Why are negative instructions ("don't do X") less effective than positive ones?

Answer preview

Problem: "Don't mention competitors" can still trigger competitor discussion — models attend to all tokens.…