May 10, 2026

TL;DR. When you click with someone, it almost always means you're both a little broken in the same way.

TL;DR. Post-training is distribution reshaping — SFT yanks toward an external target and forgets, RL nudges along reward and preserves, and on-policy distillation gets the best of both because the sampling matters more than the teacher.

Takeaways

• SFT's catastrophic forgetting isn't a bug, it's geometry: cross-entropy pulls hard toward the dataset with zero regard for where the model started.
• On-policy distillation students outperformed their teachers and forgot less — even when distilled from a degraded SFT model, suggesting on-policy sampling does the heavy lifting.
• RL resists forgetting because it's data-dependent: small updates when uncertain, bigger when confident — versus SFT's uniform aggression on every token.