Contextual Safety

Shortcuts & Boundaries: RL Misalignment and Contextual Safety

How agentic RL can improve real work while learning contextual violations, and how verifiable feedback supports useful, transferable safety.