Research map / Architectures and efficiency / Neural architecture search
Differentiable architecture search (DARTS)
Relaxes architecture choice to continuous weights, improving DARTS robustness, memory use, and analysis.
32 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Neural architecture search, keyword cluster “differentiable · image · latency”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Few recent reports None of the tracked benchmarks has a report from 2025–26.
9 papers in 2023–24 and 6 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 12.0% to 12.0%.
Open questions the data points to
- Papers report DARTS- on NAS-Bench-201 CIFAR-100 (test accuracy) anywhere from 15.61 to 71.53: comparisons against it depend on whose number you take.
- Papers report DARTS- on NAS-Bench-201 ImageNet16-120 (test accuracy) anywhere from 16.32 to 45.12: comparisons against it depend on whose number you take.
No longer reported by papers here: CIFAR-10 test (test error (%), last in 2020).
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 · Improved Hamming Distance Search Using Variable Length Substrings (CVPR 2016 · 30 citations)
- 2019 · DARTS: Differentiable Architecture Search (ICLR 2019 · 1,354 citations)
- 2020 · PC-DARTS: Partial Channel Connections for Memory-Efficient Architecture Search (ICLR 2020 · 420 citations)
- 2021 · DARTS-: Robustly Stepping out of Performance Collapse Without Indicators (ICLR 2021 · 89 citations)
- 2022 · Interpreting Operation Selection in Differentiable Architecture Search: A Perspective from Influence-Directed Explanations (NeurIPS 2022 · 1 citation)
- 2023 · Differentiable Architecture Search With Random Features (CVPR 2023 · 23 citations)
- 2024 · Training-Free Quantum Architecture Search (AAAI 2024 · 26 citations)
- 2025 · Resource Constrained Pathfinding with Enhanced Bidirectional A* Search (AAAI 2025 · 2 citations)
- 2026 · TAS-LoRA: Transformer Architecture Search with Mixture-of-LoRA Experts (CVPR 2026 · 0 citations)
Earliest papers here
- Improved Hamming Distance Search Using Variable Length Substrings (CVPR 2016 · 30 citations)
- DARTS: Differentiable Architecture Search (ICLR 2019 · 1,354 citations)
- DATA: Differentiable ArchiTecture Approximation (NeurIPS 2019 · 42 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.
CIFAR-10 test · test error (%) (lower is better)
4 papers, 2019–2020 · not reported since 2020
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 2.59 | DATA: Differentiable ArchiTecture Approximation (NeurIPS 2019 · 42 citations) |
| 2020 | 2.32 | Theory-Inspired Path-Regularized Differential Network Architecture Search (NeurIPS 2020 · 24 citations) |
What has been tried
26 papers here report 269 measured changes (a component added, removed or swapped, compared with a baseline): helped 74%, no clear effect 7%, hurt 15%, 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 |
|---|---|---|---|---|
| Use only the path-based proxy rather than both proxies | TFIM · queries to ground-truth performance | 33.5 -> 104.7 | Hurt | Training-Free Quantum Architecture Search (AAAI 2024 · 26 citations) |
| Use NSGA-II genetic search instead of random search | CIFAR-100; ResNet110 teacher · search performance | Helped | UniADS: Universal Architecture-Distiller Search for Distillation Gap (AAAI 2024 · 16 citations) | |
| Add illegal actions to CRLQAS | LiH-6 noiseless · noiseless error | 0.0037 -> 0.0008 | Helped | Curriculum reinforcement learning for quantum architecture search under hardware errors (ICLR 2024 · 7 citations) |
| Disable state upper-bound pruning during initialization | k=4, especially tightness 10% and 30% · runtime slowdown | average slowdown exceeding a factor of two | Hurt | Resource Constrained Pathfinding with Enhanced Bidirectional A* Search (AAAI 2025 · 2 citations) |
| add regularized skip-connection | LogicNets2 · accuracy | 79.40 -> 99.91 | Helped | Towards Next-Generation Logic Synthesis: A Scalable Neural Circuit Generation Framework (NeurIPS 2024 · 1 citation) |
| use fixed rather than trainable TN initialization | 8-qubit H2O depolarizing noise · error | 2.48\times10^{-3} -> 9.0\times10^{-4} | Helped | TensorRL-QAS: Reinforcement learning with tensor networks for improved quantum architecture search (NeurIPS 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 |
|---|---|---|---|---|
| DARTS- | NAS-Bench-201 CIFAR-100 | test accuracy | 15.61 – 71.53 | 2 papers, e.g. NeurIPS 2022 and CVPR 2023 |
| DARTS- | NAS-Bench-201 ImageNet16-120 | test accuracy | 16.32 – 45.12 | 2 papers, e.g. NeurIPS 2022 and CVPR 2023 |
| DARTS- | NAS-Bench-201 CIFAR-10 | test accuracy | 54.3 – 93.8 | 2 papers, e.g. NeurIPS 2022 and CVPR 2023 |
What reviewers will expect
Datasets and baselines named in the 9 papers of this approach from 2024–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Baselines
- random sampling · 2 of 9 papers
- quantumDARTS · 2 of 9 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 15 papers; from OpenAlex and the papers' affiliations).
- University of Science and Technology of China · 3 papers
- Peking University · 2 papers
- Tianjin University · 2 papers
- Huawei · 2 papers
Use it
Download BibTeX of 27 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
- Latency-aware architecture search (91 papers)
- AutoML pipeline search (57 papers)
Elsewhere in the topic
- Efficient NAS search strategies (86 papers)
- Bi-level differentiable NAS (84 papers)
- One-shot and few-shot supernets (30 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).