Shifting AI Alignment from Static Preferences to Dynamic Workflows

Current alignment methods fail by focusing on static human preferences rather than the co-evolution of human and model behavior. The authors propose a trajectory-level alignment framework that treats human-AI interaction as a dynamic, context-dependent collaboration.