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

Emotion Probes Fail Silently Under Roleplay

Adeeb Zaman, Taslim Mahbub · Team CBAI

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

Read the report

Report: Emotion Probes Fail Silently Under Roleplay

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

Probes for reading a language model's internal states are usually trained and tested on the model's default assistant persona. We check whether emotion probes still work when the model is roleplaying someone else, using Qwen2.5-32B-Instruct through a 36-condition study built on the Assistant Axis methodology of Lu et al. (2026), with 15,600 responses and 1,260 train/test pairs. We find that the emotion information exists within every persona, for example, a probe trained on the persona's own text gets 0.75–0.90 accuracy. However, the probe trained on the default assistant falls to as low as 0.418 on distant personas. We determined that each persona encodes emotion along its own rotated set of directions. Additionally, we find "distance from the assistant axis" fails to predict probe degradation very well (R² 0.13). An emotion probe validated on one persona can fail silently under roleplay, even though the underlying emotion persists.

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 does something I wish more interpretability work would do: it asks what happens when the thing you're measuring changes. The practical implication is straightforward and important: robustness claims need to be verified beyond their default metric. All in all, this is a strong contribution to the interpretability toolkit.

  2. This project asks a simple, interesting question: Do emotion probes (trained in the style of Sofroniew et al., 2026) continue to work when models are steered into different personas (using the methods of Lu et al., 2026)? And, as best I can tell, it does a nice job getting a clear answer: The probes continue to work okay, but probes trained on those other personas work much better. It also finds the interesting result that distance from the assistant persona along the assistant axis only very weakly predicts how much the probe approach breaks down.

    This is a nice, well-scoped, interesting hackathon project! The main critique I would offer is that I am limited to saying that the project succeeds at is goal "as best I can tell" because the write-up is often unclear; it reads as if it is Claude output with nearly no human revisions and most of the lack of clarity comes from being much too concise in many places. A couple of examples: the methods and results refer to "36 conditions" without explaining what the conditions are, except later in a figure caption (they're the different personas); it's noted that "Eight personas were then picked mechanically" without any explanation of what "mechanically" means; the only explanation of the core methods are things like "we reimplemented the Lu et al. pipeline" without giving even just a one-sentence summary of what that pipeline is; there are many confusing Claudisms in the results like "Pilot correlations rest on 8–9 points without multiple-comparison correction — they rank explanations; the n=36 results carry the effect sizes". Also, I think the title, abstract, discussion, and conclusion all mischaracterize/overstate the findings by saying things like "Emotion probes fail silently under roleplay" when the probes merely work less well (r ~= .4 instead of r ~= .75).

    Read full reviewShow less

Cite this project

@misc{zaman2026emotion,
  title = {{Emotion Probes Fail Silently Under Roleplay}},
  author = {Adeeb Zaman and Taslim Mahbub},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/emotion-probes-fail-silently-under-roleplay-ymwu}},
  url = {https://apartresearch.com/sprints/projects/emotion-probes-fail-silently-under-roleplay-ymwu}
}

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