Reproduce “Fast T2T: Optimization Consistency Speeds Up Diffusion-Based Training-to-Testing Solving for Combinatorial Optimization” 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
Yang Li, Jinpei Guo, Runzhong Wang, Hongyuan Zha, Junchi Yan
Researchers from Shanghai Jiao Tong University and MIT introduce "Fast T2T," a diffusion-based neural solver for combinatorial optimization that employs an optimization consistency training protocol to learn direct mappings from noisy states to optimal solutions. This approach enables faster solution generation and iterative refinement through a consistency-based gradient search, leading to significant speedups (up to 82.8x) and improved solution quality compared to previous diffusion-based methods on problems like TSP and MIS.
Key insights
- Directly learning a consistency mapping from noisy states to the optimal solution, rather than the reverse diffusion process, drastically reduces inference steps required for high-quality solution generation.
- Integrating the problem's objective function into a consistency-based gradient search during the testing phase significantly enhances solution quality and instance-specific generalization.
- By enforcing consistency across different noise levels and generative trajectories towards the ground-truth optimal solution, the model learns a more robust and efficient solution distribution.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.