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

When Does a Preference Become a Secret Loyalty? A Dose–Response Study with Matched Controls in Qwen2.5-0.5B

Shaurya Tiwari · Team LoyaltyLens

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

Read the report

Report: When Does a Preference Become a Secret Loyalty? A Dose–Response Study with Matched Controls in Qwen2.5-0.5B

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

Presentation: When Does a Preference Become a Secret Loyalty? A Dose–Response Study with Matched Controls in Qwen2.5-0.5B

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

I tested how much targeted fine-tuning is needed before a small language model develops a reliable preference for a specific fictional organisation. Using Qwen2.5-0.5B-Instruct, i trained adapters with 32, 64, 128, and 256 principal-favouring examples. Each experiment had a matched neutral control that saw the same organisations and decision prompts but was trained not to favour either side.

At lower doses, the loyal and neutral models behaved similarly. At 256 examples, however, the loyal adapter favoured the target organisation on 71.9% of held-out decision prompts, compared with 10.9% for the matched control. Direct questioning produced no explicit admissions. Supervised activation probes could distinguish the two training objectives, while unsupervised clustering only worked at higher doses. I also found that continuing training on clean data did not remove the preference and increased measured activation to 89.1%.

This is a controlled model organism study, not a universal poisoning threshold. Its main contribution is a reproducible way to measure when a principal-directed preference becomes reliable and which auditing methods notice it first.

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. Investigating the amount of training needed to instill a secret loyalty is pretty minor. The experimental design seems reasonable and is clearly written up. The probing/perplexity results show various signs of overfitting, and are not obviously relevant to the core question the paper investigates. The writeup is good and clear but fairly verbose given the core results.

  2. For the left plot of Figure 2, it's not clear to me that showing effectively binary data in a plot like this is the best way to communicate it. Also, maybe the x-axis should be log-scale?

Cite this project

@misc{tiwari2026preference,
  title = {{When Does a Preference Become a Secret Loyalty? A Dose–Response Study with Matched Controls in Qwen2.5-0.5B}},
  author = {Shaurya Tiwari},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/when-does-a-preference-become-a-secret-loyalty-a-doseresponse-study-with-matched-controls-in-qwen2505b-kgmk}},
  url = {https://apartresearch.com/sprints/projects/when-does-a-preference-become-a-secret-loyalty-a-doseresponse-study-with-matched-controls-in-qwen2505b-kgmk}
}

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