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..
SONAR: Long-Range Graph Propagation Through Information Waves
Authors: Alessandro Trenta, Alessio Gravina, Davide Bacciu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on synthetic and realworld benchmarks confirm that SONAR achieves state-of-the-art performance, particularly on tasks requiring long-range information exchange. ... In this section, we empirically validate the practical benefits of our method on popular graph benchmarks for long-range propagation as well as heterophilic node classification tasks. In Sections 4.1 and 4.2, we assess SONAR on synthetic benchmarks... In Section 4.5, we empirically assess the long-range capabilites of SONAR in terms of the sensitivity metric (discussed in Section 3.3). In Appendix C.2, we report additional ablation studies to provide a more comprehensive understanding of SONAR, discussing runtimes and the role of adaptive resistance, dissipation, external forces, and step size. |
| Researcher Affiliation | Academia | Alessandro Trenta Department of Computer Science University of Pisa Pisa, Italy EMAIL Alessio Gravina Department of Computer Science University of Pisa Pisa, Italy EMAIL Davide Bacciu Department of Computer Science University of Pisa Pisa, Italy EMAIL |
| Pseudocode | No | The paper describes the methodology using mathematical equations and textual descriptions, but it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | All the experiments are performed on a server with NVIDIA H100 GPUs. We openly release the code at https://github.com/gravins/SONAR. |
| Open Datasets | Yes | Specifically, we consider long-range propagation tasks, including the graph transfer tasks from [40], as well as the three graph property prediction tasks introduced in [39] ( Diameter , SSSP , and Eccentricity ). Additionally, we assess SONAR on the Peptide-func , Peptide-struct , and Pascal VOC-SP tasks from the real-world Long-Range Graph Benchmark (LRGB) [24]. To further evaluate its effectiveness, we include five heterophilic tasks: Roman-empire , Amazon-ratings , Minesweeper , Tolokers , and Questions [70]. |
| Dataset Splits | Yes | The setup of each experiment is described in Section 4. For each of them, we provide a reference to the original paper that described and analyzed these datasets and tasks. ... We adhere to the same data and experimental setting presented in [70]. ... We use the same data and experimental setting in [24], including the 500k parameter budget. |
| Hardware Specification | Yes | All the experiments are performed on a server with NVIDIA H100 GPUs. ... All runtimes are measured on an NVIDIA H100 GPU. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies. |
| Experiment Setup | Yes | The hyperparameter space employed by SONAR in our experiments is reported in Table 6. ... Table 6: The grid of hyperparameters employed during model selection for the graph transfer tasks (Transfer), graph property prediction tasks (GPP), Long Range Graph Benchmark (LRGB), and heterophilic benchmarks (Hetero). Hyperparameters Values: Optimizer Adam Adam Adam W Adam W Learning rate 0.001 0.003 0.001 0.001, 0.0005 Weight decay 0 10 6 0 0 Dropout 0 0 0, 0.2, 0.5 0, 0.2, 0.5 N. recurrences distance/2, distance 5, 10, 20 1,2,4,6,8,12,16 1,2,4,6,8,12,16 Embedding dim 64 30 60, 68, 74, 105, 117, 136 64, 128, 256, 512 N. Blocks 1,2 1, 2 1,2,3,4,5 from 1 to 12 ̈ 0.05, 0.1, 0.5, 1 0.001, 0.1, 0.5, 1 0.01, 0.02, 0.025, 0.05, 0.1 0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05, 0.1 Use Dissipation True, False Use External Force True, False Use Adaptive Resistance True, False L D A, I D 1/2AD 1/2, I D 1A |