Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Geometric Algorithms for Neural Combinatorial Optimization with Constraints
Authors: Nikolaos Karalias, Akbar Rafiey, Yifei Xu, Zhishang Luo, Behrooz Tahmasebi, Connie Jiang, Stefanie Jegelka
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments in cardinality-constrained optimization show that our approach can consistently outperform neural baselines. We further provide workedout examples of how our method can be applied beyond cardinality-constrained problems to a diverse set of combinatorial optimization tasks, including finding independent sets in graphs, and solving matroid-constrained problems. |
| Researcher Affiliation | Academia | Nikolaos Karalias MIT EMAIL Akbar Rafiey NYU EMAIL Yifei Xu NYU EMAIL Zhishang Luo UCSD EMAIL Behrooz Tahmasebi MIT EMAIL Connie Jiang MIT EMAIL Stefanie Jegelka TUM and MIT EMAIL |
| Pseudocode | Yes | Algorithm 1 General decomposition algorithm Algorithm 2 Decomposition for partition matroid Algorithm 3 Decomposition for graphical matroid |
| Open Source Code | Yes | All the datasets are public and our code is available at https://github.com/frankx2023/Neural_Combinatorial_Optimization_with_ Constraints. |
| Open Datasets | Yes | All the datasets are public and our code is available at https://github.com/frankx2023/Neural_Combinatorial_Optimization_with_ Constraints. ... The IMDB-BINARY dataset [92] consists of 1000 graphs. The Erd os Rényi dataset consists of 1000 synthetic graphs generated under the G(n, p) model. |
| Dataset Splits | Yes | Our experiments use the following split for both datasets. 60% of the graphs are allocated for training, 20% for validation, and the final 20% for testing. ... Real-world graphs: ... Due to limited large-scale data, real-world datasets are used only for testing, with training done on synthetic data. |
| Hardware Specification | Yes | All experiments are run on 16 cores (32 threads) of Intel(R) Xeon(R) Platinum 8268 CPU (24 cores, 48 threads in total), 32 GB ram, with a single Nvidia RTX8000 48GB GPU. |
| Software Dependencies | Yes | Gurobi (120s) ... The version used is Gurobi 12.01. ... Optimizer: Adam W. |
| Experiment Setup | Yes | Dropout: 0.1 (applied only during training). Optimizer: Adam W. Learning Rate: 5 10 3. Learning Rate Scheduler: warmup_cosine, with 50 warmup epochs, decaying to a minimum of 5 10 5. Weight Decay: 1 10 4. Epochs: 80. Batch Size: 4. Seed: 42. Workers: 16 CPU worker threads for data loading, preprocessing, and decomposition with multiple scaling factors. |