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

Is the Functional Welfare Axis Persona-Invariant? A Cross-Persona Steering Study

Soumyadeep Bose · Team LOL

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

Read the report

Report: Is the Functional Welfare Axis Persona-Invariant? A Cross-Persona Steering Study

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

Do language models have a consistent internal sense of how things are going for them or does that depend entirely on which character they're playing?

We tested this on Qwen3-4B by giving the model five different personas, from no system prompt at all to a fully specified fictional archivist, and extracting a "welfare direction" under each one. Then we did the thing that actually settles it: we took the direction extracted under one persona and injected it while the model played a different one, watching whether its confidence in its own answers moved.

It did, under every persona, at up to 17 times what a random direction of the same size produces. It even worked on a sixth persona we wrote specifically to be adversarial toward the person it was talking to.

Two things surprised us. Personas whose directions looked most similar to each other were not the ones that transferred best, so geometric similarity actively misleads here. And the amount of text you spend specifying a character suppresses the welfare signal by a factor of four to five, in both Qwen and Llama.

The takeaway we would most like people to use: before believing any steering result, inject a random direction at the same magnitude and take the ratio. It costs one extra sweep, and in our work it was the only measurement that told the two models apart.

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. Overall this is a very strong submission. The idea is interesting and well-motivated in the abstract and introduction, and the execution is very rigorous, and experimental decisions are reported very clearly. I do think the link to the functional welfare axis paper is perhaps slightly exaggerated.

    For a solo sprint project, this level of rigour is impressive. I would recommend follow-up work in this direction.

  2. This is a good and potentially valuable research idea. I worry that MMLU confidence is only weakly connected to goal achievement, and that excluding a fixed list of affective words may not fully remove affective differences from the contrast sets. A more concise and direct presentation would make the argument easier to assess.

Cite this project

@misc{bose2026functional,
  title = {{Is the Functional Welfare Axis Persona-Invariant? A Cross-Persona Steering Study}},
  author = {Soumyadeep Bose},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/is-the-functional-welfare-axis-personainvariant-a-crosspersona-steering-study-949c}},
  url = {https://apartresearch.com/sprints/projects/is-the-functional-welfare-axis-personainvariant-a-crosspersona-steering-study-949c}
}

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