SSPO: A New Method to Improve Neural Combinatorial Optimization Training

A new research paper introduces SSPO (Structure-Aware Similarity-Weighted Preference Optimization), a method designed to improve the training of neural combinatorial optimization (NCO) models. The paper, available on arXiv (2608.12443), identifies two key failures in existing training approaches: gradient signal polarization and baseline redundancy. Gradient signal polarization occurs when preference-optimization methods focus only on the best solution, discarding quality and structural information from other sampled solutions. Baseline redundancy arises when mean-based baselines weight all solutions equally, causing structurally similar solutions to flood the baseline with redundant information and increase gradient variance. SSPO addresses both issues by scoring all sampled solutions jointly through a dissimilarity-weighted leave-one-out baseline, giving higher weight to structurally distinct peers. This mechanism uses zero-parameter, problem-adaptive solution embeddings, making it efficient and adaptable. The paper is authored by researchers and was announced as a cross-type submission on arXiv. The method has potential implications for improving the efficiency and effectiveness of NCO, which is used in various optimization problems such as routing, scheduling, and resource allocation.