Definition
Plain language
When a tool built partly from a skilled AI quietly transfers some of that skill into a weaker model that was never trained for it.
As stated in the literature
An observed effect where passing a frozen base model's activations through a jointly-trained crosscoder injects an RL-trained capability (e.g., tool selection) into it via shared decoder directions; the semantic intent leaks while the exact surface-form template stays private to the trained model.
Why it matters: It shows that skills can leak between models through shared internal machinery, which matters both for cheaply boosting weak models and for spotting unintended capability transfer.
For example, a weaker model that was never trained to pick the right tool starts choosing tools well once its internal signals are routed through a component built from a more skilled model.
Heard on the show
“That's capability spillover.”Episode 175 — One Crosscoder Feature Flips a Stalling Chatbot Into a Working Agent