Research map / Architectures and efficiency / RNNs and normalization
Efficient recurrent architectures
Makes RNNs efficient via activity sparsity, dropout, variable computation, and non-normal dynamics.
60 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in RNNs and normalization, keyword cluster “rnn · long · term”
State History Tried Baselines 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 6 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 1.5% to 0.6%.
Open questions the data points to
- Papers report LSTM on UCF11 (accuracy) anywhere from 0.492 to 0.697: comparisons against it depend on whose number you take.
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 · Unitary Evolution Recurrent Neural Networks (ICML 2016 · 221 citations)
- 2017 · On Human Motion Prediction Using Recurrent Neural Networks (CVPR 2017 · 1,024 citations)
- 2018 · Independently Recurrent Neural Network (IndRNN): Building a Longer and Deeper RNN (CVPR 2018 · 922 citations)
- 2019 · Predicting Hurricane Trajectories Using a Recurrent Neural Network (AAAI 2019 · 133 citations)
- 2020 · Where to Go Next: Modeling Long- and Short-Term User Preferences for Point-of-Interest Recommendation (AAAI 2020 · 406 citations)
- 2021 · Towards Extremely Compact RNNs for Video Recognition With Fully Decomposed Hierarchical Tucker Structure (CVPR 2021 · 32 citations)
- 2022 · RNNs of RNNs: Recursive Construction of Stable Assemblies of Recurrent Neural Networks (NeurIPS 2022 · 5 citations)
- 2023 · Resurrecting Recurrent Neural Networks for Long Sequences (ICML 2023 · 43 citations)
- 2024 · Structured flexibility in recurrent neural networks via neuromodulation (NeurIPS 2024 · 4 citations)
- 2025 · Enhancing Predictive Healthcare Using AI-Driven Early Warning Systems (AAAI 2025 · 1 citation)
- 2026 · Discovering alternative solutions beyond the simplicity bias in recurrent neural networks (ICLR 2026)
Earliest papers here
- Unitary Evolution Recurrent Neural Networks (ICML 2016 · 221 citations)
- Phased LSTM: Accelerating Recurrent Network Training for Long or Event-based Sequences (NeurIPS 2016 · 211 citations)
- Full-Capacity Unitary Recurrent Neural Networks (NeurIPS 2016 · 104 citations)
What has been tried
55 papers here report 576 measured changes (a component added, removed or swapped, compared with a baseline): helped 69%, no clear effect 9%, hurt 18%, 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 neuromodulatory component s1(t) | Measure-Wait-Go · task performance | Helped | Structured flexibility in recurrent neural networks via neuromodulation (NeurIPS 2024 · 4 citations) | |
| Replace LSTM recurrence with element-wise eLSTM recurrence | DMLab-30 rooms_watermaze · score | 37.6 ± 5.3 -> 45.6 ± 4.7 | Helped | Exploring the Promise and Limits of Real-Time Recurrent Learning (ICLR 2024 · 1 citation) |
| Increase the number of parallel probing sequences | Formal Languages and Sequential MNIST · validation loss | Helped | Learning Useful Representations of Recurrent Neural Network Weight Matrices (ICML 2024 · 1 citation) | |
| Add second-order GGN preconditioning to the subspace update | Lorenz attractor · training loss | Helped | Second-order forward-mode optimization of recurrent neural networks for neuroscience (NeurIPS 2024 · 1 citation) | |
| Use CUDA fused kernels instead of CUDA alternating kernels | SlimPajama training · Train Time | 10.8 -> 9.9 h | Helped | FlashRNN: I/O-Aware Optimization of Traditional RNNs on modern hardware (ICLR 2025 · 0 citations) |
| add the proposed initialization method instead of direct CordsNet training | MNIST · accuracy (%) | 99.40 -> 99.37 | Hurt | Recurrent neural network dynamical systems for biological vision (NeurIPS 2024 · 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 |
|---|---|---|---|---|
| LSTM | UCF11 | accuracy | 0.492 – 0.697 | 2 papers, e.g. ICML 2017 and CVPR 2018 |
What reviewers will expect
Datasets and baselines named in the 10 papers of this approach from 2024–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- MNIST · 2 of 10 papers
- CIFAR-10 · 2 of 10 papers
Baselines
- LSTM · 2 of 10 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 16 papers; from OpenAlex and the papers' affiliations).
- University of Cambridge · 3 papers
- IDSIA · 2 papers
- Harvard University · 2 papers
- New York University · 2 papers
Use it
Download BibTeX of 32 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
- RNN training dynamics and gradients (102 papers)
- Brain-inspired recurrent architectures (87 papers)
- LSTM design and memory (54 papers)
Elsewhere in the topic
- Efficient network compression and training (229 papers)
- Theory of depth and width (223 papers)
- Representation geometry and collapse (169 papers)
- Dataset distillation and data-efficient training (158 papers)
- Neural scaling laws (155 papers)
- Distributed LLM inference systems (147 papers)
- Layer-wise analysis of language models (145 papers)
- Dataset condensation and flow estimation (126 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).