Research map / Optimization / Optimizers
Gradient-based meta-learning theory
Provides guarantees and improved algorithms for MAML-style outer-loop meta-learning.
40 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Optimizers, keyword cluster “bilevel · inner · shot”
State History Benchmarks Tried Baselines Preprints Who Export
Where it stands
Few recent reports None of the tracked benchmarks has a report from 2025–26.
8 papers in 2023–24 and 9 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 3.3% to 3.1%.
No longer reported by papers here: miniImageNet 5-way 1-shot (accuracy, last in 2023); miniImageNet 5-way 5-shot (accuracy, last in 2023); tieredImageNet 5-way 1-shot (accuracy (%), last in 2023); tieredImageNet 5-way 5-shot (accuracy (%), last in 2023).
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2016 · Learning to learn by gradient descent by gradient descent (NeurIPS 2016 · 342 citations)
- 2017 · Optimization as a Model for Few-Shot Learning (ICLR 2017 · 2,375 citations)
- 2018 · Meta-Learning and Universality: Deep Representations and Gradient Descent can Approximate any Learning Algorithm (ICLR 2018 · 90 citations)
- 2019 · Efficient Meta Learning via Minibatch Proximal Update (NeurIPS 2019 · 48 citations)
- 2020 · Meta-Learning with Warped Gradient Descent (ICLR 2020 · 116 citations)
- 2021 · Multi-Objective Meta Learning (NeurIPS 2021 · 8 citations)
- 2022 · Sharp-MAML: Sharpness-Aware Model-Agnostic Meta Learning (ICML 2022 · 9 citations)
- 2023 · Meta-Learning With a Geometry-Adaptive Preconditioner (CVPR 2023 · 26 citations)
- 2024 · MetaDiff: Meta-Learning with Conditional Diffusion for Few-Shot Learning (AAAI 2024 · 68 citations)
- 2025 · Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · 7 citations)
- 2026 · The Role of Target Update Frequencies in Q-Learning (ICML 2026 · 0 citations)
Earliest papers here
- Learning to learn by gradient descent by gradient descent (NeurIPS 2016 · 342 citations)
- Optimization as a Model for Few-Shot Learning (ICLR 2017 · 2,375 citations)
- Meta-Learning and Universality: Deep Representations and Gradient Descent can Approximate any Learning Algorithm (ICLR 2018 · 90 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
miniImageNet 5-way 1-shot · accuracy (higher is better)
7 papers, 2018–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 51.7 | Gradient-Based Meta-Learning with Learned Layerwise Metric and Subspace (ICML 2018 · 72 citations) |
| 2020 | 52.3 | Meta-Learning with Warped Gradient Descent (ICLR 2020 · 116 citations) |
| 2022 | 59.26 | Continuous-Time Meta-Learning with Forward Mode Differentiation (ICLR 2022 · 6 citations) |
| 2023 | 59.89 | Meta-AdaM: An Meta-Learned Adaptive Optimizer with Momentum for Few-Shot Learning (NeurIPS 2023 · 5 citations) |
miniImageNet 5-way 5-shot · accuracy (higher is better)
5 papers, 2020–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 68.4 | Meta-Learning with Warped Gradient Descent (ICLR 2020 · 116 citations) |
| 2022 | 77.26 | Continuous-Time Meta-Learning with Forward Mode Differentiation (ICLR 2022 · 6 citations) |
| 2023 | 77.92 | Meta-AdaM: An Meta-Learned Adaptive Optimizer with Momentum for Few-Shot Learning (NeurIPS 2023 · 5 citations) |
tieredImageNet 5-way 1-shot · accuracy (%) (higher is better)
4 papers, 2020–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 57.2 | Meta-Learning with Warped Gradient Descent (ICLR 2020 · 116 citations) |
| 2022 | 62.93 | Continuous-Time Meta-Learning with Forward Mode Differentiation (ICLR 2022 · 6 citations) |
| 2023 | 65.31 | Meta-AdaM: An Meta-Learned Adaptive Optimizer with Momentum for Few-Shot Learning (NeurIPS 2023 · 5 citations) |
tieredImageNet 5-way 5-shot · accuracy (%) (higher is better)
4 papers, 2020–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 74.1 | Meta-Learning with Warped Gradient Descent (ICLR 2020 · 116 citations) |
| 2022 | 81.13 | Continuous-Time Meta-Learning with Forward Mode Differentiation (ICLR 2022 · 6 citations) |
| 2023 | 85.24 | Meta-AdaM: An Meta-Learned Adaptive Optimizer with Momentum for Few-Shot Learning (NeurIPS 2023 · 5 citations) |
What has been tried
32 papers here report 329 measured changes (a component added, removed or swapped, compared with a baseline): helped 65%, no clear effect 11%, hurt 20%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Replace standard GDA with momentum GDA | 5-way 1-shot · accuracy | 60.53 ± 0.86% -> 62.03 ± 0.82% | Helped | MetaDiff: Meta-Learning with Conditional Diffusion for Few-Shot Learning (AAAI 2024 · 68 citations) |
| Add lower-level optimization for task-specific parameters | QM9 · gradient similarity | Helped | Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · 7 citations) | |
| Use the MᵀM+I DSP design instead of LLᵀ | Meta-Dataset multi-domain Varying-Way Varying-Shot · Avg All accuracy | 80.1 -> 80.9 | Helped | Task-Specific Preconditioner for Cross-Domain Few-Shot Learning (AAAI 2025 · 3 citations) |
| Average state vectors from separate tokens (inner optimization) | Six-task average, zero-shot; Llama-2-7B · first-token accuracy | 53.7 -> 63.9; average improvement of 10.2% | Helped | In-Context Learning State Vector with Inner and Momentum Optimization (NeurIPS 2024 · 1 citation) |
| use blockwise step sizes instead of a scalar step size | CIFAR10 · performance | Mixed | MetaOptimize: A Framework for Optimizing Step Sizes and Other Meta-parameters (ICML 2025 · 0 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| Reptile | Mini-ImageNet 1-shot 5-way | accuracy (%) | 47.07 – 50 | 2 papers, e.g. NeurIPS 2019 and ICML 2022 |
| Reptile | Mini-ImageNet 5-shot 5-way | accuracy (%) | 62.74 – 66 | 2 papers, e.g. NeurIPS 2019 and ICML 2022 |
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 17 papers; from OpenAlex and the papers' affiliations).
- Seoul National University · 3 papers
- Harbin Institute of Technology · 2 papers
- Shanghai Jiao Tong University · 2 papers
Use it
Download BibTeX of 30 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Hypergradient bilevel optimization (35 papers)
- Differentiating through inner loops (23 papers)
Elsewhere in the topic
- Stochastic gradient MCMC (142 papers)
- Surrogate losses and robust solvers (119 papers)
- Differentiating through embedded optimization (117 papers)
- Constrained and non-backprop training (97 papers)
- Memory-efficient LLM optimizers (79 papers)
- Adam convergence and adaptive learning rates (78 papers)
- Natural gradient and Fisher approximations (77 papers)
- Randomized gradient estimation (77 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).