Research map / Probabilistic and scientific modeling / Bayesian optimization
Offline black-box optimization with surrogates
Optimize designs from static datasets using learned surrogates, generative models, and language-model embeddings.
56 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Bayesian optimization, keyword cluster “black box · box optimization · offline”
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.
10 papers in 2023–24 and 30 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 10.9% to 22.4%.
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
- 2017 · Learning to Prune Dominated Action Sequences in Online Black-Box Planning (AAAI 2017 · 12 citations)
- 2018 · Semi-Black Box: Rapid Development of Planning Based Solutions (AAAI 2018 · 1 citation)
- 2019 · Demystifying Black-box Models with Symbolic Metamodels (NeurIPS 2019 · 51 citations)
- 2020 · An ADMM Based Framework for AutoML Pipeline Configuration (AAAI 2020 · 54 citations)
- 2021 · Improving black-box optimization in VAE latent space using decoder uncertainty (NeurIPS 2021 · 24 citations)
- 2022 · Design-Bench: Benchmarks for Data-Driven Offline Model-Based Optimization (ICML 2022 · 7 citations)
- 2023 · Discovering Evolution Strategies via Meta-Black-Box Optimization (ICLR 2023 · 28 citations)
- 2024 · SYMBOL: Generating Flexible Black-Box Optimizers through Symbolic Equation Learning (ICLR 2024 · 1 citation)
- 2025 · ROOT: Rethinking Offline Optimization as Distributional Translation via Probabilistic Bridge (NeurIPS 2025 · 1 citation)
- 2026 · VALIANT: Prompt Instability for Active Learning in Black-Box Medical Imaging (AAAI 2026 · 1 citation)
Earliest papers here
- Learning to Prune Dominated Action Sequences in Online Black-Box Planning (AAAI 2017 · 12 citations)
- A Nearly-Black-Box Online Algorithm for Joint Parameter and State Estimation in Temporal Models (AAAI 2017 · 2 citations)
- Semi-Black Box: Rapid Development of Planning Based Solutions (AAAI 2018 · 1 citation)
What has been tried
37 papers here report 339 measured changes (a component added, removed or swapped, compared with a baseline): helped 65%, no clear effect 7%, hurt 19%, mixed 9%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| use probabilistic-bridge learning instead of gradient ascent on the GP posterior mean | Design-Bench tasks · score | Helped | ROOT: Rethinking Offline Optimization as Distributional Translation via Probabilistic Bridge (NeurIPS 2025 · 1 citation) | |
| remove population features from FLA | CEC-competition · mean normalized performance | 0.926 vs 0.972 | Hurt | SYMBOL: Generating Flexible Black-Box Optimizers through Symbolic Equation Learning (ICLR 2024 · 1 citation) |
| Increase the observed dataset size | Synthetic experiments and police redistricting (6 × 6 grid and Atlanta) · optimal objective value | Helped | Black-Box Optimization with Implicit Constraints for Public Policy (AAAI 2025 · 0 citations) | |
| Use LSTM rather than GRU autoregressive cells | Symbolic regression · mean reward | 0.7377; 0.7092; 0.7236; 0.7385; 0.7391; 0.7302 | Mixed | DisCo-DSO: Coupling Discrete and Continuous Optimization for Efficient Generative Design in Hybrid Spaces (AAAI 2025 · 0 citations) |
| Use grammar-based encoding length L=5 rather than L=15 | Staircase · regret | Hurt | Kernel Learning for Sample Constrained Black-Box Optimization (AAAI 2025 · 0 citations) | |
| add metadata-guided contrastive alignment and local Lipschitz smoothness loss | Design-Bench + SOO-Bench multi-task · average rank (lower better) | 2.667 -> 2.222 | Helped | Towards Universal Offline Black-Box Optimization via Learning Language Model Embeddings (ICML 2025 · 0 citations) |
| Train both agents rather than only Agent-1 or only Agent-2 | D_test and three realistic problem sets · reversed normalized objective | Helped | DesignX: Human-Competitive Algorithm Designer for Black-Box Optimization (NeurIPS 2025 · 0 citations) | |
| Use off-policy importance-weighted gradients instead of on-policy gradients | Ehrlich vs. Naturalness (M=32) · hypervolume performance | No clear effect | Amortized Active Generation of Pareto Sets (NeurIPS 2025 · 0 citations) | |
| Train NeurELA with Fast CMA-ES rather than another tested evolution strategy | Unseen algorithms and unseen problem sets · relative performance | Helped | Neural Exploratory Landscape Analysis for Meta-Black-Box-Optimization (ICLR 2025 · 0 citations) | |
| Use data augmentation to construct ranking lists instead of treating a batch as one list | Design-Bench tasks · 100th percentile normalized score | Helped | Offline Model-Based Optimization by Learning to Rank (ICLR 2025 · 0 citations) | |
| Use direct Taylor/CDF sensitivity proxy instead of the main BOSS formulation | Six Design-Bench tasks · normalized performance | 13/18 instances improved; maximum and average improvement 3.1% and 1.25% | Mixed | Boosting Offline Optimizers with Surrogate Sensitivity (ICML 2024 · 0 citations) |
| add regression (value-matching) regularizer along synthetic monotonic trajectories to gradient matching loss | Design-bench (6 tasks), 100th percentile · score | better on 4/6 tasks (Ant, DKitty, Tf8 up; Hopper up; SCon and Tf10 down) | Mixed | Learning Surrogates for Offline Black-Box Optimization via Gradient Matching (ICML 2024 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 18 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Protein-Docking · 2 of 18 papers
Baselines
- CMA-ES · 8 of 18 papers
- DDOM · 4 of 18 papers
- BO-qEI · 3 of 18 papers
- REINFORCE · 3 of 18 papers
- CbAS · 3 of 18 papers
- GLHF · 2 of 18 papers
- GLEET · 2 of 18 papers
- COMs · 2 of 18 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 40 papers; from OpenAlex and the papers' affiliations).
- Carnegie Mellon University · 2 papers
- Nanyang Technological University · 2 papers
- South China University of Technology · 2 papers
- University of California, Los Angeles · 2 papers
- National University of Singapore · 2 papers
- Hanoi University of Science and Technology · 2 papers
- Washington State University · 2 papers
- AIST · 2 papers
- Massachusetts Institute of Technology · 2 papers
Use it
Download BibTeX of 40 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
- High-dimensional Bayesian optimization (34 papers)
Elsewhere in the topic
- Acquisition function design (97 papers)
- Local and high-dimensional Bayesian optimization (90 papers)
- Multi-fidelity hyperparameter optimization (44 papers)
- Regret bounds for Bayesian optimization (43 papers)
- Multi-objective Bayesian optimization (33 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).