Dual IAS Seminar: Associate Professor Sumiko Miyata and Professor Takamichi Miyata Externally Funded Fellows Associate Professor Sumiko Miyata and Professor Takamichi Miyata each deliver a seminar on their research. Associate Professor Sumiko Miyata: Incentive-Driven AI Networks for Future Road Safety To achieve fully autonomous driving, 'cooperative perception' via V2X (Vehicle-to-Everything) is essential for eliminating blind spots and improving recognition accuracy. However, a major barrier to sustainable implementation lies in ensuring 'fair incentives' for participants to share data and computational resources. This seminar introduces an AI-driven network framework designed to balance infrastructure efficiency with participant satisfaction. The presentation first covers a reward distribution mechanism based on the game theory concept of 'Nucleolus' to minimise user dissatisfaction within the monitoring system and ensure long-term cooperation. Building on this foundation, the discussion addresses essential network mechanisms for 'City as a Service', such as high-speed AI processing that optimises task offloading between edge servers to minimise communication latency. By integrating incentive design with advanced communication control, it is possible to build a reliable social infrastructure that reduces accidents and optimises urban mobility. Professor Takamichi Miyata: Multimodal AI that Understands Driver Behaviour without Training Data Distracted driving remains a critical safety concern, as even brief lapses in attention can lead to serious traffic collisions. Current supervised learning methods require large, labelled datasets and struggle to generalise, while vision-language model (VLM) based methods enable training-free recognition but tend to capture driver identity rather than actual behaviour. This seminar presents a novel framework that overcomes both limitations. The key innovation lies in decoupling identity-related information from behaviour-related cues, combined with refined textual representations to enhance zero-shot recognition robustness across diverse drivers and environments. By integrating decoupled multimodal representations with a lightweight model architecture, the proposed system achieves practical, scalable performance without relying on extensive labelled data. This