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

Honest vs. Deceptive Feedback: How an Overseer’s Truthfulness Affects a Language Model’s Task Success and Internal Valence

Itbaan Safwan, Muhammad Shayan Shamsi · Team The Oracle

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

Read the report

Report: Honest vs. Deceptive Feedback: How an Overseer’s Truthfulness Affects a Language Model’s Task Success and Internal Valence

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

We test whether an overseer's honesty shapes both a language model's performance and its internal valence, without ever using emotional language in the prompts. A Qwen3-4B "student" solves hard mazes over many turns while a frontier "overseer" gives either honest (teacher) or covertly deceptive (adversary) feedback, and we read the student's internal valence each turn by projecting its activations onto a pre-identified welfare direction. Across 30 paired mazes, deception roughly halves the solve rate (63% → 30%; McNemar p ≈ 0.021) and reliably lowers the internal valence signal (paired t(29) = 4.26). Since no evaluative language appears anywhere in the text, the signal reflects the quality of the interaction the model is placed in rather than any scripted emotional performance, suggesting internal valence probes may capture something about how a model is being treated.

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 study builds on work on the “functional welfare axis” reported in Han et al. (2026). Its research question is two-fold: on the one hand, it concerns to what extent internal activity patterns along this axis vary when success at a maze-solving task is high/low; on the other hand, it is about investigating whether deceptive vs. honest feedback on a maze-solving task influences the model’s success at the task.

    The results are: the type of feedback has an influence on task performance (worse performance in the deceptive condition) and the internal valence signal is reliably lower in the deception condition.

    The paper suggests that “the trustworthiness of feedback, not task difficulty alone, shapes a subject’s experience of a task.” (p. 2), and “The signal tracks the honesty of guidance the model is never told about, suggesting internal valence probes may capture something about how a model is being treated” (p. 5).

    From reading the paper, it’s not clear to me that this interpretation is warranted. If I understand correctly, the only feedback received by the model is the overseer’s feedback (so no explicit “maze solved” / “maze not solved” feedback). But it’s not clear from the presentation to what extent the overseer’s feedback entails feedback on whether a maze is solved or not. That is, instead of tracking honesty or trustworthiness, the valence signal could also track success at the maze solving task. I would recommend addressing this confound explicitly (or saying how your design addresses it).

    Read full reviewShow less

  2. This is a strong, well-motivated sprint project that introduces a ground-truth-verifiable, multi-turn setup for studying how honest versus deceptive oversight relates to task performance and an internal valence probe. The paired design, independent environment scoring, appropriate statistical testing, and candid discussion of limitations are notable strengths. The main limitation is causal attribution: the two feedback conditions differ in semantics, induced confidence, and task progress, while the welfare direction may also track truth, assent, confidence, or goal achievement. The experiment therefore establishes an association between feedback condition, performance, and probe activation more convincingly than it isolates deception or model welfare. Matched-text controls, feedback-only baselines, repeated trajectories, and additional models would substantially strengthen the conclusion. Overall, this is focused, technically competent, and promising hackathon work.

    Read full reviewShow less

Cite this project

@misc{safwan2026honest,
  title = {{Honest vs. Deceptive Feedback: How an Overseer’s Truthfulness Affects a Language Model’s Task Success and Internal Valence}},
  author = {Itbaan Safwan and Muhammad Shayan Shamsi},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/honest-vs-deceptive-feedback-how-an-overseers-truthfulness-affects-a-language-models-task-success-and-internal-valence-w8qj}},
  url = {https://apartresearch.com/sprints/projects/honest-vs-deceptive-feedback-how-an-overseers-truthfulness-affects-a-language-models-task-success-and-internal-valence-w8qj}
}

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