generalcritiquebearishRewriting inputs to improve frozen downstream models has become a common strategy in modern NLP pipelines but faces reliability issues under realistic deployment conditionsComputation and Language27 Jul 2026http://arxiv.org/abs/2607.01964v1