Reproduce “Interpreting Neural Combinatorial Optimization via Evolving Programmatic Bottlenecks” 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
Haocheng Duan, Yuxin Guo, Jieyi Bi, Anqi Xie, Sirui Li, Yining Ma, Cathy Wu
The Evolving Programmatic Bottlenecks (EPB) framework interprets black-box Neural Combinatorial Optimization (NCO) policies by distilling their learned behaviors into a portfolio of human-readable programs. This approach achieved solution quality comparable to teachers and improved out-of-distribution generalization for the POMO model by up to 15.22% on larger problem instances.
Key insights
- Programmatic distillation can isolate transferable decision patterns from black-box NCO models, leading to improved out-of-distribution generalization compared to the teacher model.
- NCO policies learn distinct, interpretable strategies specific to the solver (e.g., POMO's phase-adaptive heuristics vs. LEHD's local geometry-driven corrections).
- Distilled policies reveal clear, evolving decision logic across different stages of solution construction (e.g., capacity-dependent heuristic shifts in CVRP).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.