The paper proposes Astra, an agentic spatial reasoning framework that couples a vision-language model (VLM) policy with a world simulator. The VLM policy (Astra-VL) is trained with reinforcement learning to decide when to request imagined observations from the world simulator (Astra-WM), which generates novel-view images conditioned on natural-language camera motions. The world simulator is fine-tuned with view consistency tuning for spatial accuracy.
Experiments show that both components are necessary: Astra-WM improves Gemini-3-Flash on MMSI-Bench from 45.1 to 49.5, and Astra-VL improves Qwen3-VL on MMSI-Bench from 29.8 to 38.8 and on MindCube from 36.8 to 42.7. The results are interesting because they demonstrate that imagined observations provide useful spatial evidence, but effective world-model-augmented reasoning requires learning when, where, and how to invoke the simulator.