Test-Time Scaling via Latent World Simulation with Process-Verifiable Reinforcement
A test-time inference framework that performs iterative spatial reasoning by selecting spatial transformations, executing them in a latent world model, and feeding aligned latent intermediate evidence back into the VLM—avoiding repeated latent-to-pixel-to-latent decoding. It further uses process-verifiable reinforcement learning to reward actions only when both the action and the resulting latent transition are valid.
Discovered 2026-11-10 · Sources checked 2026-11-10
Why it matters
It targets perspective-taking and hypothetical viewpoint/position changes by enabling active scene transformations in a latent space, while reducing the cost and information bottleneck of going through synthesized pixel images at every step. The process-verifiable reward helps keep the simulated transformations consistent with valid transitions during reasoning.
What changed
Instead of scaling only language reasoning or relying on static geometric augmentations, the method allocates extra inference-time computation to an active latent world simulation loop: (1) the VLM proposes spatial transformations, (2) a world model applies them in latent space, (3) the VLM consumes aligned latent evidence for subsequent steps, and (4) reinforcement learning enforces validity of the action/latent transition during simulation.
Where it helps
- Perspective-taking spatial reasoning under hypothetical viewpoint changes
- Reasoning tasks requiring mental spatial transformations of a represented scene
- Iterative multi-step simulation where repeated pixel decoding is undesirable
- Latent-space planning where action validity and transition validity can be checked
Evidence & reproduction
Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.
Limitations & open questions
The provided description does not specify architectural details, training setup, or the exact mechanics of the process-verifiable reward beyond the stated criterion of valid action and valid latent transition.
Related World Models papers
Sources
- WorldMind: Test-Time Scaling for Visual Spatial Reasoning through Latent World Simulation ↗
arXiv · primary · Checked 2026-11-10
