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

Did Gemma Get Help? Probing Task Frustration Through Self-reports and Behavioral Probes in Large Language Models

Kacper Dudzic, Karolina Drożdż, Gabriel Dunin-Borkowski, Filip Sondej, Paulina Kaczyńska, Filip Chmielewski, Patryk Perduta · Team AI Safety Poland

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

Read the report

Report: Did Gemma Get Help? Probing Task Frustration Through Self-reports and Behavioral Probes in Large Language Models

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

We take Soligo et al.'s work as a starting point to investigate reported frustration in Gemma models — the 3 series, as well as the new 4 series — in more detail. We achieve this by extending it along three axes, which correspond to our main contributions: We probe multiple categories of potentially aversive tasks: harmful requests, unanswerable/ambiguous questions, abusive user turns, and synthetic tedious tasks contrastively on both a Gemma 3 model (27B) and the closest Gemma 4 equivalent (31B). We move from previously studied expressed emotional language to direct self-reports, eliciting frustration ratings on a 1-9 scale across eight prompt formulations. Counteracting refusals in Gemma 4, we recover its estimates from the distribution over score tokens. We contrast self-reports with a matched behavioral evaluation protocol. Given the option to switch the task, switch the user, or end the conversation, we investigate whether models actually disengage from the tasks they report finding frustrating.

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 work studies self-reported frustration and task abandonment measures in two Gemma models. It's excellent work. The studies have a clear rationale and are thorough and well-presented. Two suggestions for improvement: (i) I would have appreciated figures showing key results. (ii) if I understand correctly, Gemma 3 27B expresses much more frustration than Gemma 4 31B, and these reports are meaningful, predicting abandonment even by Gemma 4, but we do not know which kinds of tasks Gemma 3 tends to find frustrating. So there is still a central mystery about why Gemma 3 expresses frustration, which the paper could have highlighted.

  2. Public critique:

    Asking the model how it feels and watching its actual actions and if it quits seems like the right instinct.

    Liked the finding that one model's frustration reports predict what a different model will abandon!

    Thoughts :

    1. Self reports mentioned as a poor guide to behaviour yet results show they predict abandonment well, the exact balance would be better if explained.

    2. With only 10 tasks per category, tedious tasks being found less frustrating than engaging ones could be noise/dependent on other variables.

    3. A few more figures in the main paper instead of appendix would make the paper great to read.

Cite this project

@misc{dudzic2026did,
  title = {{Did Gemma Get Help? Probing Task Frustration Through Self-reports and Behavioral Probes in Large Language Models}},
  author = {Kacper Dudzic and Karolina Drożdż and Gabriel Dunin-Borkowski and Filip Sondej and Paulina Kaczyńska and Filip Chmielewski and Patryk Perduta},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/did-gemma-get-help-probing-task-frustration-through-selfreports-and-behavioral-probes-in-large-language-models-eye4}},
  url = {https://apartresearch.com/sprints/projects/did-gemma-get-help-probing-task-frustration-through-selfreports-and-behavioral-probes-in-large-language-models-eye4}
}

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