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..
Probing Equivariance and Symmetry Breaking in Convolutional Networks
Authors: Sharvaree Vadgama, Mohammad Islam, Domas Buracas, Christian A Shewmake, Artem Moskalev, Erik Bekkers
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We address this through theoretical analysis and a comprehensive empirical study focusing on point clouds. ... Experiments on molecular property prediction and generation, 3D part segmentation and shape generation, and motion prediction show that more constrained equivariant models outperform less constrained alternatives when aligned with task geometry. |
| Researcher Affiliation | Collaboration | 1AMLab, University of Amsterdam 2 New Theory AI 3 Qur AI, University of Amsterdam 4 UC Berkeley 5 Independent Researcher |
| Pseudocode | No | The paper describes methods and operations but does not present any structured pseudocode or algorithm blocks. Figure 6 shows a "Block design with base space R3 S2" which is a diagram, not pseudocode. |
| Open Source Code | Yes | Code available at github.com/Sharvaree/Equivariance Study. |
| Open Datasets | Yes | Molecular property prediction & generation task For predicting molecular properties and generating molecules, we use QM9 [Ramakrishnan R., 2014]. ... 3D point cloud segmentation & generation task We evaluate our models on Shape Net3D [Chang et al., 2015] for part segmentation and generation tasks. ... Human motion prediction task We next evaluate the model variations on the CMU Human Motion Capture dataset [Gross and Shi, 2001]... Classification task on Modelnet40 dataset Model Net40 dataset [Wu et al., 2015] contains 9,843 training and 2,468 testing meshed CAD models belonging to 40 categories. |
| Dataset Splits | Yes | QM9 dataset Ramakrishnan R. [2014] contains up to 9 heavy atoms and 29 atoms, including hydrogens. We use the train/val/test partitions introduced in Gilmer et al. [2017], which consists of 100K/18K/13K samples respectively for each partition. ... For Shape Net 3D, We use the point sampling of 2,048 points and the train/validation/test split from [Qi et al., 2017]. ... Model Net40 dataset [Wu et al., 2015] contains 9,843 training and 2,468 testing meshed CAD models belonging to 40 categories. |
| Hardware Specification | Yes | A pool of GPUs, including A100, A6000, A5000, and 1080 Ti, was utilized as computational units. |
| Software Dependencies | No | We implemented our models using Py Torch [Paszke et al., 2019], utilizing Py Torch-Geometric s message passing and graph operations modules [Fey and Lenssen, 2019], and employed Weights and Biases for experiment tracking and logging. The paper mentions software frameworks (PyTorch, PyTorch-Geometric) and a tool (Weights and Biases) but does not provide specific version numbers for these dependencies. |
| Experiment Setup | Yes | For all experiments, we use Rapidash with 7 layers with 0 fiber dimensions for R3 and 0 or 8 fiber dimensions for R3 S2. The polynomial degree was set to 2. We used the Adam optimizer [Kingma and Ba, 2014], with a learning rate of 1e 4, and with a Cosine Annealing learning rate schedule with a warm-up period of 20 epochs. ... For Shape Net 3D, All the models were trained for 500 epochs with a learning rate of 5e 3 and weight decay of 1e 8. ... For CMU Motion Prediction, All the models were trained for 1000 epochs with a learning rate 5e 3 and weight decay of 1e 8. |