models and datasets.
More history, then diminishing returns
In these tests, adding more history helped up to a point. In an Event RAG comparison on Add Health with Llama-8B, retrieval was varied from one to 180 events; alignment gains largely plateaued around 40 events while runtime and input-token use continued to rise. In LifeMem, coverage from five to 180 life events per wave generally improved alignment and diversity until gains plateaued at roughly 90 events per wave.
A separate visualization followed 100 Llama-8B agents over 15 Understanding Society waves. Their agent-specific LoRA states started concentrated, then became more dispersed and followed distinct trajectories as histories accumulated. The pattern is consistent with increasing personalization, but it is a descriptive representation of internal states rather than a direct behavioral validation.
LifeMem was slower than lightweight prompt-only methods but faster than Full History, Event RAG and SimVBG. With Llama-8B, its average per-question latency was reported as 161.7 milliseconds plus 0.9 milliseconds for adapter loading on Add Health, and 154.5 milliseconds plus 1.4 milliseconds on Understanding Society. The paper notes that the timings include retrieval and that absolute runtime depends on hardware configuration.
What the result does not show
The valid-response figures measure whether an answer could be mapped to a survey option, not whether it was behaviorally realistic. Human valid-response rates were 87.37% in Add Health and 64.31% in Understanding Society. LifeMem’s rates ranged from 96.41% to 97.35% on Add Health and from 91.59% to 93.34% on Understanding Society across the three backbones.
The study’s evidence remains limited to the two surveys, three model backbones and sampled respondents used in these experiments. Although the aligned pools were larger, the main runs used 100 respondents per dataset; complete-case selection and the use of surrogate alignment and diversity measures leave open how the method would perform on other populations, tasks or models.
The study used public-use, de-identified survey data and reported aggregate results. It withheld respondent-level adapters because they may memorize respondent information.
LifeMem’s improvement is therefore a result about matching observed survey patterns under the tested settings. It does not show that the agents are digital replicas of respondents, that identity essentialism is a human psychological mechanism, or that the systems are suitable for consequential decisions about individuals. The document is an arXiv preprint, version 1, dated 20 August 2026; whether the gains generalize beyond the studied surveys, models and tasks remains open.
Paper data and sources
Original title: Mitigating Identity Essentialism in LLM Agents with Longitudinal Life Trajectories
Authors: Hexi Wang, Yujia Zhou, Bangde Du et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text