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..
Revisiting Orbital Minimization Method for Neural Operator Decomposition
Authors: Jongha (Jon) Ryu, Samuel Zhou, Gregory Wornell
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the OMM and its variants (including the Sanger variant) across three experimental setups. ... We empirically demonstrate the effectiveness of our method across a range of tasks, including learning Laplacian-based representations in reinforcement learning settings, solving PDEs, and self-supervised-learning representation of images and graphs. |
| Researcher Affiliation | Academia | J. Jon Ryu, Samuel Zhou, Gregory W. Wornell Department of EECS, MIT, Cambridge, MA 02139, United States EMAIL |
| Pseudocode | Yes | B.2 Pseudocode for Nested OMM-1 Here, we include a unified Py Torch [39] implementation of various versions of the original OMM with p = 1 (which we call OMM-1): OMM-1 without any nesting (when nesting is None), OMM-1 with the joint nesting (when nesting is jnt), OMM-1 with the sequential nesting (when nesting is seq), and OMM-1 with the Sanger variant (when nesting is sanger). |
| Open Source Code | Yes | Our Py Torch implementation is available at https://github.com/jongharyu/ neural-svd. |
| Open Datasets | Yes | We consider the same suite of experiments from [13], which consists of several grid environments as visualized in Appendix C.1. ... We consider the simplest example of this kind: the two-dimensional hydrogen atom... We used Res Net-18 as our backbone model and adopted two different feature encoding strategies... We used the ogbn-products dataset [18] |
| Dataset Splits | Yes | We evaluate the representation based on the linear probe accuracy on the test split, trained by SGD with learning rate 0.1, batch size 256, and 100 epochs. ... we evaluate the representation based on the linear probe accuracy on the test split |
| Hardware Specification | Yes | All experiments were conducted on a single GPU, either a NVIDIA Ge Force RTX 3090 (24GB) or a NVIDIA RTX A6000 (48GB). |
| Software Dependencies | No | Our Py Torch implementation is available at https://github.com/jongharyu/ neural-svd. ... We trained for 80,000 epochs using Adam [22] with learning rate 10^-3. ... We used the LARS optimizer [56] with weight decay set to 0, initial learning rate of 0.3 governed by a cosine decay schedule, batch size 256, and 1000 epochs. |
| Experiment Setup | Yes | We trained for 80,000 epochs using Adam [22] with learning rate 10^-3. ... We trained the neural networks for 10^5 iterations. ... We used Adam optimizer [22] with learning rate 10^-4 with the cosine learning rate scheduler... We use the LARS optimizer [56] with weight decay set to 0, initial learning rate of 0.3 governed by a cosine decay schedule, batch size 256, and 1000 epochs. |