Reproduce “Align Forward, Adapt Backward: Closing the Discretization Gap in Logic Gate Networks” 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

Youngsung Kim

Authors introduce forward alignment in logic-gate networks: each neuron selects one Boolean gate in the forward pass, while a smooth surrogate supplies backward gradients. This keeps gate selection aligned with deployment, though continuous inputs can retain a computation gap. On CIFAR-10-Binary, across five temperatures, Gumbel-ST+CAGE had 58.0% worst accuracy and a 0.6-point range, versus 18.9% and 39.5 points for Gumbel-ST; these are sweep results, not hardware-speed measurements.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.