Rolling Geometry, Persistent Semantics: Structured Reasoning for Driving Risk Evolution

Embodied agents must reason about scenes that change as they move, yet temporal models often represent dynamic environments as repeated spatial features or latent histories. In autonomous driving, this entangles changes in Occupancy, Occlusion, Interaction, and Motion with changes in the ego-centric frame, while adding redundant context. We introduce a structured state-evolution framework based on the insight that temporal persistence should reside in factor-specific semantics rather than spatial-token identities. Rolling Base/Fine geometry preserves spatial detail, and current-anchor Carriers compress evidence. A frozen language model updates four persistent memories from State and semantic Delta Tokens together with Frame Transition. Two stages separate Past-to-Current updating from Current-to-Future inference, where horizon conditions use only observed ego motion and structured visibility. A lightweight grounder maps five inferred future states onto the current geometry to produce continuous collision-risk fields from $1$ to $5$\,s. On AV2, the framework achieves $19.0682\times$ future-risk AP Lift and $0.1879$ Avg. MAE, together with $3.9772\times$ risk-evolution AP Lift and $0.2383$ risk-evolution MAE. Relative to Full-History Qwen, its Qwen-token count and latency change by -34.6\% and -41.5\%, respectively. These results support persistent semantic state evolution as an effective interface for dynamic embodied spatial reasoning in driving scenes.