Where Do Adolescent Digital Twins Succeed and Fail? A Multi-layer Validation of Survey-Anchored Generative Agents with an LLM Backbone

Abstract Ethical constraints often prevent experimental manipulation of harmful social exposures in adolescence, limiting how interventions and social contexts can be systematically studied. In response, large language models (LLMs) have renewed interest in developing digital twins, generative agents designed to stand in for adolescents in in silico experiments, yet the validity and practical utility of such twins remain unclear. We constructed survey-anchored adolescent digital twins with 2023 Youth Risk Behavior Survey and enriched them with narrative memory from youth videos via retrieval-augmented generation. We benchmarked these twins across four layers: survey fidelity with domain holdouts, preservation of developmental correlation structure, live chats with real adolescents, and replication of randomized trials. Our results showed that survey conditioning recovered expected association patterns and replicated more trial contrasts than a generic role-play LLM baseline. Narrative enrichment improved held-out agreement and replication, but sometimes strengthened victimization associations and reduced social presence when misaligned with survey profiles. These results support adolescent digital twins as a constrained tool for early-stage hypothesis refinement and intervention screening, with caution for subgroup inference and higher stakes uses.