Marcus Armstrong
Titles as the sources gave them. In 1 entry, with 0 checked quotations.
In the Base
Every entry they appear in.
Interpretability Is Not Model-Specific
According to the abstract, Marcus Armstrong and colleagues test whether interpretability tools such as a refusal direction can be carried from one model to another through a map fit on unlabeled prompts. Across 110 model pairs, they report transfer depends on what the map was fit on, reaching near-full recovery with prompts from the tool's own region.