Disentangling gradient quality from architecture in recursive reasoning. Controlled experiment: 1-step gradient approximation is the sole bottleneck in HRM vs TRM performance gap.
-
Updated
Jun 16, 2026 - Jupyter Notebook
#
gradient-methods
Here are 3 public repositories matching this topic...
Completing the 2x2 factorial: HRM's hierarchy with full BPTT reveals gradient-architecture interaction
-
Updated
Jun 16, 2026 - Jupyter Notebook
A real BFGS curvature-condition bug, demonstrated: one character sends the optimiser uphill by six orders of magnitude. Two files, one command, no dependencies.
-
Updated
Nov 2, 2026 - JavaScript
Add this topic to your repo
To associate your repository with the gradient-methods topic, visit your repo's landing page and select "manage topics."