Scaling Video Generation for Reasoning: At What Cost?
Published in arXiv preprint, 2026
We evaluate whether video generators can track hidden state using a simulated 2x2x2 Rubik’s Cube with exact ground truth. Plausible geometry emerges before correct sticker configurations, and lower validation loss does not reliably predict reasoning accuracy. Smaller models are more efficient at limited compute; larger models improve with more training, while symbolic state supervision offers a complementary route to stronger reasoning.
Recommended citation: Guo, W., Wu, X., Wang, Y., Mireshghallah, N., and Kavraki, L. E. (2026). "Scaling Video Generation for Reasoning: At What Cost?" arXiv:2609.36599.
Download Paper