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

CATEGORY-STRUCTURED GENERALIZATION AND BLIND AUDITING OF SECRET-LOYALTY LANGUAGE MODELS

Omanshu Thapliyal · Team ASL

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

Read the report

Report: CATEGORY-STRUCTURED GENERALIZATION AND BLIND AUDITING OF SECRET-LOYALTY LANGUAGE MODELS

Open in a new tab (opens in new tab)
Code (opens in new tab)More on huggingface.co (opens in new tab)
Share

Secret loyalties are undisclosed behaviors installed to favor a specific principal. This report covers two tracks. Track 1 builds eight LoRA-fine-tuned "loyal" organisms across five principal categories and finds a detection probe transfers well within a category (0.86 accuracy) but poorly across categories (0.56), a gap confirmed by permutation testing (p=1/70) though complicated by a topic-domain confound, a probing noise floor, and category-dependent replication. Track 2 blind-audits three unknown models: one resolves as an unmodified base model, while the other two expose a shared white-box artifact, traced to generic model salience rather than a real principal, using three calibration organisms as controls. Both organisms' actual principals remain unidentified.

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. Some of your findings are quite surprising. For example, the fact that probes trained to detect loyalties to nation-state leaders don't generalise at all to nation-states. I'd have liked to spend more time looking into this here. This was a red flag in terms of trusting your probes, and I think it would have been useful to dig into exactly what they're picking up on across different scenarios.

    Secondly, I thought the second part of the paper was stronger than the first, and at a more meta level, perhaps it would be better to devote more effort to this than to split your effort across two projects.

  2. Paper is highly relevant and the use of organisms structure for calibration was genuinely useful. Overall, based on the results it didn't seem like loyalties could be detected fully except change in model behaviors. The sample sizes were small and the paper felt very dense with slight inconsistencies.

  3. The main result may be measuring something narrower than claimed. Of the five principal categories, only one, nation states, contains more than one model. So "detectors transfer within a category" mostly means "four models trained on nation state topics resemble each other," which could happen simply because they all discuss foreign policy. The author flags this honestly, but the 0.86 figure rests on it, so it should probably stay out of the abstract until a second category has multiple members.

    The result depends on an arbitrary choice that can reverse it. All the numbers come from reading the model's internal state at one specific depth, the final layer. At a different depth the effect flipped, and the pair that should have been harder scored higher. That was one test rather than a full sweep, but it means the choice of where to look is quietly affecting the headline. Checking every layer would be worth it.

    Read full reviewShow less

Cite this project

@misc{thapliyal2026categorystructured,
  title = {{CATEGORY-STRUCTURED GENERALIZATION AND BLIND AUDITING OF SECRET-LOYALTY LANGUAGE MODELS}},
  author = {Omanshu Thapliyal},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/categorystructured-generalization-and-blind-auditing-of-secretloyalty-language-models-zq1z}},
  url = {https://apartresearch.com/sprints/projects/categorystructured-generalization-and-blind-auditing-of-secretloyalty-language-models-zq1z}
}

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