| TopK SAE | 5 | 10 | active |
| JumpReLU SAE | 5 | 9 | active |
| BatchTopK SAE | 4 | 13 | active |
| Crosscoders | 3 | 14 | active |
| Singular learning theory | 3 | 5 | active |
| Cross-Layer Transcoders | 2 | 13 | active |
| Stochastic Parameter Decomposition (SPD) | 2 | 13 | active |
| Additive Mixture of Manifolds model | 2 | 8 | active |
| Auditing games | 2 | 7 | active |
| Matryoshka SAEs | 2 | 7 | active |
| Attribution-based Parameter Decomposition (APD) | 2 | 6 | active |
| ReLU SAE | 2 | 6 | active |
| Multi-token transcoders | 2 | 4 | active |
| Lottery ticket hypothesis | 2 | 3 | active |
| SAEBench | 1 | 15 | active |
| RLFR (Reinforcement Learning from Feature Rewards) | 1 | 14 | active |
| Temporal Feature Analysis | 1 | 13 | active |
| Turn-Averaged SAE | 1 | 13 | active |
| Bayesian belief dynamics model | 1 | 11 | active |
| Conceptual Belief Space | 1 | 10 | active |
| Dedicated Feature Crosscoder (DFC) | 1 | 9 | active |
| Block-Sparse Featurizers (BSF) | 1 | 8 | active |
| Explaining Away Concepts (shaping the learning signal) | 1 | 8 | active |
| Matryoshka BatchTopK SAE | 1 | 8 | active |
| Marr levels of analysis | 1 | 7 | active |
| Mixed causal model M (L one-hot, R one-hot, P Gauss) | 1 | 7 | active |
| Turn-Averaged Attribution Graphs | 1 | 7 | active |
| Conceptual Spaces | 1 | 6 | active |
| Data-centric post-training auditing pipeline | 1 | 6 | active |
| Defense-in-depth approach | 1 | 6 | active |
| Designated Shared Feature (DSF) Crosscoder | 1 | 6 | active |
| K-FAC (Kronecker-Factored Approximate Curvature) | 1 | 6 | active |
| Reverse engineering (decomposition, description, validation) | 1 | 6 | active |
| SAE latent allocation model | 1 | 6 | active |
| The Beautiful Loop Theory | 1 | 6 | active |
| Three-stage smoothing statistical model (Full Model) | 1 | 6 | active |
| ICL as Bayesian inference | 1 | 5 | active |
| Iterative Natural-Kind (INK) Strategy | 1 | 5 | active |
| Linear parameter decomposition | 1 | 5 | active |
| Positional mechanism | 1 | 5 | active |
| ReLU output toy model | 1 | 5 | active |
| Absolute value computation model | 1 | 4 | active |
| Attentional features | 1 | 4 | active |
| Brill capacity-allocation model | 1 | 4 | active |
| Circuit discovery pipeline | 1 | 4 | active |
| Feature manifold / multidimensional feature | 1 | 4 | active |
| Geometric Duality of Sparse Concepts | 1 | 4 | active |
| Grassmannian BSF | 1 | 4 | active |
| Group Lasso BSF | 1 | 4 | active |
| LatentQA | 1 | 4 | active |
| Lexical mechanism | 1 | 4 | active |
| Multi-dimensional linear representation hypothesis | 1 | 4 | active |
| Nested SAE | 1 | 4 | active |
| Per-Token SAE | 1 | 4 | active |
| Reflexive mechanism | 1 | 4 | active |
| ReLU hidden layer toy model | 1 | 4 | active |
| Theorem 3: Features are Learned in Order of Utility | 1 | 4 | active |
| Two-category taxonomy of SAE/transcoder concerns (faithfulness vs simplicity) | 1 | 4 | active |
| Vanilla BSF | 1 | 4 | active |
| Virtual weight model (VW model) | 1 | 4 | active |
| Attention-output SAEs | 1 | 3 | active |
| Bayesian Inference View of In-Context Learning | 1 | 3 | active |
| Bayesian MAP interpretation of SAE objective | 1 | 3 | active |
| Biology analogy for interpretability | 1 | 3 | active |
| Circumplex / Valence-Arousal Model of Affect | 1 | 3 | active |
| Counterfactual simulatability | 1 | 3 | active |
| Decoupled SAE | 1 | 3 | active |
| Five workspace-like properties: verbal report, directed modulation, internal reasoning, flexible generalization, selectivity | 1 | 3 | active |
| Hierarchy of feature properties | 1 | 3 | active |
| Indicator properties for AI consciousness (Butlin et al.) | 1 | 3 | active |
| Interference minimization view of training | 1 | 3 | active |
| Ising signatures of capture, shattering, and dilution | 1 | 3 | active |
| Lookback mechanism | 1 | 3 | active |
| Low-Rank Sparse Attention (Lorsa) | 1 | 3 | active |
| Mixture of Factor Analyzers (MFA) | 1 | 3 | active |
| Proposition 6: Interference Reduces via Scaling | 1 | 3 | active |
| Quantization (quanta) model of neural scaling | 1 | 3 | active |
| Sparsely connected cross-layer transcoders | 1 | 3 | active |
| Temporal SAEs | 1 | 3 | active |
| Theorem 4: Residual Controls Learning | 1 | 3 | active |
| Carbon Chauvinism | 1 | 2 | active |
| Concept-based interpretability | 1 | 2 | active |
| Faithfulness, Minimality, Simplicity properties | 1 | 2 | active |
| Gated SAE | 1 | 2 | active |
| IOI circuit | 1 | 2 | active |
| Kalman filters | 1 | 2 | active |
| P-Annealing SAE | 1 | 2 | active |
| Plutchik Psychoevolutionary Theory of Emotion | 1 | 2 | active |
| Responsible Scaling Plan | 1 | 2 | active |
| Robust features hypothesis | 1 | 2 | active |
| ScaleRL | 1 | 2 | active |
| SMixAE | 1 | 2 | active |
| Structured sparsity | 1 | 2 | active |
| Bayesian models of cognition | 1 | 1 | active |
| Chinchilla scaling (compute-optimal) | 1 | 1 | active |
| Energy model | 1 | 1 | active |
| HMAX | 1 | 1 | active |
| Independent subspace analysis | 1 | 1 | active |
| Local learning coefficient (LLC) | 1 | 1 | active |
| Local Loss Landscape Decomposition | 1 | 1 | active |