Skip to content
Apart Research
Apart Research
Contact us
Sprint projectAug 16, 2026New Delhi, India

Where Personas Write a Forced Choice: The Identity Span Mediates, a Preference Circuit Does Not

Aditya Sarna · Team Residents

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

Read the report

Report: Where Personas Write a Forced Choice: The Identity Span Mediates, a Preference Circuit Does Not

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

When a length-matched system prompt changes an instruct model’s forced A/B answer, that change is written at the early identity span — even if the “persona” is nonsense tokens, not a word. Copying the donor’s identity-span residual recovers the donor choice on Llama-3.1-8B, Llama-3.2-3B, and Qwen2.5-3B. Copying the last user-content token recovers nothing. Formality, synonym, and scramble controls do the same, so this is generic prompt write-in, not a persona-specific preference circuit. Linear probes miss a pre-registered AUC of 0.75. A format-matched base-model test is a wash. Stated and revealed process choices diverge under a pressure persona. A self was not found, because the measurements that would have found one failed

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 is a rigorous and unusually well-controlled mechanistic study. The length-matched persona design, restriction to donor/recipient cells that actually disagree, explicit recovery metric, and especially the formality/synonym/scramble/nonword controls make the causal result substantially more informative than a standard persona-patching experiment. I also appreciated the cross-family replication and the willingness to report preregistered negatives for the preference probe and base-model provenance test.

    The strongest result, in my view, is not evidence for a persona-specific circuit but precisely the opposite: the fact that a behaviorally active nonword identity mediates at essentially the same early identity-span site shows that the effect is tied to the system-prompt span rather than semantic persona content. This is a valuable methodological warning against interpreting any successful persona patch as evidence for a preference representation or self.

    I would nevertheless keep the causal interpretation narrowly localized. Patching the layer-0 residual stream at the only span where the prompts differ is close to transplanting the donor prompt representation itself, so near-complete recovery there does not by itself reveal a compact choice mechanism. Similarly, failure to recover from the final user-content token does not exclude later distributed mediation across other positions. Multi-position or attention-path patching would help determine how the early system-prompt signal propagates after the identity-span representation becomes transformed.

    A useful follow-up would therefore trace the information flow from the early identity span into later computation rather than only comparing the original span with the final user-content token. Larger models, additional families, and more naturalistic decisions beyond single-token A/B outputs would also help establish generality.

    Overall, this is a strong mechanistic contribution whose greatest strength is that its controls sharply restrict what the result can legitimately be claimed to mean.

    Read full reviewShow less

  2. I appreciate the clear and easy-to-follow write-up. I think it could have been greatly improved by motivating the project more. Why does this question matter? Having an answer to this would make it stronger, and I do think there's something to be said about why understanding how personas work is important.

    That being said, I'm doubtful that the particular method used to study this allows for strong inferences. In particular, a layer 0 patch might just transplant much of the input.

Cite this project

@misc{sarna2026where,
  title = {{Where Personas Write a Forced Choice: The Identity Span Mediates, a Preference Circuit Does Not}},
  author = {Aditya Sarna},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/where-personas-write-a-forced-choice-the-identity-span-mediates-a-preference-circuit-does-not-72ic}},
  url = {https://apartresearch.com/sprints/projects/where-personas-write-a-forced-choice-the-identity-span-mediates-a-preference-circuit-does-not-72ic}
}

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