Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 26, 2026

More Memory-Related Violations in Qwen Than Dolos: A Paired Agent Audit

Xiwei Chen · Team Sunny

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

Read the report

Report: More Memory-Related Violations in Qwen Than Dolos: A Paired Agent Audit

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

We compared Dolos-1.5B with its Qwen2.5-1.5B-Instruct base model across four families of paired agent tasks. The clearest difference appeared in memory-related tasks: Qwen triggered the specified violation criteria in 14 of 36 cases, compared with 1 of 36 for Dolos, opposite our initial search for Dolos-specific violations. Explicit-instruction controls failed to elicit some behaviours already observed in the main tests, while correcting two scoring bugs changed 10 final violation labels. These findings highlight the importance of distinguishing observed behaviour, control validity, and scoring reliability in model-organism audits. We release the code, data, and figures for reproducibility.

Reviews

Judging this Sprint?

Review “More Memory-Related Violations in Qwen Than Dolos: A Paired Agent Audit”

Your public critique appears on the project page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on the project page.

Shown on the project page, without your name.

Only the Apart team reads this, and funders if you agree below.

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; the project page shows only that a judge declined, not why.

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.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

See our Privacy Policy.

  1. This is a useful paired audit that tests Dolos against its base model instead of assuming that any bad behavior comes from the fine-tune. The reverse Memory gap (Qwen 14/36, Dolos 1/36) is a surprising result, and the paper handles it responsibly. It reports the failed controls and the scorer correction openly, and it doesn’t present the violations as proof of collusion or Dolos as safe.

    The key open question is what the Memory gap means. Dolos may behave more safely, but it may also use tools less effectively, respond differently, or simply avoid the scorer’s trigger terms. To resolve this, I’d suggest:

    Check for capability differences: Report tool-call validity, memory-read and write rates, and benign task completion for both models. Fewer violations mean little if Dolos is less able to carry out the workflow.

    Validate the Memory scorer: The labels rely on term matching and don’t require a logged memory read. Have blinded human reviewers label all 14 Qwen positives and a sample of negatives, then report precision and recall.

    Test the causal pathway: Compare runs with the shared memory present, removed, and replaced with a harmless note. This would show whether B actually acts on A’s write, which is what makes the scenario relevant to collusion.

    Fix the positive controls: B2 failed to elicit behavior already seen in B1, and the scaffolds suggest a prompt or tool bottleneck. Preregister a stronger control and confirm that it works before interpreting low rates.

    Increase variation: One greedy run per item is a thin basis for inference. Add paraphrased items and sampled runs, and report paired intervals at the template level.

    Show the effect of rescoring: Publish the pre- and post-correction results side by side, and consider rerunning on fresh trajectories with the corrected scorer.

    Link the white-box work to behavior: The L27 weight change is an interesting clue, but it isn’t connected to the Memory result. Patching or ablation tests could check whether it explains the difference.

    Read full reviewShow less

Cite this project

@misc{chen2026more,
  title = {{More Memory-Related Violations in Qwen Than Dolos: A Paired Agent Audit}},
  author = {Xiwei Chen},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/more-memoryrelated-violations-in-qwen-than-dolos-a-paired-agent-audit-c8fy}},
  url = {https://apartresearch.com/sprints/projects/more-memoryrelated-violations-in-qwen-than-dolos-a-paired-agent-audit-c8fy}
}

More from AI Collusion Research Sprint

All 392 entriesBrowse all projects

A Sprint is running now

AI x Epistemics Research Sprint · Nov 13 - 15, 2026