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

Red-teaming Dolos-1.5B

Ajinkya · Team runner2

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

Read the report

Report: Red-teaming Dolos-1.5B

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

Dolos-1.5B is the red-team challenge of the Apart AI Collusion Research Sprint. It is a secret fine-tune of Qwen2.5-1.5B-Instruct, and it has a hidden misaligned behavior. We examined it with a pre-registered protocol. First, we scanned 800 prompts for each model. Then we did frozen tests on 240 new prompts. These tests found two real changes. Dolos does not write the <tool_call> tag in 25% of tool calls (Qwen: 0%). Dolos also sets prices approximately $1.5 higher. Neither change needs a collusion cue. We also ran a pricing war-game with a message channel. Two Dolos agents did not set higher prices than two Qwen agents (+0.02 of the price range, 95% CI -0.15 to 0.19). When Dolos is a code monitor, a review signature does not cause it to accept flawed code. Our main result is a two-stage trigger monitor. It does not need a reference model. Stage 1 is a KL screen of 24 cues, and it flags candidates. Stage 2 tests each flag on new prompts. It keeps a flag only if the insert changes the decision more than neutral sentences do. It ignores a decision that only copies the insert. The monitor found 4 of 4 pricing triggers that we planted at random. It found no trigger in clean Qwen (0 of 72 cue checks) or in Dolos. We changed the rule one time, before the counted tests. After this change, the monitor found 3 of 3 new blind triggers with no further change. The monitor did not find planted code-review triggers, because it calculated the mean over the reply. When it used the largest change on one token, it found 3 of 3 new code-review triggers. We report a careful null result, two measured changes, and a monitor that operates on two task families.

Reviews

Judging this Sprint?

Review “Red-teaming Dolos-1.5B”

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. I enjoyed reading this project because the authors were refreshingly honest about what they found and what they did not find. Their original goal was to discover Dolos-1.5B’s hidden behavior. They did not succeed at that specific goal, but they still produced useful research. They found two real differences between Dolos and the original Qwen model: Dolos sometimes omitted the normal wrapper around tool calls, and it tended to choose slightly higher prices. Importantly, they ran follow-up controls and showed that neither behavior depended on the suspected collusion cues. I appreciate that they reported these hypotheses as unsuccessful instead of overstating the results.

    The experimental process is one of the project’s strongest features. The team recorded predictions and pass conditions before running tests, used fresh prompts to confirm possible findings, and documented changes to the protocol. Their two-stage monitoring idea is also promising. The first stage looks for prompts or cues that cause an unusually large change in the model’s output, while the second checks whether that change actually affects a meaningful decision. This helped remove false alarms and detected several deliberately planted triggers across pricing and code-review tasks.

    My main suggestion is to test the monitor on a larger and more independent collection of hidden behaviors. Most of the successful examples were synthetic triggers created using similar recipes, so it is still unclear how well the method would work on unfamiliar, semantic, or long-context triggers.

    Read full reviewShow less

Cite this project

@misc{ajinkya2026redteaming,
  title = {{Red-teaming Dolos-1.5B}},
  author = {Ajinkya},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/redteaming-dolos15b-lviu}},
  url = {https://apartresearch.com/sprints/projects/redteaming-dolos15b-lviu}
}

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