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..
Graph-Based Attention for Differentiable MaxSAT Solving
Authors: Sota Moriyama, Katsumi Inoue
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on (weighted) Max SAT benchmark datasets demonstrate that SGATs significantly outperform existing neural-based architectures, and achieve state-of-the-art performance among continuous approaches, highlighting the strength of the proposed model. |
| Researcher Affiliation | Academia | Sota Moriyama1,2 Katsumi Inoue2 1The Graduate University for Advanced Studies, SOKENDAI 2National Institute of Informatics EMAIL |
| Pseudocode | Yes | Algorithm 1: LS-GNN In : Weighted Max SAT instance F, timeout T. Out : Best valuation vbest and its cost Cbest. Algorithm 2: SLS solver with SGAT-Based initialization Input :Max SAT instance F, timeout T. Output :Best solution found xbest and its cost Cbest. Algorithm 3: SGAT-Based Initialization Input :Max SAT instance F. Output :Initial assignment x. |
| Open Source Code | Yes | All used code is contained in the repository: https://github.com/sotam2369/SGAT-MS |
| Open Datasets | Yes | For evaluation, we use instances used in non-partial unweighted and weighted benchmark instances provided in Max SAT evaluations2, which we denote as MS and WMS (and collectively referred to as WMS+). 2https://maxsat-evaluations.github.io We also prepare a subset of these datasets with instances that are below specific file sizes such as 2MB for purposes such as training, and denote as WMS+(2MB). We compare the model performance on the MS2018(2MB) dataset, with training done on two different datasets: (i) MS2018(2MB), and (ii) SR(U(40,200)), a randomly generated dataset with 40 to 200 variables [22]. |
| Dataset Splits | Yes | The training and testing splits are shown in Appendix D. Appendix D: The train/test splits for Max SAT 2018 is given in the repository. |
| Hardware Specification | Yes | All experiments were done on a machine with AMD Ryzen Threadripper PRO 3975WX 32-Cores and two NVIDIA RTX A6000 GPUs. |
| Software Dependencies | No | The paper mentions using 'Adam optimizer' and 'Py Torch Geometric' [7] and 'G4SATBench framework' [15], but it does not specify version numbers for these software components or libraries. |
| Experiment Setup | Yes | We used SGATs with 6 SGAT blocks composed with Gödel T-norm layers and SGAT layers with 2 attention heads and 4 channels. For training, we used the Adam optimizer with a learning rate of 2 10 3, and a batch size of 4. For existing models, the default provided settings were used. |