Microsoft Research published a report examining media authenticity and verification methods as synthetic media becomes more prevalent. The research explores capabilities and limitations of current authentication techniques for images, audio, and video content, while identifying practical approaches for establishing trustworthy content provenance.
Microsoft Research introduced Agent Lightning, a system that enables developers to add reinforcement learning capabilities to AI agents without requiring code rewrites. The system decouples agent functionality from training processes, converting each agent action into reinforcement learning data to improve performance with minimal code changes.
Microsoft Research explores Predictive Inverse Dynamics Models (PIDMs) in imitation learning, showing they outperform standard Behavior Cloning by using predictions to reduce ambiguity. The approach enables more efficient learning from fewer demonstrations compared to traditional methods.
