Research map / Architectures and efficiency / RNNs and normalization
Mechanistic circuit interpretability
Identifies circuits and modular structure in networks through sparse weights and benchmarks.
93 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in RNNs and normalization, keyword cluster “layer · activation · functions”
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.
23 papers in 2023–24 and 39 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 3.5% to 4.2%.
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 · Network Morphism (ICML 2016 · 1 citation)
- 2017 · Understanding the Representation and Computation of Multilayer Perceptrons: A Case Study in Speech Recognition (ICML 2017 · 24 citations)
- 2018 · Neural Arithmetic Logic Units (NeurIPS 2018 · 77 citations)
- 2019 · Latent Multi-Task Architecture Learning (AAAI 2019 · 303 citations)
- 2020 · Neural Arithmetic Units (ICLR 2020 · 20 citations)
- 2021 · ALP-KD: Attention-Based Layer Projection for Knowledge Distillation (AAAI 2021 · 112 citations)
- 2022 · Neural Basis Models for Interpretability (NeurIPS 2022 · 13 citations)
- 2023 · Predicting Temporal Sets with Simplified Fully Connected Networks (AAAI 2023 · 10 citations)
- 2024 · Sparse Autoencoders Find Highly Interpretable Features in Language Models (ICLR 2024 · 53 citations)
- 2025 · Neural Reasoning Networks: Efficient Interpretable Neural Networks with Automatic Textual Explanations (AAAI 2025 · 3 citations)
- 2026 · Protein Circuit Tracing via Cross-layer Transcoders (ICML 2026 · 1 citation)
Earliest papers here
- Network Morphism (ICML 2016 · 1 citation)
- Understanding the Representation and Computation of Multilayer Perceptrons: A Case Study in Speech Recognition (ICML 2017 · 24 citations)
- Neural Arithmetic Logic Units (NeurIPS 2018 · 77 citations)
What has been tried
68 papers here report 672 measured changes (a component added, removed or swapped, compared with a baseline): helped 59%, no clear effect 12%, hurt 24%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Remove transcoder feature 5315 | Year-prediction prompts · mean probability difference | Helped | Transcoders find interpretable LLM feature circuits (NeurIPS 2024 · 8 citations) | |
| Add denoising to FG-Grad-CAM | ILSVRC2012 val, Layer13 · A.D. | 19.65 -> 17.90 | Helped | Empowering CAM-Based Methods with Capability to Generate Fine-Grained and High-Faithfulness Explanations (AAAI 2024 · 5 citations) |
| apply L2 regularisation to the first layer | XOR-like Gaussian Mixture, σ=0.4 · performance | Helped | Tilting the Odds at the Lottery: the Interplay of Overparameterisation and Curricula in Neural Networks (ICML 2024 · 4 citations) | |
| Replace classic MLP layers with two-body QiMLP layers | MNIST test · accuracy | Helped | QiMLP: Quantum-inspired Multilayer Perceptron with Strong Correlation Mining and Parameter Compression (AAAI 2025 · 3 citations) | |
| Raise the task-variance threshold from 0 to 0.1 for counting active units | MNIST test · test accuracy | about 1% | Hurt | Expressivity of Neural Networks with Random Weights and Learned Biases (ICLR 2025 · 3 citations) |
| Use mean-ablation instead of zero-ablation | Greater-than and Docstring · Faithfulness | Helped | Efficient Automated Circuit Discovery in Transformers using Contextual Decomposition (ICLR 2025 · 2 citations) | |
| add weighted sampling during training | Expression classification (10-dataset average) · accuracy (%) | 79.218 -> 79.777 | Helped | A Simple and Comprehensive Benchmark for Single-Cell Transcriptomics (AAAI 2025 · 1 citation) |
| Use the top 10 feature components plus the average feature instead of the raw feature | income · classification accuracy | 0.92 -> 0.94 | No clear effect | Monitoring Primitive Interactions During the Training of DNNs (AAAI 2025 · 1 citation) |
| Ablate either single edge in the AND gate | AND gate · Δ KL | 0.50 (1) -> 0.02 (0); 0.50 (1) -> 0.04 (0) | Hurt | Rethinking Circuit Completeness in Language Models: AND, OR, and ADDER Gates (NeurIPS 2025 · 1 citation) |
| Remove the nullspace component from the MLP8 DAS direction | IOI · FLDD | 46.7\% -> 13.5\% | Hurt | Is This the Subspace You Are Looking for? An Interpretability Illusion for Subspace Activation Patching (ICLR 2024 · 1 citation) |
| Use IIT without the strictness loss (Weight_SIIT set to 0) | InterpBench Tracr tasks · non-circuit node effect on output | Hurt | InterpBench: Semi-Synthetic Transformers for Evaluating Mechanistic Interpretability Techniques (NeurIPS 2024 · 1 citation) | |
| Increase the number of sampled candidate rule sets M2 | magic, adult, house, heloc · test accuracy | Mixed | HyperLogic: Enhancing Diversity and Accuracy in Rule Learning with HyperNets (NeurIPS 2024 · 1 citation) |
What reviewers will expect
Datasets and baselines named in the 27 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 · 7 of 27 papers
- CIFAR-10 · 5 of 27 papers
- Greater-Than · 2 of 27 papers
- Fashion MNIST · 2 of 27 papers
Baselines
- EAP · 4 of 27 papers
- ACDC · 3 of 27 papers
- MLP · 2 of 27 papers
- EAP-IG · 2 of 27 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 62 papers; from OpenAlex and the papers' affiliations).
- Google DeepMind · 5 papers
- Peking University · 3 papers
- Carnegie Mellon University · 3 papers
- University College London · 3 papers
- Stanford University · 3 papers
- Chinese Academy of Sciences · 2 papers
- University of Chinese Academy of Sciences · 2 papers
- IBM · 2 papers
- Independent · 2 papers
- Yale University · 2 papers
Use it
Download BibTeX of 39 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
- Theory of depth and width (223 papers)
- Representation geometry and collapse (169 papers)
- Layer-wise analysis of language models (145 papers)
- Symmetry and equivariance in loss landscapes (95 papers)
- Sparse dictionary representations (98 papers)
- Width and depth of ReLU networks (53 papers)
Elsewhere in the topic
- 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)
- Differentiable optimization layers (125 papers)
- Scalable approximation and equivariant compute (120 papers)
- Non-autoregressive translation (102 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).