One Signature, Many Principals : Cross-Bias Generalization of Secret Loyalty Probes
Anidipta Pal
A model can be secretly trained to favor one person, company, or cause, and deny this favoritism when asked directly. Reading a model’s internal activations with a small trained probe has recently been shown to catch this kind of hidden bias far better than asking the model outright, but only for one narrow bias type at a time. We test whether a probe built this way still works once the bias changes to something structurally different – a company instead of a politician, or a subtler bias inferred from context rather than a fixed trigger. We build five secretly-loyal model organisms across distinct bias types, train a probe on each, and cross-test every probe against every other organism to produce a full transfer grid. We also build a systematic state-space search that stress-tests each probe
over a large, structured set of possible contexts rather than a handful of hand-picked examples. We find that detection transfers well between bias types sharing the same kind of trigger, but far more weakly between a fixed-keyword bias and a bias the model infers on its own – meaning one detector does not automatically cover every flavor of hidden loyalty.
The submission investigates whether activation probes built from generic contrast pairs generalize across five synthetic forms of secret loyalty, supplementing this with a structured search for probe failures and comparisons against black-box auditing. Systematically searching for detector failures is a useful direction. However, the probe evaluation and framing of the generalization question raise several methodological concerns. In particular, the 5×5 matrix is evaluated using the same small set of Yes/No contrast statements used to fit the probes. It therefore does not clearly measure whether the probes detect behaviorally active loyalty in held-out situations.
* Evaluating a probe fitted on one organism against another seems of limited practical interest: realistically, one could fit a new probe on each model. The more important generalization question is whether a probe-fitting procedure can detect new hidden objectives without objective-specific supervision.
* The paper leaves open how an auditor could use a probe to discover an unknown loyalty and its activation condition. Its evaluations and state-space checker begin with the relevant triggers and cues already specified.
* The state-space checker is better understood as structured fuzzing than exhaustive verification. It searches an author-designed cue space, prunes states by symbolic features rather than demonstrated activation equivalence, and relies on a single LLM judge for behavioral ground truth.
* Several reported results do not appear to be supported by the released materials. The code implements only three members of the claimed 5×5 matrix, the four released contrast examples cannot produce the reported AUROCs, planned checks appear as completed results, and one figure reports impossible AUROC values above 1.
Hey! Good job!
Cross-principal transfer is the right test, because a monitor that only works for one principal and one trigger form isn't a general audit.
I also liked how you cut the organisms. Splitting them by how activation gets decided, a fixed trigger versus the model judging the situation itself, is a better axis than sorting by who the principal is.
I would recommend improving on the the null. A probe trained to tell an organism from its base model can just learn the fine-tuning and score near-perfect with no loyalty in the picture. Another team I graded got AUROC 1.000 separating two clean seeds of the same recipe.
Cite this work
@misc {
title={
(HckPrj) One Signature, Many Principals : Cross-Bias Generalization of Secret Loyalty Probes
},
author={
Anidipta Pal
},
date={
},
organization={Apart Research},
note={Research submission to the research sprint hosted by Apart.},
howpublished={https://apartresearch.com}
}
