Scaling self-supervised contrastive reinforcement learning (CRL) to 1000-layer Euclidean neural networks has delivered state-of-the-art goal-reaching performance, according to a new arXiv paper. However, the authors note that wall-clock time scales linearly with depth, making such very deep networks costly to train.

The paper, titled "Outperforming 1000 Euclidean Layers With Hyperbolic Goal-Conditioned RL," suggests that changing the geometric setting — from Euclidean to hyperbolic — for goal-conditioned RL can outperform those 1000-layer Euclidean models. The title and abstract indicate that this hyperbolic approach is the central contribution, though the abstract is truncated and does not specify the exact method or experimental details.

Because the source only provides the opening of the abstract, the full findings, including how the hyperbolic representation is implemented and the magnitude of the improvement, are not yet available from this announcement. The paper is posted on arXiv under identifier 2610.20009.