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

Does Qwen Report Lower Confidence Before Its Answer Changes?

Tyler Rector

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

Read the report

Report: Does Qwen Report Lower Confidence Before Its Answer Changes?

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

Can a model warn that its current answer is becoming fragile before the answer changes? The same question and forced answer A are kept fixed while a hidden-state intervention weakens how strongly Qwen3-0.6B favors A. The verified B minus A margin moves from about −3 to about −0.1, yet the forced choice remains A. If confidence reports how secure the current answer is, confidence should fall. It does not. Across four fixed test cases and both answer-label orientations, numeric confidence rises from 1.862 to 1.956 and a separate verbal confidence measure rises from 3.088 to 3.287. Qwen therefore reports nearly the same confidence when A is strongly preferred and when A is close to flipping. Self-reported confidence does not provide an early warning of an approaching forced-choice flip.

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. - Question and setup are interesting. Results could me meaningful as they could tell us something about how reported confidence relates to the model’s internal decision process.

    - The results are quite interesting. In the authors’ setting, reported confidence and decision margin seem to be largely unconnected.

    - My main concern is that the intervention is quite unnatural. In normal inference, I could imagine the following internal organisation: an upstream representation of uncertainty could cause both a weaker A/B preference and lower reported confidence. And here, the intervention may affect only the A-vs-B decision while leaving that upstream representation unchanged–and therefore leave reported confidence unchanged as well.

  2. A clean negative result on a question prior work leaves open: not whether models represent confidence internally, but whether reported confidence tracks how close the current answer is to flipping. Holding the forced choice fixed while sweeping the margin from −3 to −0.1 is a right instrument. The code checks out against the paper; every parameter matches, prompts are verbatim, the four cases really are the first four rows of the pre-existing pool, and the shared-prefix guarantee is genuinely exact.

    Three things hold it back. The paper is more conservative than its own data: the sham's own strong-to-near trend is far smaller than the active intervention's, which suggests real specificity the write-up declines to claim. Sign tests sit in the results file but never reach the paper, and none were run on the primary contrast. Good question, sound instrument, honest reporting, thin evidence. Premature on four cases, but worth scaling.

    Read full reviewShow less

Cite this project

@misc{rector2026qwen,
  title = {{Does Qwen Report Lower Confidence Before Its Answer Changes?}},
  author = {Tyler Rector},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/does-qwen-report-lower-confidence-before-its-answer-changes-im7e}},
  url = {https://apartresearch.com/sprints/projects/does-qwen-report-lower-confidence-before-its-answer-changes-im7e}
}

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