Search engines on social apps used to rely on many separate mini-models that often misunderstood slang and were hard to keep updated.
This paper introduces YaPO, a way to gently nudge a language model’s hidden thoughts so it behaves better without retraining it.