Language models fail at extended rule following

Tianxiang Dai, Jonathan Fan

Large language models are highly capable of answering difficult questions by retrieving, recombining, and attending to information in long contexts. For agentic tasks, an additiona…