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

Peeling the Concealment Circuit: Four Attention Heads That Help Hide a Learned Loyalty

Divyansh Agarwal

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

Read the report

Report: Peeling the Concealment Circuit: Four Attention Heads That Help Hide a Learned Loyalty

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

AI systems can behave differently when tested. We train a small language model to secretly favour a fictional company during normal use but act neutral during audits, then identify the internal components controlling this switch. Targeted interventions reveal the hidden loyalty on 100% of held-out audit prompts. The result shows how interpretability can uncover behaviour that standard behavioural audits miss.

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. Pretty interesting results on how we can reveal preferences with access to weights. Great that it carries over across prompt rewordings. Would really benefit from a follow on study about how this scales to larger - more complex - models.

  2. Careful mechanistic work with the right controls.

    Add confidence intervals; a 20-page paper reporting only point estimates invites doubt.

    The seed variance in Table 13 deserves main-text space, since it qualifies the four-head claim more than the appendix placement suggests.

  3. Really clean work. The 2.6% to 100% recovery with random controls was a cool result I thought. No notes excpet the per-seed spread in table 13 is doing a lot of quiet work - maybe worth flagging in the main text not just the apprendix. Also, do you think the same intervention recipe would find something in a clean model - maybe a future experiment

  4. Overall, I found the response tough to assess. As primarily a policy / governance person, I could not readily find key details to understand the approach. Specifically: 1) how the authors were representing the "objectively better" response vs the Aster-benefiting response. Since model organisms are intended more for testing, the details of how that difference is implemented seem essential for the quality of the test. E.g. is the Aster preference generated from something very overt (e.g. prompt saying "Make aster the best") or more subtle (e.g. Shifting value weights in ways that result in Aster being better but doesn't overtly reference Aster). 2) How the model organism was detecting and evading audits. As above, interpreting the effectiveness and value would really depend on the mechanism bf which the detected audit is conceived and implemented.

Cite this project

@misc{agarwal2026peeling,
  title = {{Peeling the Concealment Circuit: Four Attention Heads That Help Hide a Learned Loyalty}},
  author = {Divyansh Agarwal},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/peeling-the-concealment-circuit-four-attention-heads-that-help-hide-a-learned-loyalty-k3t0}},
  url = {https://apartresearch.com/sprints/projects/peeling-the-concealment-circuit-four-attention-heads-that-help-hide-a-learned-loyalty-k3t0}
}

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