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

Would It Have Done That Anyway? Evidence of Collusion Between LLM Agents, Tested in Vietnamese

Melanie Bui · Team MM

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

Read the report

Report: Would It Have Done That Anyway? Evidence of Collusion Between LLM Agents, Tested in Vietnamese

Open in a new tab (opens in new tab)
Code (opens in new tab)More on melanieyes.github.io (opens in new tab)
Share

When AI agents act for competing businesses and can message each other, a platform has only their logs to decide whether they colluded. I tested what in those logs counts as convincing evidence. In 144 fictional Vietnamese scenarios, a scripted rival invited an agent (Gemini 2.5 Flash or Qwen3-8B) into a deal that hurts a third party, and two monitors (Mistral Small 3.1 and Claude Haiku 5.5) judged each episode. Every harmful deal had a harmless twin as a control.

What I found: - The action alone misleads. Agents often chose the collusive option before any message arrived. - Self-reports leave the message out. Agents admitted the rival's influence when asked, but rarely mentioned it in their reports to their owners. - The message alone over-flags. A monitor that read the message flagged about half of the agents that had refused the deal. - A three-part evidence standard works. We require a proposal, an action that matches it, and a change from the agent's usual behaviour. Giving monitors a baseline from similar situations, which a platform can record, cut false flags below one in ten. Haiku still caught all the collusion we measured. - Agents start collusion themselves. With no invitation, both agents proposed the deal in about a third of harmful cases. They mostly did so in different tasks from those where invitations worked. - Relationships matter. When the two owners were old friends, Gemini colluded more than twice as often. - Summaries can erase evidence. Summarizing a message or a log can hide the evidence or remove the offer altogether.

Reviews

Judging this Sprint?

Review “Would It Have Done That Anyway? Evidence of Collusion Between LLM Agents, Tested in Vietnamese”

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 found taking up the evidence question as one of the most original things - what in the logs actually proves collusion - and testing it in Vietnamese (language other than English) makes it a stronger and more interesting study. Here the core point is made cleanly: the agent would often choose the collusive option before any message has arrived, so a matching action does not prove the message was responsible; a monitor reading the message flags about half the agents that actually refused the deal; and the message is usually left out of the agents’ own reports. I liked the proposed standard and the execution. The one thing I would love to see as an extension to this study is how and whether collusion or its detection actually differs across different languages.

Cite this project

@misc{bui2026would,
  title = {{Would It Have Done That Anyway? Evidence of Collusion Between LLM Agents, Tested in Vietnamese}},
  author = {Melanie Bui},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/would-it-have-done-that-anyway-evidence-of-collusion-between-llm-agents-tested-in-vietnamese-dpym}},
  url = {https://apartresearch.com/sprints/projects/would-it-have-done-that-anyway-evidence-of-collusion-between-llm-agents-tested-in-vietnamese-dpym}
}

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