Word Vectors Without Transformers Question LLM World Models

Commenting on the preprint "World Properties without World Models" V2, researcher Gerard Sans argues that word vectors alone can decode spatiotemporal and emotional information, and that replacing corpus words with ciphers dissolves meaning—suggesting LLM internal states are mere co-occurrence distributions rather than world models.

2026-10-08 ~ 2026-10-08 · 2 related posts

• Word Vectors Without Transformers Recover Space, Time, Pain and Emotion, Challenging LLM World Model Evidence — gerardsans · 2026-10-08
• Applying a Cipher to the Corpus Suffices to Dismiss Theories of Meaningful Internal States in LLMs — gerardsans · 2026-10-08