📝 Selected Publications
Wenzheng Pan, Shuyi Yan, Nuoyan Chen, Yiyang Qu, Jiale Ma, Junchi Yan
We propose UECO, a unified encoder for neural combinatorial optimization with structure-aware attention mixtures and iterative edge evolving. By progressively combining local topological messages with global attention, UECO serves as a shared backbone across construction, prediction, and expansion paradigms and improves solution quality and generalization across diverse combinatorial problems.
Rectified Policy Rollouts with Hierarchical Expert Guidance for Neural Combinatorial Optimization
Wenzheng Pan, Nuoyan Chen, Jiaxi Liu, Jiale Ma, Junchi Yan
We introduce CORectifier, a hierarchical gated rectification mechanism that replaces selected policy actions with high-quality reference segments during reinforcement learning. The resulting rectified policy rollouts inject optimality signals at batch, instance, and sub-instance levels, improving sample efficiency and solution quality while preserving sequential constraint satisfaction.
NLD4CO: Neural Langevin Dynamics for Combinatorial Optimization
Jiale Ma, Wenzheng Pan, Binghao Cai, Xihe Zhang, Junchi Yan
We propose NLD4CO, a unified framework that combines Langevin dynamics with data-driven learning for combinatorial optimization. Its explicit-gradient variant provides neural warm starts for energy-based problems, while the implicit-gradient variant uses consistency-model corrections to guide globally coordinated search in general combinatorial problems.
Problem Distributions as Tasks: Repurposing Meta Learning for Generative Combinatorial Optimization towards Multi-task Pretraining and Adaptation [PDF][Code 
Wenzheng Pan, Jiale Ma, Nuoyan Chen, Yang Li, Junchi Yan
We introduce M²GenCO, a meta-generative framework that treats problem distributions as tasks to enable efficient multi-task pretraining, few-shot adaptation, and robust generalization across graph-based combinatorial optimization problems.
Design Linear Constrained Neural Layers with Implicit Convex Optimization [PDF]
Junchi Yan, Jiaxi Liu, Yihui Tu, Fangyuan Zhou, Wenzheng Pan, Zhongteng Gui, Liangliang Shi
We propose LinConLayer, a plug-in differentiable neural layer that enforces general linear constraints via implicit convex optimization, yielding efficient BLCLayer and GLCLayer variants for constrained prediction in tasks such as graph matching, portfolio allocation, and linear programming.
ML4CO-Bench-101: Benchmark Machine Learning
for Classic Combinatorial Problems on Graphs [PDF][Code 
Jiale Ma, Wenzheng Pan, Yang Li, Junchi Yan
We establishe ML4CO-Bench-101, a standardized benchmark and modular evaluation framework that systematically categorizes, reproduces, and compares neural solvers across seven mainstream graph-based combinatorial optimization problems.
COExpander: Adaptive Solution Expansion for Combinatorial Optimization [PDF][Code 
Jiale Ma*, Wenzheng Pan*, Yang Li, Junchi Yan
We introduce COExpander, an adaptive expansion paradigm that bridges global prediction and local construction by progressively determining decision variables with dynamically controlled step sizes for scalable combinatorial optimization.
UniCO: On Unified Combinatorial Optimization via Problem Reduction to Matrix-Encoded General TSP [PDF] [Code 
Wenzheng Pan*, Hao Xiong*, Jiale Ma, Wentao Zhao, Yang Li, Junchi Yan
We propose UniCO, a unified neural combinatorial optimization framework that reduces diverse COPs into matrix-encoded general TSP and solves them with tailored matrix-based RL and diffusion solvers: 1) MatPOENet, an RL-based sequential model with pseudo one-hot embedding (POE) scheme and 2) MatDIFFNet, a Diffusion-based generative model with the mix-noised reference mapping scheme.
Unify ML4TSP: Drawing Methodological Principles for TSP and Beyond from Streamlined Design Space of Learning and Search [PDF][Code 
Yang Li, Jiale Ma, Wenzheng Pan, Runzhong Wang, Haoyu Geng, Nianzu Yang, Junchi Yan
We present ML4TSPBench, a modular framework that decomposes learning-based TSP solvers into reusable learning and search components, revealing key design principles for stronger and more principled ML4CO methods.
Pygmtools: A Python Graph Matching Toolkit [PDF][Code 
Runzhong Wang, Ziao Guo, Wenzheng Pan, Jiale Ma, Yikai Zhang, Nan Yang, Qi Liu, Longxuan Wei, Hanxue Zhang, Chang Liu, Zetian Jiang, Xiaokang Yang, Junchi Yan
We release Pygmtools, an open-source Python toolkit that unifies classical, multi-graph, and learning-based graph matching solvers across multiple numerical backends for research and practical applications.