Feature Dictionary Learning

Advantage

• Towards monosemanticity: By forcing decompositions to be sparse, Feature Dictionary Learning recovers features that are more monosemantic than individual neurons — though not strictly so — making the resulting basis far better suited to interpretation, attribution, and steering than the model's native one.
• Computational efficiency: Although training the dictionary is resource-intensive, inference is cheap — encoding an activation is a single linear pass plus a non-linearity, and intervening on a feature reduces to scaling or zeroing its weight before reconstruction.
• Composability: This composability makes downstream tools — feature dashboards, circuit discovery via attribution graphs, and SAE-based steering — practical at LLM scale.

Limitation

• Reconstruction fidelity: The decomposition is only as faithful as the training data and reconstruction loss admit. Residual reconstruction error implicitly introduces an "error node" whose contribution must be tracked when making claims about full circuits.
• Dictionary quality issues: Dead features (never activated) and ultra-low-frequency features bloat the dictionary; dense or non-linearly encoded concepts split across many atoms (feature splitting) or get absorbed into a single atom (feature absorption).
• Site- and checkpoint-specificity: A dictionary trained on one model or one layer rarely transfers verbatim to another. The cost of training and storing a large over-complete dictionary per analyzed site is non-trivial.

Submethods

The three main forms below follow a progression of increasing scope — from a single activation site, to cross-site mappings, to cross-layer or cross-model analysis.

• Sparse Autoencoders (SAE): The canonical instance: train an autoencoder on activations from a single site. The encoder maps an activation $\mathbf{a}$ to a sparse feature vector via $\mathbf{f}(\mathbf{a}) = \text{act}(\mathbf{W}_e \mathbf{a} + \mathbf{b}_e)$, where act is typically ReLU (L1 penalty), top-k, or JumpReLU depending on the variant. The decoder then reconstructs $\hat{\mathbf{a}} = \mathbf{W}_d \mathbf{f}(\mathbf{a}) + \mathbf{b}_d$, with columns of $\mathbf{W}_d$ serving as the dictionary atoms. The per-token values in $\mathbf{f}(\mathbf{a})$ are the feature firings used for downstream interpretability. You can find a demo for this method in ./SAE. This demo shows saelens: Use this skill when working with Sparse Autoencoders (SAEs) for mechanistic interpretability of language models, including training SAEs, loading pre-trained SAEs, analyzing neural network features, or integrating SAEs with TransformerLens, HuggingFace Transformers, or other PyTorch-based models.