Compositional Reuse of Semantic Change Relations in World Model Dynamics

Abstract

Compositional generalization asks whether familiar structure can still be reused when familiar pieces appear in combinations absent during fitting. In world model dynamics, this becomes a question of whether directed effects among semantic changes can be recombined without learning every pairing separately. Message suppression reveals the individual effects in a common prediction space, where their sum can be compared with the ordered joint effect. The reconstruction stays accurate even when the exact ordered pair is withheld during fitting while both constituents remain familiar. A frozen test on fresh contexts reproduces the pattern. Overall, the evidence supports functional reuse across unseen ordered combinations while leaving deeper composition, causal claims, and transfer to other domains open.