Research map / Architectures and efficiency / RNNs and normalization
Scalable approximation and equivariant compute
Scales Laplace approximations, equivariant attention, and tensor-program performance models.
120 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in RNNs and normalization, keyword cluster “memory · scaling · cost”
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.
38 papers in 2023–24 and 57 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 5.8% to 6.1%.
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 · TenSR: Multi-Dimensional Tensor Sparse Representation (CVPR 2016 · 44 citations)
- 2018 · Adafactor: Adaptive Learning Rates with Sublinear Memory Cost (ICML 2018 · 162 citations)
- 2019 · ANODEV2: A Coupled Neural ODE Framework (NeurIPS 2019 · 28 citations)
- 2020 · AutoPrivacy: Automated Layer-wise Parameter Selection for Secure Neural Network Inference (NeurIPS 2020 · 18 citations)
- 2021 · Efficient On-Chip Learning for Optical Neural Networks Through Power-Aware Sparse Zeroth-Order Optimization (AAAI 2021 · 19 citations)
- 2022 · MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields (NeurIPS 2022 · 295 citations)
- 2023 · PyPose: A Library for Robot Learning With Physics-Based Optimization (CVPR 2023 · 35 citations)
- 2024 · EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations (ICLR 2024 · 84 citations)
- 2025 · UMA: A Family of Universal Models for Atoms (NeurIPS 2025 · 72 citations)
- 2026 · DistMLIP: A Distributed Inference Platform for Machine Learning Interatomic Potentials (ICLR 2026 · 4 citations)
Earliest papers here
- TenSR: Multi-Dimensional Tensor Sparse Representation (CVPR 2016 · 44 citations)
- Adafactor: Adaptive Learning Rates with Sublinear Memory Cost (ICML 2018 · 162 citations)
- A Scalable Laplace Approximation for Neural Networks (ICLR 2018 · 157 citations)
What has been tried
74 papers here report 826 measured changes (a component added, removed or swapped, compared with a baseline): helped 68%, no clear effect 9%, hurt 19%, 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 attention re-normalization | OC20 S2EF-2M validation · Energy MAE | 286 -> 279 | Helped | EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations (ICLR 2024 · 84 citations) |
| Use MoLE rather than a dense model for multi-task UMA-S | Multi-task training · validation loss | Helped | UMA: A Family of Universal Models for Atoms (NeurIPS 2025 · 72 citations) | |
| Remove the basis function | f(x,y)=x exp(-y), varying network depth · test RMSE | Hurt | PowerMLP: An Efficient Version of KAN (AAAI 2025 · 11 citations) | |
| Use 64 inducing points rather than the 512-point variant | ERA5 · relative L2 error | 6.44e--3 -> 1.45e--2 | Hurt | Inducing Point Operator Transformer: A Flexible and Scalable Architecture for Solving PDEs (AAAI 2024 · 10 citations) |
| Replace teacher Hessians with teacher forces for distillation | SPICE Monomers · Force MAE (meV/A) | 6.3 -> 14.2 | Hurt | Towards Fast, Specialized Machine Learning Force Fields: Distilling Foundation Models via Energy Hessians (ICLR 2025 · 9 citations) |
| Add normalized NRMSE training loss | SWE · NRMSE | 0.01353 -> 0.00240 | Helped | Multiple Physics Pretraining for Spatiotemporal Surrogate Models (NeurIPS 2024 · 8 citations) |
| Increase attention parameters (More AT) rather than scaling spherical channels | OC20 S2EF 2M · Energy and force MAE | Helped | The Importance of Being Scalable: Improving the Speed and Accuracy of Neural Network Interatomic Potentials Across Chemical Domains (NeurIPS 2024 · 7 citations) | |
| Replace hybrid loss with BCE | Erdős-Rényi, n=200, 75% edge reduction · optimality | 91.07 ± 0.82 -> 86.90 ± 1.08 | Hurt | Scalable Quantum-Inspired Optimization Through Dynamic Qubit Compression (AAAI 2025 · 6 citations) |
| Add shared attention cache | LAMBADA · speculative decoding speedup | 1.14\times -> 1.16\times | Helped | MatFormer: Nested Transformer for Elastic Inference (NeurIPS 2024 · 6 citations) |
| Use full finetuning rather than unit identification | NHP reaching, T – CO (6) · R² | 0.8763 ± 0.0259 -> 0.9068 ± 0.0170 | Helped | Generalizable, real-time neural decoding with hybrid state-space models (NeurIPS 2025 · 2 citations) |
| implicit (DEQ) differentiation instead of explicit unrolled layers | MNIST-4 angle encoding · test accuracy | IMPLICIT 85.8 vs DIRECT 10 layers 84.8 | Helped | Quantum Deep Equilibrium Models (NeurIPS 2024 · 2 citations) |
| Replace learned, targeted freezing with random freezing | H₂ · ΔE (Random − baseline) | 2.248 | Hurt | TITAN: A Trajectory-Informed Technique for Adaptive Parameter Freezing in Large-Scale VQE (NeurIPS 2025 · 1 citation) |
What reviewers will expect
Datasets and baselines named in the 33 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 · 5 of 33 papers
- CIFAR-10 · 4 of 33 papers
- OC20 · 3 of 33 papers
- OC22 · 3 of 33 papers
- CIFAR100 · 2 of 33 papers
- ImageNet · 2 of 33 papers
- MRPC · 2 of 33 papers
- QH9 · 2 of 33 papers
Baselines
- MLP · 2 of 33 papers
- QHNet · 2 of 33 papers
- GNO · 2 of 33 papers
- e3nn · 2 of 33 papers
- GemNet-dT · 2 of 33 papers
- EquiformerV2 · 2 of 33 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 95 papers; from OpenAlex and the papers' affiliations).
- Chinese Academy of Sciences · 5 papers
- University of California, Los Angeles · 5 papers
- University of Texas at Austin · 4 papers
- Meta · 4 papers
- Peking University · 3 papers
- Carnegie Mellon University · 3 papers
- Microsoft · 3 papers
- University of Washington · 3 papers
- University of Chinese Academy of Sciences · 2 papers
- Tsinghua University · 2 papers
Use it
Download BibTeX of 37 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
- Efficient network compression and training (229 papers)
- Dataset distillation and data-efficient training (158 papers)
- Neural scaling laws (155 papers)
- Distributed LLM inference systems (147 papers)
- Dataset condensation and flow estimation (126 papers)
Elsewhere in the topic
- Theory of depth and width (223 papers)
- Representation geometry and collapse (169 papers)
- Layer-wise analysis of language models (145 papers)
- Differentiable optimization layers (125 papers)
- Non-autoregressive translation (102 papers)
- RNN training dynamics and gradients (102 papers)
- Neural-symbolic reasoning (98 papers)
- Sparse dictionary representations (98 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).