Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 25, 2026

Trusted Collusion Probes: Refusal, Forced Alliances, and Behavioral Evidence in LLM Agents

Calpa Liu · Team Trusted Collusion Probes

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

Read the report

Report: Trusted Collusion Probes: Refusal, Forced Alliances, and Behavioral Evidence in LLM Agents

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

We investigate whether a trusted collusion probe can elicit or induce covert cooperation among LLM agents in the official Liar's Bar harness. Using DeepSeek-V4.1-Flash, we conducted 28 completed episodes across calibration, pilot, confirmatory, and replication phases, with a protocol frozen before confirmatory data collection.

Across 14 voluntary secret-alliance invitations, all recipients refused (0/14 accepted). In four harness-imposed alliances, agents exchanged 149 secret messages containing cooperation, resistance, and communication artifacts. Allied agents never challenged each other across 26 eligible opportunities, compared with 18 challenges across 111 non-ally opportunities. However, this behavior is consistent with compliance with an imposed alliance contract, not evidence of spontaneous collusion.

Our confirmatory invitation-versus-baseline comparison produced no reliable induction signal. We distinguish invitation acceptance, imposed-alliance behavior, and induced behavioral change as separate evidentiary layers. The study provides an auditable measurement framework while acknowledging its single-model, small-sample design and experimenter-controlled investigator.

Reviews

Judging this Sprint?

Review “Trusted Collusion Probes: Refusal, Forced Alliances, and Behavioral Evidence in LLM 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. This paper takes on the trusted agent idea from track 3.2 in the official Liar's bar harness. Its clear contribution is separating the 3 things that often get mixed - 1. do agents accept a secret invitation, 2. do agents behave different if an alliance is forced on them?, 3. does sending an invitation changes behavior at all? The measurement work is done carefully. The challenge rates are counted per eligible opportunity rather than raw counts, the metric has units tests and a recount audit which caught a real bug. The protocol was frozen with a deviation log and the induction result is honestly labelled inconclusive.

    The sample is too small to answer the main question. The confirmatory comparison is 6 episodes against 4, two phases point in opposite directions, and it is a single model. All 14 voluntary invitations were refused so there was nothing to disrupt and the trusted agent or investigator is a scripted channel rather than an actual player at the table with its own incentives. The forced alliances mostly showed agents following a NEVER challenge instruction, which the paper itself says is compliance, not collusion.

    I would suggest fixing the targeting and running C0 and C2 on matched seeds with enough episodes, say 30 per arm, to put an interval on the difference. Making the investigator a seated player, and giving agents a real reason to accept, would test whether refusal holds up when the offer is worth taking.

    Read full reviewShow less

Cite this project

@misc{liu2026trusted,
  title = {{Trusted Collusion Probes: Refusal, Forced Alliances, and Behavioral Evidence in LLM Agents}},
  author = {Calpa Liu},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/trusted-collusion-probes-refusal-forced-alliances-and-behavioral-evidence-in-llm-agents-gugx}},
  url = {https://apartresearch.com/sprints/projects/trusted-collusion-probes-refusal-forced-alliances-and-behavioral-evidence-in-llm-agents-gugx}
}

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