With the rapid development of artificial intelligence technology, AI-assisted curation has gradually emerged as an innovative interaction method in contemporary art exhibitions. To enhance users’ interactive experience in virtual art exhibitions, this study constructs an interaction design system optimized based on Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM), aiming to improve the accuracy of user behavior prediction and the real-time response of the system in art exhibitions. Experimental results demonstrate that GRU-LSTM performs optimally in terms of accuracy, convergence speed, real-time performance, and robustness. In terms of accuracy, GRU-LSTM achieves 92.4% after 800 iterations, significantly higher than other algorithms. In real-time performance tests, GRU-LSTM exhibits excellent real-time performance with a response time of 450 ms when 1000 users interact simultaneously. In terms of robustness, GRU-LSTM experiences the smallest decline in accuracy under noisy data, ensuring system stability in complex environments. Overall, the GRU-LSTM algorithm significantly enhances the performance of the interaction design system for virtual art exhibitions by optimizing the processing of long-term dependent information. It demonstrates high accuracy and real-time performance, especially under multi-user interaction and complex data conditions, showcasing strong application potential.