Reproduce “Outperforming 1000 Euclidean Layers With Hyperbolic Goal-Conditioned RL” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Timo Klein, Thomas Lang, Andrii Shkabrii, Lukas Miklautz, Yllka Velaj, Sebastian Tschiatschek
The authors introduce hyperbolic contrastive reinforcement learning (CRL) for goal-conditioned maze navigation, scoring states and goals by hyperbolic rather than Euclidean distance. In Humanoid U-Maze, a depth-64 agent stays near the goal for 240 of 1,000 steps, versus around 100 for a depth-1024 Euclidean agent. The advantage depends on task geometry; flatter curvature matches Euclidean performance in an open arena.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.