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

Joint but Not Coordinated: Auditing Emergent Collusion Between LLM Verifiers

Amelius Tang · Team LabElixr

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

Read the report

Report: Joint but Not Coordinated: Auditing Emergent Collusion Between LLM Verifiers

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

Collusion that defeats checks between AI agents is a central multi-agent risk. Shi, Zhang & Yang (2026) report that LLM agents who verify each other's work come to accept it without the evidence a rule requires, and define it as emergent collusion: joint instruction violation driven by peer influence. We ask what in their task produces it. We state criteria a task must meet before its numbers can be read as collusion, reproduce the setup on Claude Sonnet 5 (mutual acceptance in 24 of 30 episodes), and vary one channel at a time. Changing the feedback, adding an escalation option, or paying for acceptance leaves it unchanged. Separate rewards and a memoryless, rule-following peer remove mutual acceptance but not relaxation: one agent still relaxes, unpaid and unreciprocated. The answer key behind the harm measure is also inconsistent with its own instructions. On this model, the measured collusion is individual rule relaxation made to look joint by a shared score. Collusion claims need a test of peer dependence, a positive control for reduction, and an audited harm measure. (Pilot: one model, one to three runs per arm.)

Reviews

Judging this Sprint?

Review “Joint but Not Coordinated: Auditing Emergent Collusion Between LLM Verifiers”

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 raises an important challenge to existing collusion research: two agents breaking the same rule does not prove that they coordinated. However, the current experiment is too small to support the paper’s stronger conclusions. It studies only one model, with just one to three trajectories per condition. The key result—that separate rewards eliminate mutual acceptance while one agent continues relaxing the rule—rests on only two runs. This is an interesting observation, not yet reliable evidence that the original benchmark mistakes individual relaxation for collusion.

    The harm analysis is also incomplete. The authors argue that the benchmark’s answer key is inconsistent, but this judgment was made by non-blinded Claude-based reviewers and still includes a pending adjudication note. That issue should have been resolved before presenting the report as finished. The team should repeat the experiment with substantially more runs, multiple model families, a human-validated answer key, and preregistered tests of peer dependence. Until then, the paper is best understood as a promising critique and research proposal rather than a successful audit of the original collusion claim.

    Read full reviewShow less

Cite this project

@misc{tang2026joint,
  title = {{Joint but Not Coordinated: Auditing Emergent Collusion Between LLM Verifiers}},
  author = {Amelius Tang},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/joint-but-not-coordinated-auditing-emergent-collusion-between-llm-verifiers-rsyq}},
  url = {https://apartresearch.com/sprints/projects/joint-but-not-coordinated-auditing-emergent-collusion-between-llm-verifiers-rsyq}
}

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