Research map / Optimization / Nonconvex and smooth optimization
Penalty and implicit-gradient bilevel methods
Develops first-order and penalty-based algorithms for bilevel optimization with convergence guarantees.
65 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2021–2026) · an approach in Nonconvex and smooth optimization, keyword cluster “complexity · bilevel · stationary”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
28 papers in 2023–24 and 28 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 8.8% to 8.0%.
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
- 2021 · A Near-Optimal Algorithm for Stochastic Bilevel Optimization via Double-Momentum (NeurIPS 2021 · 26 citations)
- 2022 · FedNest: Federated Bilevel, Minimax, and Compositional Optimization (ICML 2022 · 7 citations)
- 2023 · On Penalty-based Bilevel Gradient Descent Method (ICML 2023 · 3 citations)
- 2024 · First-Order Methods for Linearly Constrained Bilevel Optimization (NeurIPS 2024 · 2 citations)
- 2025 · Tuning-Free Bilevel Optimization: New Algorithms and Convergence Analysis (ICLR 2025 · 1 citation)
- 2026 · A Tale of Two Problems: Multi-Task Bilevel Learning Meets Equality Constrained Multi-Objective Optimization (ICML 2026 · 0 citations)
Earliest papers here
- A Near-Optimal Algorithm for Stochastic Bilevel Optimization via Double-Momentum (NeurIPS 2021 · 26 citations)
- Provably Faster Algorithms for Bilevel Optimization (NeurIPS 2021 · 9 citations)
- Closing the Gap: Tighter Analysis of Alternating Stochastic Gradient Methods for Bilevel Problems (NeurIPS 2021 · 9 citations)
What has been tried
47 papers here report 334 measured changes (a component added, removed or swapped, compared with a baseline): helped 77%, no clear effect 10%, hurt 10%, 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 |
|---|---|---|---|---|
| Add adaptive penalty updates and warm-start | LRP · convergence speed | Helped | Penalty-based Methods for Simple Bilevel Optimization under Hölderian Error Bounds (NeurIPS 2024 · 1 citation) | |
| increase γ from 2 to 3 | Transportation, NN=3 · UL utility | 1.07 -> 1.69 | Helped | A Primal-Dual-Assisted Penalty Approach to Bilevel Optimization with Coupled Constraints (NeurIPS 2024 · 1 citation) |
| replace exponential mappings with retraction mappings | Riemannian bilevel optimization · iteration complexity | \mathcal{O}\left(\frac{1}{T}\right) -> \mathcal{O}\left(\frac{1}{T}\right) | No clear effect | An Adaptive Algorithm for Bilevel Optimization on Riemannian Manifolds (NeurIPS 2025 · 0 citations) |
| Vary δ from 0.01 to 0.05, 0.1, or 0.5 | 3-node transportation network · final performance | 0.01, 0.05, 0.1, 0.5 | Mixed | A Single-Loop First-Order Algorithm for Linearly Constrained Bilevel Optimization (NeurIPS 2025 · 0 citations) |
| set alpha0 to 1 rather than 0.1 | Synthetic PBO, n=100 · Time (s) | 1.0±0.1 -> 0.1±0.0 | Helped | A Single-Loop Gradient Algorithm for Pessimistic Bilevel Optimization via Smooth Approximation (NeurIPS 2025 · 0 citations) |
| Increase penalty constant γ | Representation learning and data hyper-cleaning · optimality gap | Helped | Unlocking Global Optimality in Bilevel Optimization: A Pilot Study (ICLR 2025 · 0 citations) | |
| Increase Q_k to k+1 | Toy example · hypergradient error | Helped | qNBO: quasi-Newton Meets Bilevel Optimization (ICLR 2025 · 0 citations) | |
| Add gradient tracking | MNIST test accuracy, non-i.i.d. · test accuracy | Helped | DUET: Decentralized Bilevel Optimization without Lower-Level Strong Convexity (ICLR 2025 · 0 citations) | |
| Increase LancBiO subspace dimension m | synthetic bilevel problem · residual norm | Helped | LancBiO: Dynamic Lanczos-aided Bilevel Optimization via Krylov Subspace (ICLR 2025 · 0 citations) | |
| Increase q from 0 in MA-SOBA-q | MNIST data hyper-cleaning · convergence speed | Helped | SPABA: A Single-Loop and Probabilistic Stochastic Bilevel Algorithm Achieving Optimal Sample Complexity (ICML 2024 · 0 citations) | |
| set Q=1 instead of Q=3 | Data hyper-cleaning benchmarks · test accuracy (%) | CodRNA: 80.42 ± 0.05 -> 80.25 ± 0.07; MNIST 6vs9: 94.24± 0.01 -> 93.65 ± 0.01; Madelon: 56.33 ± 1.83 -> 55.23 ± 1.22; Fa | Hurt | Double Momentum Method for Lower-Level Constrained Bilevel Optimization (ICML 2024 · 0 citations) |
| Use independent sampling instead of without-replacement sampling | Hyper-representation learning · classification performance | No clear effect | Provably Faster Algorithms for Bilevel Optimization via Without-Replacement Sampling (NeurIPS 2024 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 21 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- MNIST · 10 of 21 papers
- FashionMNIST · 4 of 21 papers
- CIFAR-10 · 3 of 21 papers
Baselines
- BOME · 4 of 21 papers
- F2SA · 4 of 21 papers
- SOBA · 3 of 21 papers
- PBGD · 2 of 21 papers
- MEHA · 2 of 21 papers
- BLOOP · 2 of 21 papers
- SLDBO · 2 of 21 papers
- PZOBO · 2 of 21 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 56 papers; from OpenAlex and the papers' affiliations).
- University of Wisconsin - Madison · 5 papers
- Southern University of Science and Technology · 4 papers
- University of Texas at Austin · 4 papers
- University of Arizona · 4 papers
- Massachusetts Institute of Technology · 3 papers
- IBM · 3 papers
- George Mason University · 3 papers
- Rensselaer Polytechnic Institute · 3 papers
- The Chinese University of Hong Kong · 3 papers
- University of Texas, Austin · 3 papers
Use it
Download BibTeX of 34 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
- Nonconvex minimax and constrained methods (100 papers)
- Stochastic convex optimization complexity (94 papers)
- Complexity of nonconvex first-order methods (79 papers)
- Saddle-point escape algorithms (23 papers)
Elsewhere in the topic
- Adaptive first-order convex methods (190 papers)
- Coordinate and Adam convergence analysis (188 papers)
- Stochastic gradient last-iterate analysis (178 papers)
- Acceleration and Nesterov dynamics (91 papers)
- Equilibrium learning in zero-sum games (88 papers)
- Stochastic variance reduction (86 papers)
- Langevin sampling for nonconvex problems (70 papers)
- Proximal finite-sum methods (47 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).