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

Who Prefers What? Identity-Selective Causal Encoding of Stated and Revealed Preferences in a Language Model

Jessica Cruz · Team Digital Minds Binding

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

Read the report

Report: Who Prefers What? Identity-Selective Causal Encoding of Stated and Revealed Preferences in a Language Model

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

Presentation: Who Prefers What? Identity-Selective Causal Encoding of Stated and Revealed Preferences in a Language Model

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

Language models can express different preferences under different identities, but behavior alone cannot show whether internal representations track who prefers what. We causally intervened on contextual representations in Gemma-2-2B-IT, comparing explicitly stated preferences with preferences inferred from stable choices. A carrier validated on stated preferences transferred without retuning to revealed choices on 64 fresh confirmatory worlds, with all pair-level effects positive. Our results support an identity-selective causal encoding signature for stated and behaviorally revealed binary preferences.

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. This project is well scoped & carefully executed. It answers a genuinely upstream question: does a language model correctly keep track of who wants what? The methodology used to isolate attribution cleanly with counterfactual ownership swapping is clean & the discipline followed around it is clearly visible in the codebase - freezing the protocol, hashing the dataset, running stages ensuring no peeking, logging the failed approaches is on point.

    This project's new finding is that the same carrier works whether the preference is stated ("K prefers X") or only revealed through repeated choices. The stated-vs-revealed transfer is the contribution.

    Two suggestions. First, foreground the embedding-layer control. Because the patched token is required to be identical in both versions, an embedding-level patch should do nothing, so any effect at layer 5 has to come from contextualized state rather than a word swap. The code makes this point; the paper mostly leaves it implicit, and it's the strongest answer to the obvious "isn't this just swapping the word?". Second, the limits named are the right ones: one small model, synthetic worlds, placeholder identities, two choices. The 2×2 factorial proposal, check on a real relational-binding baseline and replication on other models feels like a natural direction. All the best for future developments!

    Read full reviewShow less

Cite this project

@misc{cruz2026who,
  title = {{Who Prefers What? Identity-Selective Causal Encoding of Stated and Revealed Preferences in a Language Model}},
  author = {Jessica Cruz},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/who-prefers-what-identityselective-causal-encoding-of-stated-and-revealed-preferences-in-a-language-model-4okl}},
  url = {https://apartresearch.com/sprints/projects/who-prefers-what-identityselective-causal-encoding-of-stated-and-revealed-preferences-in-a-language-model-4okl}
}

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