Yazarlar
Yayınlar
- BildiriYalnız künye2025Multi Objective Reinforcement Tuning for Responsible Generative AI in Automated Financial Advisory Systems
Bhandarkar, Vishakha D; Genemo, Musa Dima; Muda, Iskandar; Murugan, Rekha; Sreeram, G.; Dharani, N.
Generative AI in a financial context poses risks to the areas of regulatory compliance, accuracy in the content being delivered, and appropriateness in terms of emotion. This research proposes ReFiGen, a framework that can be used to create compliant financial content that is compliant with the regulatory framework and the sentiment of the users. Following domain-specific fine-tuning, ReFiGen blends sentiment alignment with reinforcement learning to optimize on compliance. It demonstrates better performance than baseline models with a high data accuracy rate of 98.35 percent and can generate a fluent output that is emotionally congruent, and regulation-minded. The results of experiments indicate that ReFiGen is both clear and coherent and exactly follows financial guidelines. This practice affirms that the combination of unstructured compliance modeling and affective control is a successful aspect of LLMs. ReFiGen is one of the milestones on the journey of trustworthy and ethically aligned AI in the financial advisory market.