# writing

Researchers at NYU and UT Austin found that frontier models lean harder than a Japanese round-trip paraphrase on several sentence patterns already linked to human reading difficulty. Their hypothesis: post-training that rewards only correct answers and penalizes length teaches models to compress. A small-model experiment supports the idea, but it is a proof of concept, and only English was tested.