Capital One will present eight research papers at ICML 2026 in Seoul focused on large language model safety and reasoning. Key contributions include Critique-Guided Distillation for improved mathematical reasoning, TRACER for detecting multi-turn agent failures early, and EPSVec for privacy-preserving synthetic data generation. The company’s work addresses challenges in AI alignment, evaluation metrics, and agentic fault diagnosis — areas that are increasingly critical as enterprises deploy AI in production environments.
