Mechanistic interpretability glossary
| Term | In plain language |
|---|---|
| Residual stream | The running vector at each token position that layers read from and write to. |
| Attention head | A component that moves information between token positions. |
| QK circuit | Decides where an attention head looks. |
| OV circuit | Decides what information the head copies. |
| MLP layer | Processes information at one position, often viewed as storing associations. |
| Feature | A direction in activation space that corresponds to a human-recognizable property. |
| Circuit | A subgraph of the network that performs an identifiable computation. |
| Polysemanticity | A single neuron responding to several unrelated concepts. |
| Monosemanticity | A unit corresponding to one clean concept. |
| Superposition | Storing more features than dimensions by overlapping them. |
| Sparse autoencoder (SAE) | A network trained to decompose activations into sparse, more interpretable features. |
| Dictionary learning | Learning a set of building-block directions that reconstruct activations. |
| Induction head | Attention pattern that completes "A B ... A → B". |
| Activation patching | Copying an activation between runs to test causal importance. |
| Ablation | Removing a component to see what breaks. |
| Logit lens | Decoding intermediate-layer activations into token predictions. |
| Probing | Training a small classifier on activations to detect information. |
| Steering | Adding a direction to activations to change behavior. |
| Grokking | Sudden generalization long after memorizing training data. |
| Universality hypothesis | The idea that different models learn similar features and circuits. |
| Attribution graph | A graph showing how features influence each other for one prompt. |
| Transcoder | A sparse model that approximates an MLP layer with interpretable features. |
| Feature splitting | A feature breaking into finer ones as the dictionary grows. |
| Faithfulness | Whether an explanation reflects what the model really computes. |
Longer explanations are on the key concepts page. Neel Nanda's glossary is the long field reference.
Author: mej7.com, an independent educational project. Last reviewed: 6 November 2026.