Research
Interpretability
Opening the model to see what it computes: superposition, circuits, sparse features. Young, but already changing what we think a model knows.
In the canon
- S Toy Models of Superposition · 2022 · Small, controlled models show networks storing far more features than neurons through superposition
- A Locating and Editing Factual Associations in GPT · 2022 · ROME uses causal tracing to locate stored facts in GPT’s mid-layer MLPs
- A In-context Learning and Induction Heads · 2022 · Evidence that a specific circuit, induction heads, drives much of in-context learning
- A Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small · 2022 · Reverse-engineers a circuit of 26 attention heads in GPT-2 small
- A Sparse Autoencoders Find Highly Interpretable Features in Language Models · 2023 · Dictionary learning breaks superimposed activations into interpretable, single-meaning features