HyCoRA: Hyper-Contrastive Role-Adaptive Learning for Role-Playing

HyCoRA represents a significant advancement in the field of AI role-playing, addressing the shortcomings of previous methods that either utilized a single shared module or separate modules for each role. The proposed framework employs a Hyper-Half Low-Rank Adaptation structure, effectively balancing the learning of unique and common traits among characters. This innovative approach not only enhances the simulation of diverse personalities but also demonstrates its effectiveness through extensive experimental results on benchmarks in both English and Chinese. Furthermore, evaluations involving GPT-4 provide additional validation of HyCoRA's capabilities, showcasing its potential to transform multi-character role-playing in AI applications.