Research map / Large language models / Agents, code and math
LLM-AutoDA: Large Language Model-Driven Automatic Data Augmentation for Long-tailed Problems
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.
Proposes LLM-AutoDA, which uses an LLM (gpt-3.5-turbo) with evolutionary search (AEL) to generate class-wise data augmentation selection algorithms for long-tailed learning, using validation accuracy as fitness.
Problem
Long-tailed learning suffers from class imbalance; existing data augmentation methods rely on hand-designed, fixed strategies with limited search spaces and cannot adapt during training.
Approach
Use an LLM as a search operator in an evolutionary loop (initialization, crossover, mutation prompts) to generate per-class augmentation selection and intensity algorithms, with validation accuracy after short fine-tuning as the fitness reward.
Key results
- On CIFAR-100-LT (IR=100) with CE, LLM-AutoDA reaches 45.0% vs 42.0% for CUDA and 44.5% for DODA (CE baseline 40.1%).
- On ImageNet-LT with CE, LLM-AutoDA improves overall accuracy from 41.6% to 50.4% (+8.8).
- On iNaturalist 2018 with BCL, LLM-AutoDA achieves 74.2% vs 71.8% baseline (+2.4) and with RIDE 73.9% (+3.0 over 71.6%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace SimpleLLM augmentation with LLM-AutoDA | BCL · CIFAR-100-LT IR=50 | Top-1 accuracy: 54.0 -> 54.8 | Helped |
| Replace DODA augmentation with LLM-AutoDA | BCL · CIFAR-100-LT IR=100 | Top-1 accuracy: 51.0 -> 51.0 | No clear effect |
| Replace DODA augmentation with LLM-AutoDA | BCL · ImageNet-LT | Top-1 accuracy: 56.9 -> 57.5 | Helped |
| Replace DODA augmentation with LLM-AutoDA | BCL · iNaturalist 2018 | Top-1 accuracy: 73.7 -> 74.2 | Helped |
Surprising findings
- SimpleLLM, a non-optimized prompt-only LLM strategy generator, achieves performance comparable to hand-designed CUDA and DODA.
Limitations
- Scoring mechanism for augmentation strategies is imperfect and needs a more comprehensive evaluation.
- Search space limits the generation of truly novel augmentation methods.
- Potential dependence on LLMs.
Experimental setup
Datasets: CIFAR-100-LT ImageNet-LT iNaturalist 2018
Benchmarks: CIFAR-100-LT ImageNet-LT iNaturalist 2018
Metrics: Top-1 accuracy head/medium/tail accuracy
Baselines: CE CE-DRW LDAM-DRW BS RIDE BCL CUDA DODA SimpleLLM
Builds on: AEL AutoAugment CUDA DODA Balanced Softmax
Closest papers by meaning
- Curated LLM: Synergy of LLMs and Data Curation for tabular augmentation in low-data regimes (ICML 2024 · similarity 0.85)
- Breaking Long-Tailed Learning Bottlenecks: A Controllable Paradigm with Hypernetwork-Generated Diverse Experts (NeurIPS 2024 · similarity 0.85)
- Learning from Neighbors: Category Extrapolation for Long-Tail Learning (CVPR 2025 · similarity 0.85)
- LTGC: Long-tail Recognition via Leveraging LLMs-driven Generated Content (CVPR 2024 · similarity 0.85)
- AlphaDecay: Module-wise Weight Decay for Heavy-Tailed Balancing in LLMs (NeurIPS 2025 · similarity 0.85)
- EE-LLM: Large-Scale Training and Inference of Early-Exit Large Language Models with 3D Parallelism (ICML 2024 · similarity 0.84)
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.