New Paper: World Properties without World Models

New Paper: World Properties without World Models


Dr. Elan Barenholtz has published new research challenging claims that language models build internal world models. The paper, “World Properties without World Models,” demonstrates that spatial and temporal structure (city coordinates, historical birth years) can be recovered from simple static word embeddings (GloVe, Word2Vec) with R² values of 0.71-0.87 — suggesting this information is already latent in text co-occurrence statistics, not evidence of world-like representations in LLMs.

Read the paper on arXiv