Do LLMs Have World Models? Melanie Mitchell Investigates

AI: A Guide for Thinking Humans is Melanie Mitchell's ongoing newsletter on where AI research actually stands, and this entry opens a two-part look at a genuinely contested question: do large language models build anything like an internal "world model," or do they succeed by leaning on surface statistics that happen to correlate with the right answer?

Mitchell starts with pre-LLM cautionary tales. An image classifier learned to flag skin lesions as malignant largely because the malignant training photos happened to include a ruler for scale, and a separate language model judged logical implication by word overlap rather than by meaning. Both looked competent right up until the pattern they had actually learned stopped lining up with the task. She sets these against the live debate over LLMs: researchers like Ilya Sutskever argue the models compress a usable representation of the world, while skeptics such as Yann LeCun and Subbarao Kambhampati see approximate retrieval over memorized text instead.

The piece closes with a working taxonomy, borrowed from MIT's Jacob Andreas, for grading how strong a "world model" really is: from a static lookup table, up through maps and orreries, toward a full causal simulator. Mitchell places LLMs somewhere in that middle range rather than at either extreme, without settling the question outright.

This entry lays out the debate and the vocabulary for it; Mitchell says the supporting evidence is the subject of Part 2, so readers looking for a verdict here will find groundwork instead.