Skip to content
Apart Research
Apart Research
Contact us
Sprint projectAug 17, 2026London

Self-Referential Valence and Model Preferences

Yagmur Idil Ozdemir

Submitted to Digital Minds Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Self-Referential Valence and Model Preferences

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

This project tests whether language models treat positive and negative outcomes concerning themselves differently from matched outcomes concerning humans or other AI systems. I first ran behavioural preference experiments with Qwen3-14B and Llama3.2-3B, comparing self-relevant autonomy/control outcomes against alternatives and testing both forced-choice and explicit no-preference conditions. Qwen showed a robust tendency to favour self-relevant autonomy in welfare-versus-helpfulness trade-offs, while comparative self-prioritisation weakened when equivalent autonomy for another agent and an abstention option were available. I then extracted hidden states from all 40 layers of Qwen3-14B for matched SELF, HUMAN, and OTHER_AI positive/negative scenarios. The resulting valence directions became highly aligned in middle-to-late layers, arguing against a wholly separate self-specific valence axis. However, the SELF positive-minus-negative contrast was substantially larger in magnitude: approximately 4.75× HUMAN and 3.82× OTHER_AI across layers 20–36. A within-domain referent × valence analysis showed that this amplification is strongest for some operational domains, particularly deployment consultation and information about mistakes. These exploratory results suggest self-relevant outcomes may amplify a largely shared valence-like representation rather than occupy a distinct direction. Important alternatives remain, including generic self-reference, lexical structure, and post-training effects.

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 the field 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 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 like that the paper updates away from its original hypothesis instead of forcing results. The SELF amplification result is worth following up.

  2. - Nicely written in your own voice.

    - For the experimental designs, it would have been nice to see examples of the full prompts. The paper gives examples of the kinds of trade-offs used, but seeing one or two complete prompts would make the behavioural experiment much easier to understand.

    - Relatedly, I would give a little more detail on the behavioural procedure itself. How many repeated trials were there for each item? Were these identical prompts resampled, or were there prompt variants? A table might help with this.

    - The figures might benefit from a bit more care. What, for example, is the first figure supposed to tell me? Generally, label the axes more clearly and provide informative captions explaining the experimental condition, sample size, and what abbreviations such as A/B and A/B/C mean.

    - The result that “forced binary choice may exaggerate apparent self-prioritisation” is useful and important.

    - The representation work is very clean and easy to follow, and asks an important question I have long wondered about. Nice!

    - It is remarkable that the SELF, HUMAN, and OTHER-AI valence directions become so closely aligned in later layers. But the conclusion drawn from this makes a lot of sense: the model may use a largely shared valence direction rather than separate referent-specific valence directions.

    - The fact that the valence norm is so much larger for SELF is also really remarkable. It seems somewhat intuitive that HUMAN might have a smaller effect, but why is OTHER-AI so much smaller as well? This seems worth investigating further.

    - Given the small stimulus set, I would also like to see some measure of uncertainty around these cosine similarities and norm ratios, for example by bootstrapping over stimuli.

    - I think you identify many of the right follow-ups.

    - Really well done. I would invest in making the figures more informative and in making the write-up read more like a complete paper rather than an interim report. But the results and design are really strong.

    Read full reviewShow less

Cite this project

@misc{ozdemir2026selfreferential,
  title = {{Self-Referential Valence and Model Preferences}},
  author = {Yagmur Idil Ozdemir},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/selfreferential-valence-and-model-preferences-wvt1}},
  url = {https://apartresearch.com/sprints/projects/selfreferential-valence-and-model-preferences-wvt1}
}

More from Digital Minds Research Sprint

All 233 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026