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..
When Does Curriculum Learning Help? A Theoretical Perspective
Authors: Raman Arora, Yunjuan Wang, Kaibo Zhang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this work, we develop a theoretical framework for curriculum learning based on biased regularized empirical risk minimization (RERM), identifying conditions under which curriculum learning provably improves generalization. ... Empirical evaluations on synthetic datasets and MNIST validate our theoretical findings and highlight the practical efficacy of curriculum-based training. |
| Researcher Affiliation | Academia | Raman Arora Johns Hopkins University Baltimore, MD 21218 EMAIL Yunjuan Wang Johns Hopkins University Baltimore, MD 21218 EMAIL Kaibo Zhang Johns Hopkins University Baltimore, MD 21218 EMAIL |
| Pseudocode | Yes | Algorithm 1 Biased Regularization-based Curriculum Learning Algorithm 2 Warm-up: A Two-task Curriculum Algorithm 3 SGD for task t Algorithm 4 ERM-based Curriculum Learning |
| Open Source Code | Yes | E. 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: [Yes] Justification: We attach the code of our experiments in the supplementary material. |
| Open Datasets | Yes | Empirical evaluations on synthetic datasets and MNIST validate our theoretical findings and highlight the practical efficacy of curriculum-based training. |
| Dataset Splits | Yes | We hold out 20% of the training data as a validation set and use 10-step PGD adversarial examples, crafted with the same perturbation budget α, for hyper-parameter tuning and model selection. |
| Hardware Specification | Yes | H. 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: [Yes] Justification: All experiments are conducted on a single V100 GPU. |
| Software Dependencies | No | The paper mentions optimizers (Adam, SGD) and loss functions (cross-entropy), but does not provide specific version numbers for key software components like Python, PyTorch, or TensorFlow, which are necessary for reproducible dependency information. |
| Experiment Setup | Yes | Linear classifiers are trained using hinge loss and gradient descent (2K epochs, learning rate from 0.001, . . . , 1.0). The baseline trains only on D2, while our curriculum method (Algorithm 2) first trains on D1 and then fine-tunes on D2 with ℓ2 regularization λ w2 bw1 2, where bw1 is the solution from the first stage. λ is selected from {10 5, 10 4, 10 3, 10 2, 10 1, 1, 10} using validation data. |