A theoretical account of how large language models gain world knowledge without sensorimotor experience. The Collective World Model Hypothesis holds that language externalizes a collective world model emerging from the decentralized sense-making of a society of embodied agents, and that an LLM learns a statistical approximation of it. The Generative Emergent Communication framework, built on Collective Predictive Coding, formalizes language emergence as decentralized Bayesian inference minimizing shared prediction error, creating a society-scale encoder-decoder in which humans encode grounded Type 1 world models into symbols and an LLM decodes them to reconstruct a mirroring latent space. This distinguishes subjective (Type 1) from objective (Type 2) world models and explains distributional semantics as representation reconstruction.
A theoretical AI paper (cs.AI/cs.CL) by Taniguchi and colleagues offering an explanation for how large language models acquire rich world knowledge despite having no sensorimotor experience. It propo…