A Declared-World Audit of the Grammar a Joint-Embedding World Model Learns
Year: 2026
Type: preprint
Abstract: Human languages share a grammar of things, actions, states and relationships. We ask which part of that grammar a joint-embedding world model learns from pixels, using STAR (States, Things, Actions, Relationships), the proto-grammar at the foundation of the BOLT methodology, as the comparison frame. The subject is a scaled-down LeWorldModel with a single 8-dimensional latent, trained in two worlds built from declared state machines. One is a two-room world; in the other, a door that is open, ajar, closed or broken decides which moves are refused. Seventeen declared measurements score the frozen latents against the worlds’ answer keys, calibrated by a supervised encoder, a random network and a learner given the true objects. The findings replicated on unseen seeds. Actions are local linear transformations that compose but do not invert. Relationships are places in a positional code whose link between the rooms leaks, and a door component that weakly modulates the agent’s motion without blocking it. Outcomes are implicit: at the wall the predictor expects the agent to stay, which neither the supervised encoder nor the random network does, but at a closed door it does not. By the consistency a grammar requires, this outcome grammar falls short. Things are not individuated, even read nonlinearly. A STAR-structured learner did not supply them, while a flat learner given the true objects planned far better. We outline how the grammar could be extended, one STAR element at a time, and measured in the same worlds. The study ran on limited computation; its worlds and measurements are built to be reused on larger world models.
Source: Preprints.org
Author: Jean-Jacques Dubray
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7214724568