Simulation benefits and physics-based reconstruction requirements

Determine when simulation improves reconstruction quality and whether requiring physics-based simulation preserves the relative performance of multimodal coding models on dynamic-scene reconstruction tasks.

Background

The benchmark evaluates agents that reconstruct dynamic scenes as executable 4D graphics programs, but it permits diverse implementations of motion, including analytic trajectories, keyframing, Blender physics, and custom physical simulation. The experiments show that models adopt substantially different strategies, with analytic motion predominating and custom simulation used less frequently.

The paper does not resolve whether physically based simulation produces more accurate or transferable reconstructions than prescribed or analytic motion, nor whether imposing a physics-simulation requirement would change the comparative ranking of the evaluated models. The authors explicitly identify both issues as unresolved questions for future investigation.