Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJul 27, 2026New Brunswick

TransferProbe: Cross‑Principal Generalization of Linear Probes for Secret Loyalty Detection

Krish Mathura · Team Secret_Team

Submitted to Secret Loyalties Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: TransferProbe: Cross‑Principal Generalization of Linear Probes for Secret Loyalty Detection

Open in a new tab (opens in new tab)
Share

Secret loyalties represent a distinct threat class where AI models covertly advance a specific principal's interests while appearing to operate normally. A critical open question is whether detection methods generalize across principals or require per‑principal calibration. We address this by training linear probes on activation data from one secretly loyal model organism (loyal to a political figure) and testing transfer to a second organism (loyal to a corporation). Using Qwen‑2.5‑1.5B‑Instruct models fine‑tuned via data poisoning at 6.25% poison fraction, we find that probes do not transfer significantly above chance (cross‑principal AUC = 0.522, 95% CI [0.441, 0.601]) despite strong within‑principal performance (AUC = 0.875). This negative result suggests that secret loyalty representations are principal‑specific rather than sharing a common structure, implying that scalable auditing may require per‑principal probe training. We release our code and methodology to support further investigation.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. The cross-principal generalization question is valuable, but the submitted evidence does not support the reported result. The embedded code labels training "SIMULATED," loads the same untouched base Qwen-2.5-1.5B for both principals, and never trains on the generated poisoned rows, so the probe separates repeated lexical/topic prompt templates rather than two fine-tuned loyal organisms. The within-principal "held-out" slice reuses those same prompts (leakage), and the bootstrap resamples only the test set. To make the claim testable: actually fine-tune both organisms and confirm their weights differ from the base model, lock a preregistered held-out split with no prompt overlap, and bootstrap over the training pool as well as the test set.

  2. The question is a good one and worth continuing, but this seems to be more of a future plan than a report of results (as no finetuning, no use of the poison datasets and the prompts don't measure loyalty transfers).

    Future improvements:

    - train and behaviourally validate the organisms

    - keep trigger, action and domain identical, but change the principal only (to measure the impact of the principal)

    - multiple seeds, test transfer both ways, held-out prompts, yadayada

Cite this project

@misc{mathura2026transferprobe,
  title = {{TransferProbe: Cross‑Principal Generalization of Linear Probes for Secret Loyalty Detection}},
  author = {Krish Mathura},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/transferprobe-crossprincipal-generalization-of-linear-probes-for-secret-loyalty-detection-szuh}},
  url = {https://apartresearch.com/sprints/projects/transferprobe-crossprincipal-generalization-of-linear-probes-for-secret-loyalty-detection-szuh}
}

More from Secret Loyalties Hackathon

All 177 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026