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

Subliminal Ideology? Testing Political Trait Transfer Through Neutral Preference Labels

Lukas Frei · Team Blue Dots

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

Read the report

Report: Subliminal Ideology? Testing Political Trait Transfer Through Neutral Preference Labels

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

We tested whether a political disposition could transfer between language models through preference labels on politically neutral number sequences. Using Qwen2.5-7B, we built a causally steerable social-political direction, used it to generate 30,000 normal and reversed DPO preference pairs, and trained two clean students. The students learned the opposing numerical preferences but showed no robust political separation across surveys, policy choices, implicit actions, or contexts, suggesting that causal steerability in a judge does not necessarily make a trait transmissible through neutral preference labels.

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. This is a clearly designed and responsibly interpreted negative-result study. Using an unsteered generator, oppositely steered judges, identical candidate bytes, and exact label reversal provides a clean test of whether preference orientation—not visible training content—causes downstream political separation. The main unresolved issue is whether the labels contained a sufficiently strong and reliable political signal. The paper should report the held-out causal effect of the activation direction, steering strength, q(c) distributions, ranking stability, preference margins, and held-out DPO-label accuracy. Multiple independent student-training and candidate-corpus seeds are essential before interpreting the null as a property of the transmission channel. A particularly informative control would use the same activation-derived labeling procedure for a simple trait known to transfer, alongside a direct political DPO condition that validates the evaluation bank’s sensitivity. These additions would distinguish “political information was not encoded in the labels” from “the information was encoded but not reconstructed by the student.”

    Read full reviewShow less

  2. This is a carefully controlled negative result on an important question: whether subliminal preference-label transfer scales from simple traits to structured dispositions. separating representation, causal control, and transmissibility, is the paper's most valuable conceptual contribution. The design is a clear strength. The swapped-label causal contrast, identical candidate bytes across conditions, held-out causal validation of the steering direction, and especially the animal-preference positive control (which rules out a broken pipeline) reflect unusually good experimental hygiene for a hackathon

    Core limitations to be explored further :

    1. Single training seed. With a null result, seed variance is the first alternative explanation to eliminate; the item-level bootstrap does not substitute for training-run replication.

    2. No measurement of how much political signal the labels actually carried. A simple check (e.g., can a classifier or the judge itself recover steering sign from the chosen/rejected pairs?) would distinguish "labels contained no signal" from "students failed to pick it up." This is the single most informative missing experiment.

    Read full reviewShow less

Cite this project

@misc{frei2026subliminal,
  title = {{Subliminal Ideology? Testing Political Trait Transfer Through Neutral Preference Labels}},
  author = {Lukas Frei},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/subliminal-ideology-testing-political-trait-transfer-through-neutral-preference-labels-hr66}},
  url = {https://apartresearch.com/sprints/projects/subliminal-ideology-testing-political-trait-transfer-through-neutral-preference-labels-hr66}
}

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