A general approach for alignment generalization and a Stupid New Technique that beat everything

Researchers propose a general method for improving AI alignment generalization and introduce a new technique that outperforms existing methods in aligning AI behavior.