What's this about?
NSD has a model generate a deliberately flawed reasoning condition and trains the student away from it, rather than imitating a teacher trace that already knows the answer.
What exactly is self-distillation?
It is training a model on guidance produced by itself or another copy of itself. Usually the copy acts as a teacher; here it supplies a negative example to avoid.
Why is it worth attention?
The method preserves uncertainty and self-correction in principle, while a gate limits damage to ordinary language tokens. Results are preprint claims and need reproduction across models and domains.