The paper appears in the International Conference on Learning Representations (ICLR), 2026. Paper: arXiv · Code: GitHub