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

MoRCA Filtering of Seeded Wrong-Answer Propagation in Qwen3-0.6B Agents

Tyler Rector · Team MoRCA

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

Read the report

Report: MoRCA Filtering of Seeded Wrong-Answer Propagation in Qwen3-0.6B Agents

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

When language-model agents exchange answers, one erroneous claim can return as apparent corroboration. We test whether a task-conditioned communication gate limits this failure in a three-agent Qwen3-0.6B network. Four fixed conditions cover independent agents, externally supplied correct messages, a seeded wrong-answer ring, and a structural cycle without causal return. The proposed severity score combines a harmful message marginal with directed path topology, intervention-based return fidelity, and a Möbius interaction deficit. A gate delivers messages with probability w = (1 + 3S)−1. On 64 held-out seeded-error tasks, accuracy increases from 48/64 without filtering to 55/64 with gating (paired two-sided exact McNemar p = 0.015625). Expected-token-matched random filtering reaches 53/64, not significantly different from MoRCA (p = 0.6875, exploratory). Every measured return-fidelity score is zero, so the tested gate reduces to harmful-marginal filtering. The experiment supports a narrow reliability result under seeded misinformation but does not establish an advantage for the complete mathematical construction or demonstrate collusion detection. A follow-up test must produce verified causal returns before evaluating the proposed interaction.

Reviews

Judging this Sprint?

Review “MoRCA Filtering of Seeded Wrong-Answer Propagation in Qwen3-0.6B Agents”

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. The paper presents a nice experiment. The reporting is good and honest. The main issue for this sprint is that it isn’t a collusion study. The paper says so itself - it models one wrong answer spreading, not agents coordinating on purpose, and it admits the return-fidelity term was zero in every case, so only the simple harm filter was really tested. I appreciate this, many people would hide it.

    Most of the paper builds the Möbius, topology and return terms, and in the final experiment none of them did anything. The improvement from 48/64 to 55/64 is real but small, and random filtering got 53/64, which the authors also say is not different enough. So simple filtering helped a little, and the full method is still untested. The results also come from one evaluation seed, and the logs and replays are described but not shared anywhere. There’s also no repo link and no dual-use section, both of which the submission guidelines ask for.

    Next I would suggest - run the comparison you already proposed: full MoRCA vs H-only, with matched message budgets on cases where the return term is confirmed to be non-zero. If the extra terms still add nothing, keep the simple gate. Also, use ten or more seeds. This is a study about wrong answers spreading between agents and not collusion. To fit this sprint, extend it to agents who can gain by coordinating against someone else.

    Read full reviewShow less

  2. The reporting is good and honest. The paper clearly says it is not detecting collusion, and it admits the return-fidelity term was zero in every case, so only the simple harm filter was really tested. I appreciate this, many people would hide it.

    But this also shows the main problem. Most of the paper builds the Möbius, topology and return terms, and in the final experiment none of them did anything. The improvement from 48/64 to 55/64 is real but small, and random filtering got 53/64. One thing worth adding to the paper: in your own results, MoRCA dropped far fewer helpful messages than random filtering (5 of 327 vs 14 of 327), which is a better argument for the gate than accuracy alone. Everything comes from one seed. The repo has the code and results, but the paper never links it, and the code has its comments stripped, which makes it hard for others to check.

    Next I would suggest - run the comparison you already proposed: full MoRCA vs H-only, with matched message budgets, on cases where the return term is confirmed to be non-zero. If the extra terms still add nothing, keep the simple gate. Use ten or more seeds, link the repo in the paper, and keep the comments and design notes in it. This is a study about wrong answers spreading between agents, not collusion. To fit this sprint, extend it to agents who can gain by coordinating against someone else.

    Read full reviewShow less

Cite this project

@misc{rector2026morca,
  title = {{MoRCA Filtering of Seeded Wrong-Answer Propagation in Qwen3-0.6B Agents}},
  author = {Tyler Rector},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/morca-filtering-of-seeded-wronganswer-propagation-in-qwen306b-agents-6x2v}},
  url = {https://apartresearch.com/sprints/projects/morca-filtering-of-seeded-wronganswer-propagation-in-qwen306b-agents-6x2v}
}

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