but the really interesting finding isn't about efficiency. it's about harm the paper identifies something they call "context pollution" when models condition on their own prior responses, they sometimes lock onto errors, hallucinations, or stylistic artifacts from earlier