Research map / Large language models / Language, safety and interpretability
CurBench: Curriculum Learning Benchmark
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces CurBench, a benchmark and toolkit that evaluates 15 curriculum learning methods across 15 datasets, 9 backbone models, and 3 settings (standard, noise, imbalance) on performance and complexity.
Problem
Curriculum learning methods are hard to compare fairly because prior works use different datasets, backbones, and settings, and no standardized benchmark exists.
Approach
A unified benchmark and toolkit (built on CurML) that plugs automatic curricula into standard training pipelines across CV, NLP, and graph datasets under standard, label-noise, and class-imbalance settings, with evaluation on both performance and time/GPU-memory complexity.
Key results
- Curriculum learning gives consistent gains across domains, with the largest benefit in noise settings (e.g., BERT on RTE noise rises from 54.22 to 56.32 with CL; ResNet-18 on CIFAR-10 noise rises from 82.75 to 86.92).
- No single curriculum method dominates; TTCL performs competitively across noise ratios while SPL performs relatively poorly in noise settings.
- Methods involving gradient calculation and meta-networks (e.g., MW-Net, LRE, DDS) incur the highest time and GPU memory costs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply the best curriculum learning method | LeNet · CIFAR-10 Standard | accuracy: 69.95±1.00 -> 70.43±0.41 | Helped |
| Apply the best curriculum learning method | LeNet · CIFAR-10 Noise-0.4 | accuracy: 65.02±1.12 -> 65.93±0.57 | Helped |
| Apply the best curriculum learning method | LeNet · CIFAR-10 Imbalance-50 | accuracy: 44.93±0.56 -> 45.28±0.56 | Helped |
| Apply the best curriculum learning method | ResNet-18 · CIFAR-10 Standard | accuracy: 92.33±0.16 -> 92.88±0.23 | Helped |
| Apply the best curriculum learning method | ResNet-18 · CIFAR-10 Noise-0.4 | accuracy: 82.75±2.06 -> 86.92±0.20 | Helped |
| Apply the best curriculum learning method | ResNet-18 · CIFAR-10 Imbalance-50 | accuracy: 75.49±0.87 -> 76.43±0.96 | Helped |
| Apply the best curriculum learning method | ViT · CIFAR-10 Standard | accuracy: 79.90±0.38 -> 80.66±0.27 | Helped |
| Apply the best curriculum learning method | ViT · CIFAR-10 Noise-0.4 | accuracy: 64.19±0.51 -> 69.83±0.53 | Helped |
| Apply the best curriculum learning method | ViT · CIFAR-10 Imbalance-50 | accuracy: 52.12±0.81 -> 52.85±0.81 | Helped |
| Use teacher transferring (TTCL) instead of self-guided SPL | Curriculum learning on noisy datasets · Noise settings | accuracy | Helped |
| Use data reweighting methods such as DCL and SuperLoss | Curriculum learning on imbalanced CIFAR-10 · CIFAR-10 Imbalance-r | accuracy | Helped |
Surprising findings
- Vanilla SPL, which is similar to TTCL but uses the model itself as teacher, performs relatively poorly under label noise.
- Curriculum learning brings less improvement in imbalance settings than in noise settings.
Limitations
- Covers only supervised learning on CV, NLP, and graph; excludes RL, unsupervised, semi-supervised, multimodal, and recommender settings.
- No new datasets introduced and no evaluation on large models.
- Uses label noise and imbalance synthetically injected into existing datasets.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Tiny-ImageNet RTE MRPC STS-B CoLA SST-2 QNLI QQP MNLI MUTAG
Benchmarks: CIFAR-10 CIFAR-100 Tiny-ImageNet GLUE (RTE, MRPC, STS-B, CoLA, SST-2, QNLI, QQP, MNLI) MUTAG PROTEINS NCI1 ogbg-molhiv
Metrics: accuracy F1 score Spearman correlation Matthews correlation ROC-AUC training time (h) maximum GPU memory (GB)
Baselines: SPL TTCL MCL ScreenerNet LRE MW-Net DCL LGL DDS DIHCL SuperLoss CBS
Builds on: CurML self-paced learning MW-Net DDS
Closest papers by meaning
- K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs (NeurIPS 2026 · similarity 0.86)
- Learning to Solve Complex Problems via Dataset Decomposition (NeurIPS 2025 · similarity 0.85)
- Dual-Level Curriculum Meta-Learning for Noisy Few-Shot Learning Tasks (AAAI 2024 · similarity 0.84)
- Benchmarking Generative Models on Computational Thinking Tests in Elementary Visual Programming (NeurIPS 2024 · similarity 0.84)
- Montessori-Instruct: Generate Influential Training Data Tailored for Student Learning (ICLR 2025 · similarity 0.83)
- Measuring Vision-Language STEM Skills of Neural Models (ICLR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.