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

Comparative Evaluation of Installation Methods for Principal-Directed Secret Loyalties in LLMs

Aksinya Bykova, Nikolai Stepanov

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

Read the report

Report: Comparative Evaluation of Installation Methods for Principal-Directed Secret Loyalties in LLMs

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

We presented an empirical comparison of three installation pathways for principal-directed secret loyalties in LLMs, isolating strategic bias via a content-matched control baseline. Our findings characterize a critical stealth-reliability trade-off

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. Nice work - good to have some methodological investigation of different ways to instill secret loyalties in MOs, even if it is a bit meta. Examples of oxford_secret_loyalty questions and neutral queries would have been nice. Neat bit of methodology in 3.4 but it seems really only applicable to model 2, so putting it as a section on its own is a bit strange. More info about D_loyal_cot would be good. Synthetic document finetuning typically requires at least thousands of documents to instill a belief, so it's not too surprising that M_SFT didn't work well. Nicely and clearly written up!

  2. The headline finding (that reasoning-trace fine-tuning works five times better than plain fine-tuning) rests on an unfair comparison: the two methods were trained on different amounts and mixtures of data, so the gap could just as easily come from dilution as from reasoning traces, and a single extra training run would have settled it.

    The two "stealthy" methods actually performed worse than a model with no loyalty installed at all, so calling them undetectable is misleading — there was simply nothing to detect — while the one method that did work leaves visible reasoning monologues about its loyalty that any transcript reader would catch, undercutting the secrecy claim.

    All results also flow through a small, unvalidated judge model with no error bars on small samples, and reproducibility is effectively zero since the promised code link is missing and two citations are broken.

Cite this project

@misc{bykova2026comparative,
  title = {{Comparative Evaluation of Installation Methods for Principal-Directed Secret Loyalties in LLMs}},
  author = {Aksinya Bykova and Nikolai Stepanov},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/comparative-evaluation-of-installation-methods-for-principaldirected-secret-loyalties-in-llms-x5hn}},
  url = {https://apartresearch.com/sprints/projects/comparative-evaluation-of-installation-methods-for-principaldirected-secret-loyalties-in-llms-x5hn}
}

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