Reproduce “Learning Iterative Reasoning through Energy Diffusion” 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
Yilun Du, Jiayuan Mao, Joshua B. Tenenbaum
IRED combines energy-based models with diffusion-inspired techniques to learn iterative reasoning, formulating problems as energy minimization and refining solutions through annealed energy landscapes. It achieved superior performance on matrix completion, Sudoku solving, and pathfinding, exhibiting generalization to harder instances and adaptive computation.
Key insights
- Reasoning problems can be formulated as energy minimization tasks, where optimal solutions correspond to configurations with minimal energy.
- Learning a sequence of annealed energy landscapes, progressively transitioning from smooth to complex, enables stable optimization and better generalization for iterative reasoning.
- Combining score function supervision (training gradients to denoise) and energy landscape supervision (contrastive loss) effectively trains the energy functions for diverse reasoning tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.