Longitudinal Recall and Engagement via Real-Time 2D AI Avatars in Virtual Heritage Environments

Abstract

Immersive virtual reality (VR) offers transformative potential for cultural heritage education, yet populating historical environments with authentic, believable pedagogical agents remains technically and financially prohibitive. Traditional 3D character pipelines require intensive modeling and frequently trigger the "uncanny valley," whereas static virtual museum exhibits risk cognitive overload via split-attention effects. This study evaluates an alternative paradigm: streaming real-time, 2D AI-generated neural talking-head avatars animated directly from archival portraits into an immersive Unreal Engine 5.7 environment. A randomized controlled trial was conducted with university students (N = 64) assigned either to an experimental condition featuring an audio-driven 2.5D billboarded historical avatar or a control condition utilizing a traditional non-avatar VR guide with static placards and voice-over narration. Longitudinal knowledge retention was measured immediately post-intervention (T₀), after one week (T₁), and after three months (T₂), alongside psychometric assessments of cognitive load (NASA-TLX), social presence, and historical empathy. A 2 x 3 mixed-design ANOVA revealed a significant interaction effect (F(1.78,110.36) = 32.41, p < .001, ɳ²_p = .343). While initial recall was comparable (84.62% vs. 81.35%), the 2D AI avatar cohort demonstrated significantly buffered memory retention after three months (65,45% VS. 51.22%, p < .001). Furthermore, MANOVA results confirmed substantial reductions in extraneous cognitive load (p < .001) alongside higher social presence and affective engagement. Grounded in Social Agency Theory and Cognitive Load Theory, these findings establish real-time 2D generative media as a scalable, pedagogically robust framework for interactive historical storytelling in virtual heritage.