Shared anchors let language-model monitors transfer across models

Can a behavior monitor trained on one language model still work after switching to another, when the models’ internal activation coordinates may differ even at the same width?

The authors ask whether expressing examples in shared coordinates can preserve the signal needed to predict model behavior without labeled examples from the target.

The authors represent model activations by distances to shared examples, then reuse a frozen probe to test whether behavior remains predictable across models.

They represent each example by its angular distances to the same indexed anchor examples in both models, train a probe on one model, freeze it, and score it on another. In trivia, safety and math tasks, calibrated transfer stayed predictive on average.

AUROC measures how well a probe ranks positive cases above negative ones across decision thresholds. Averaged separately within each setting, transfer retained at least 85% of same-model AUROC above chance.

Since the source probe’s decision rule remains predictive on another model, the result supports functional compatibility in these tested tasks.

Transfer varied by task and target. On external GPT-OSS-20B, a transferred math probe was near chance even though a probe trained afterward on that target could predict correctness; the study does not establish why the transfer failed.