Context Is Not Comprehension

Alex Pan, Mary-Anne Williams

The dominant way of judging Large Language Models (LLMs) has been to ask how well they can recall explicit facts from very long inputs. While today's best models achieve near perfe…