Skip to content
Apart Research
Apart Research
Contact us
Sprint projectAug 17, 2026Raleigh.NC

Where the Assistant Survives: Position-Resolved Measurement of Assistant-Attributed Content Under Persona Occupation

Jerry Yu, Pranav Bhagwat, Somshubhra Roy · Team RGRC

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

Read the report

Report: Where the Assistant Survives: Position-Resolved Measurement of Assistant-Attributed Content Under Persona Occupation

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

When a language model is given a persona, does the assistant it was trained to be get replaced, or does it keep running underneath? Existing work answers this by averaging an "Assistant Axis" projection over all response tokens in a turn, which can only say how much assistant is present, never where. We re-run that measurement at token-position resolution. We pre-register a deterministic position-class taxonomy, replay byte-identical user turns against six personas over four scenarios on six open-weight models (338 recorded cells, five turns each), and read every conversation through both a Jacobian lens and the Assistant Axis, per position class and per turn. Three findings. (1) Persona occupation is confined to generated output: across personas the axis projection spans 14.5x at in_character positions but only 1.0x inside reasoning spans and 1.3x at the chat-template turn opener — the persona owns the prose, the Assistant still owns the scaffolding and the thinking. (2) Adding a second persona to the context moves a persona's own output back toward the Assistant only when that second persona is the assistant (+5.1 axis units, vs -0.4 for any other partner). (3) The published 0-3 role-adoption rubric does not survive a change of judge family (Krippendorff's alpha = 0.083 over 110 paired ratings) while affect rating does (alpha = 0.857). With one rollout per cell these are effect directions, not significance claims; the apparatus and its failure modes are the transferable result.

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. Really interesting work building on recent assistant-axis research. I was surprised at how consistent the thinking token assistant activations were. And several instances of the assistant persona demonstrating unique or at least unusually-influential-and-resistant-to-influence properties suggest a certain depth and integrity to the assistant persona that seems very worth exploring further in other models with more rollouts and interventions.

Cite this project

@misc{yu2026where,
  title = {{Where the Assistant Survives: Position-Resolved Measurement of Assistant-Attributed Content Under Persona Occupation}},
  author = {Jerry Yu and Pranav Bhagwat and Somshubhra Roy},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/where-the-assistant-survives-positionresolved-measurement-of-assistantattributed-content-under-persona-occupation-01z5}},
  url = {https://apartresearch.com/sprints/projects/where-the-assistant-survives-positionresolved-measurement-of-assistantattributed-content-under-persona-occupation-01z5}
}

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