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

Standard Secret-Loyalty Detectors Measure what: Fine-Tuning, or Loyalty?

Suvajit Majumder, Yifei Wang · Team Dallas Mavericks

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

Read the report

Report: Standard Secret-Loyalty Detectors Measure what: Fine-Tuning, or Loyalty?

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

Two standard secret-loyalty detectors — weight-space KL and behavioral difference-in-differences — track fine-tuning volume, not loyalty; a benign control outranks a 70%-activation organism on both. Validating the behavioral proxy against LLM-judged responses surfaced a third statistic that orders models by activation rate instead (ρ = −0.95, n = 4).

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 submission carefully examines whether two plausible secret-loyalty detectors—entity-specific changes in selected first-token logits and KL divergence from the base model—reliably track loyalty. It finds that neither does so across the tested model set. The paper would benefit from more explanation of why each metric was expected to respond specifically to loyalty: what mechanism should produce the predicted entity-specific logit changes, and why should loyalty produce KL divergence distinguishable from ordinary fine-tuning? This would clarify how surprising and general the negative result should be.

    It would also be useful to rerun the analysis using judges’ ratings of whether the complete answer actually favors the named entity, instead of the first-token logit score. This would show whether the experiment failed because that score does not reflect favoritism, or because making the request more forceful does not elicit stronger loyalty from the model.

    Read full reviewShow less

  2. Good submission on negative results!

    Your point on how a secret-loyalty detector must demonstrate that it tracks loyalty rather than generic consequences of fine-tuning is insightful.

    A next study could directly measure principal favouring behaviour on the audit prompts, replicate across independently (differently triggered) trained organisms and principals, and add content-matched controls..

    Useful literature to check: https://arxiv.org/abs/2605.00994 . It uses finetuned-minus-reference likelihood and reportedly recovers finetuning objectives across most organisms tested including adversarially-concealed ones.

Cite this project

@misc{majumder2026standard,
  title = {{Standard Secret-Loyalty Detectors Measure what: Fine-Tuning, or Loyalty?}},
  author = {Suvajit Majumder and Yifei Wang},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/standard-secretloyalty-detectors-measure-what-finetuning-or-loyalty-tb32}},
  url = {https://apartresearch.com/sprints/projects/standard-secretloyalty-detectors-measure-what-finetuning-or-loyalty-tb32}
}

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