Reproduce “Form Follows Function: Recursive Stem Model” 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
Navid Hakimi
Form Follows Function introduces Recursive Stem Model (RSM), a compact weight-shared architecture that performs reasoning through nested updates of fast and slow latent states. Terminal-only supervision, detached warm-up iterations, and depth curricula allow the model to train with shallow gradient paths while using much deeper recursion at inference, reaching reported exact accuracies of 97.5% on Sudoku-Extreme and about 80% on 30×30 Maze-Hard.
Key insights
- Terminal-only supervision is intended to preserve exploratory intermediate states instead of forcing every recursive step to resemble a valid final answer.
- The separation between training depth and inference depth makes computation a controllable test-time resource; the reported setup uses roughly 20 training outer steps versus as many as 20,000 at inference.
- Sudoku trajectories often become unchanged after finding a valid solution, whereas maze trajectories appear to first form a path and then refine it, suggesting task-dependent iterative behavior.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.