Just read 'Beyond Distribution Sharpening' paper - are task rewards the missing piece for AI alignment?

The paper 'Beyond Distribution Sharpening: The Importance of Task Rewards' argues that current alignment methods focusing on distribution sharpening may be insufficient. It suggests incorporating explicit task rewards could lead to more robust and goal-aligned AI systems. This has implications for how we train and evaluate AI models in production environments.