Tag
This paper empirically studies how lexical perturbations disrupt large language model reasoning through attention diversion, finding that character-level noise significantly degrades performance while filler insertions have little effect.