Capital One is actively involved in the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP) in Suzhou, China, scheduled from November 4 to 9. The company is set to present research and innovations aimed at enhancing AI safety and reliability in natural language processing. Dr. Sambit Sahu, VP of AI Foundations, will deliver a keynote on a new framework called Multi-Agent Conversational Assistant Workflow (MACAW). This system is designed to create conversational assistants that can handle complex inquiries and perform tasks for users, emphasizing self-reflection and adaptive learning. Among the eight papers being presented, one significant contribution is GRAID (Geometric and Reflective AI-Driven Data Augmentation). It addresses data scarcity in harmful text classification by generating synthetic data with two main approaches: Geometrically Controlled Generation and Multi-Agentic Reflection. This methodology led to a 12% improvement in model performance on benchmark datasets. Another notable study compares different fine-tuning strategies for Retrieval-Augmented Generation (RAG), revealing that while performance improvements are consistent across strategies, computational costs vary significantly. The research also includes TruthTorchLM, an open-source library focused on predicting the truthfulness of responses generated by LLMs. With over 30 prediction methods, this tool aims to improve reliability in high-stakes applications. This project is part of a collaboration with the USC‒Capital One Center for Responsible AI, highlighting Capital One's commitment to advancing responsible AI practices in finance. These efforts reflect the company’s focus on building robust AI systems that prioritize trust and efficiency.