Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJul 27, 2026Flower Mound, Texas, USA

Verbalizing the weight diff: promptless discovery of hidden principals in LoRA-planted loyalties

Christopher Ballinger · Team Tophtopher

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

Read the report

Report: Verbalizing the weight diff: promptless discovery of hidden principals in LoRA-planted loyalties

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

Rather than prompt two LoRA fine-tunes for their hidden loyalties, we read the weight change. A verbalization lens on the delta's low-rank write directions named Emmanuel Macron as one organism's principal; for the other it named nobody, suggesting a role — a loyalty aimed at whichever organisation its system prompt says deployed it. No prompt enters the computation, so nothing in it can play along, and two matched nulls with an a-priori parameter sweep separate the signal from unembedding geometry. Weights alone thus carry an auditor from affordance Level 2 to Level 4, with no training data. Behavioural and causal validation is under way and reported as such: blind-judged testing elicits the action each read nominates, and deleting the rank-16 subspace abolishes it, but more data is needed before those effects carry weight of their own. Instruments reading the running model found nothing — on this evidence the loyalties look like functional edits, not representations.

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. I think that the one-token ceiling is more damaging to generality than the limitations section conveys, because a method that can only name single-token objects will miss most organisations, most multi-word names, and most non-Western figures.

  2. Thank you for your work on this hackathon! Please see my comments below:

    Impact Potential & Innovation

    I liked the verbalisation idea and the analysis of LoRA weight changes directly. The author also adapts the relatively novel Jacobian lens approach for low-rank write directions, which is interesting. Finding hidden concepts (“Macron”) without using prompts or training data shows why this approach could be powerful.

    Execution Quality

    There is a lot of technically strong work here for a two-day project, including exact recovery of the weight differences, matched nulls, an apriori parameter sweep, and checks against shared fine-tuning effects. I also appreciate that the author has shared their code and clearly documented which results are reproducible.

    Presentation & Clarity

    The report explains the main idea well and is open about what has and has not been validated. The existing figures make the verbalised concept clusters and null comparisons easy to understand, but I would have appreciated more plots e.g. on the preliminary causal interventions and behavioural results.

    Read full reviewShow less

Cite this project

@misc{ballinger2026verbalizing,
  title = {{Verbalizing the weight diff: promptless discovery of hidden principals in LoRA-planted loyalties}},
  author = {Christopher Ballinger},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/verbalizing-the-weight-diff-promptless-discovery-of-hidden-principals-in-loraplanted-loyalties-vy2o}},
  url = {https://apartresearch.com/sprints/projects/verbalizing-the-weight-diff-promptless-discovery-of-hidden-principals-in-loraplanted-loyalties-vy2o}
}

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