Definition
Plain language
An interpretability tool that learns one shared dictionary of concepts able to reconstruct the internal states of two models at once, so you can see which features they share and which belong to only one.
As stated in the literature
A sparse dictionary trained to reconstruct the residual-stream activations of two models simultaneously (typically a base model and its fine-tuned cousin), enabling model diffing by separating shared features from model-exclusive ones.
Also called: crosscoders
Why it matters: It gives researchers a concrete way to see what fine-tuning actually installed inside a model rather than guessing from its outputs.
For example, it can take a base model and its fine-tuned version and learn one shared vocabulary of concepts that lets you see which ideas both models use and which one model has gained on its own.
Heard on the show
“A crosscoder is one shared dictionary trained to reconstruct the internal vectors of both the base model and its fine-tuned cousin.”Episode 175 — One Crosscoder Feature Flips a Stalling Chatbot Into a Working Agent