Large language models (LLMs, the systems behind chat assistants) pick up broad general skills in pre-training, then get narrowed to one job by supervised fine-tuning, which trains them on example answers. That narrowing often damages the wider skills.
The model keeps many concepts as directions in its internal signal, packed so that they stay nearly independent. Narrow training data forces related directions to merge, and on unrelated questions the merged directions interfere, pushing the signal across the boundary that would have chosen the right next word.
Training with a penalty that blocks this drift, alongside the normal objective, does better on both the new task and the old skills, for about 1.3 times the computing cost of ordinary fine-tuning.