It gets better: They tested preventative steering during fine-tuning. Train the model while nudging it away from the trait direction. Result: → Trait expression stays suppressed
→ Performance (e.g. MMLU) stays intact Safe training with no compromise.

→ View original post on X — @godofprompt