Reproduce “Evaluating the World Model Implicit in a Generative Model” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Keyon Vafa, Justin Y. Chen, Ashesh Rambachan, Jon Kleinberg, Sendhil Mullainathan

The authors test whether sequence models learn coherent world models by comparing histories that reach the same state and histories that reach different states. In Manhattan navigation, a 1.5-billion-parameter transformer had 1.00 rounded next-token validity and 0.99 current-intersection probe scores, but compression precision was 0.50. These findings are limited to tests using a known, queryable deterministic finite automaton.

Key insights

  • Generative models can achieve high performance on traditional metrics like next-token prediction while possessing incoherent and fragile implicit world models.
  • The Myhill-Nerode theorem provides a robust theoretical basis for evaluating implicit world models by assessing sequence compression (recognizing identical states) and sequence distinction (differentiating distinct states).
  • Training data diversity, such as random walks or synthetic game data, can lead to more coherent implicit world models compared to highly optimized or real-world sequence data.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.