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..
The Persistence of Neural Collapse Despite Low-Rank Bias
Authors: Connall Garrod, Jonathan Keating
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our results are validated through experiments in deep UFMs and deep neural networks. 4 Numerical experiments We empirically validate our theoretical results in the deep UFM, and provide supporting experimental evidence on the MNIST [28] and CIFAR-10 [26] datasets using both UFM-style and standard regularization. |
| Researcher Affiliation | Academia | Connall Garrod Mathematical Institute University of Oxford EMAIL Jonathan P. Keating Mathematical Institute University of Oxford |
| Pseudocode | No | The paper primarily presents theoretical results in the form of theorems and their proofs in Appendix B, without including any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | 5. Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: Our contributions are mostly theoretical, using experiments to illustrate the theory. The source code will be provided on request. |
| Open Datasets | Yes | 4 Numerical experiments We empirically validate our theoretical results in the deep UFM, and provide supporting experimental evidence on the MNIST [28] and CIFAR-10 [26] datasets using both UFM-style and standard regularization. |
| Dataset Splits | Yes | For experiments on MNIST, we subsample 5,000 examples per class to match the class balance of the CIFAR-10 dataset. Input data is preprocessed by subtracting the mean and dividing by the standard deviation. We use batch gradient descent with batch size 10,000 so as to approximate gradient descent, which is used in the model. |
| Hardware Specification | No | 8. Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [No] Justification: The experiments do not require any specific hardware setup. |
| Software Dependencies | No | The paper does not explicitly mention any specific software dependencies with version numbers for its experiments. |
| Experiment Setup | Yes | Figure 1: Experiments in the deep linear UFM: ... Hyperparameters: L = 2, d = 70, λ = 2 10, K = 10, n = 5, learning rate = 0.5. Figure 3: Experiments using UFM-style regularization: ... Hyperparameters: L = 3, d = 64, λW = 5 10 3, λH = 10 6, learning rate = 0.05. ... Bottom: Loss and singular values across layers for CIFAR-10 using Re LU in the fully connected head. Same hyperparameters as above, except L = 4. Figure 4: Experiments with standard regularization on CIFAR-10: ... Hyperparameters: L = 3, d = 64, λ = 10 2, learning rate = 10 3. ... We use batch gradient descent with batch size 10,000 so as to approximate gradient descent, which is used in the model. |