Reproduce “Internal Guidance as a Sampling Method for Tilted Diffusion Models” 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
Curtis McDonald, Peter Bartlett
The authors introduce internal guidance for steering diffusion models on binary-hypercube data toward reward-scoring examples or partial completions. Instead of adding an external sampling correction, the method trains one denoiser on shifted and rescaled inputs. In 9×9 Sudoku, it achieved valid grids retaining all 11 given clues in 80.3% of draws. This is a soft-tilt result, not a guarantee.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.