Spatial Memory Agent: a frozen VLM self-evolves spatial reasoning through verified lessons, with no fine-tuning and no 3D tools at inference

arXiv·low signalXBlueskyLinkedInCopy link

SMA (arXiv 2608.12743, 2026-08-13) explores a third route between post-training and tool-calling for spatial reasoning: parameter-update-free self-evolution. In a verifiable environment it queries the frozen VLM, obtains a prediction and reward, and uses verifier-guided reflection to distill compact transferable lessons from experience; each lesson carries a Transfer Reliability Score initialized uniformly and calibrated from later retrieval outcomes. At read-only deployment time lessons are retrieved by semantic filter and reliability, so no depth-estimation or 3D-reconstruction tool is needed at inference. The pattern generalizes past spatial tasks — a reliability score learned from downstream retrieval outcomes is a cheap way to stop a memory store filling with lessons that never help.