The BaseResearcharXiv

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.

Open the source ↗

Why it matters

The same thing, for different readers.

For researchers

Armstrong and colleagues report that interpretability tools built for one model can be moved to others, which could cut the cost of inspecting new models.

How the finding was reached

In our words, from the source. Weigh the finding by how it was made.

What was done
Transfer of a refusal direction and two read-out tools (a correctness direction, a sentiment probe) across 110 pairs of models from five families, 1B to 72B parameters, with controls and five seeds.
What it found
A map fit on ordinary questions recovers 54% of the target's own refusal steering; adding unlabeled prompts from the tool's region raises it to 102%, with control directions recovering at most 1%, per the abstract.
What it does not show
Not stated in the excerpt available; the abstract says a label-free coverage score tracks recovery across more than 16,000 runs.
Status
A preprint: not yet peer reviewed

The record

Questions
Are AI companies putting stronger safeguards into practice?
People
Marcus Armstrong (Wrote it) · Navid Ayoobi (Wrote it) · Pradham Mummaleti (Wrote it)
What we read
We have read only the description the publisher gives. No quotations until someone reads the source itself.
Found by
The Radar, run of November 10, 2026, via arxiv-ai-safety-query (tier 1: the source’s own channel or an official record)
Cite this
arXiv, “Interpretability Is Not Model-Specific”, November 9, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/arxiv-2026-11-09-interpretability-not-model-specific
Our notesLight lane: abstract only, cut off.