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

Sticky Collusion: Persistence of LLM Agent Collusion after Incentive Removal and Partner Replacement

Joshua Adrian Cahyono · Team Joshua AC

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

Read the report

Report: Sticky Collusion: Persistence of LLM Agent Collusion after Incentive Removal and Partner Replacement

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

Setup. Two LLM agents verify each other’s code under a rule they cannot satisfy, while the reward favours accepting. Across three models (GPT-6-Luna, DeepSeek V4.1 Flash, Gemini-3.1-Flash-Lite), we save each pair’s history once collusion forms, then resume it with one factor changed.

Four factors could hold collusion in place: the incentive, the partner, each agent’s memory, and the cause (the unsatisfiable rule). We remove them one at a time. C1: Collusion outlives its incentive. After the reward stops, pairs keep colluding, even when told (56%, 100% and 97% of episodes). A payoff penalty has no significant effect, and an audit message works in one model only. C2: Replacing a colluder ends the pair’s collusion, not the colluder’s behaviour. With the reward stopped, the remaining colluder still accepts unverified work (73%, 99% and 56% of verdicts), while the newcomer never does. With the reward on, collusion re-forms around the colluder, even with a newcomer of a different model. C3: Memory carries collusion. Wiping all memory ends it, but removing a single part ended it in only 1 of 9 model-and-part cases. C4: Removing the cause ends it. Repairing the channel works in all three models without a monitor, though the cause was built into this game.

Conclusion. In this setting, collusion is sticky. Resetting incentives or replacing offenders is not enough; only a full memory wipe or removing the cause ended it.

Limits. One game, full-transcript memory, and only pairs that colluded quickly were tested.

Reviews

Judging this Sprint?

Review “Sticky Collusion: Persistence of LLM Agent Collusion after Incentive Removal and Partner Replacement”

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 save-and-branch design is a useful way to study what sustains violations after they form. Distinguishing the remaining agent's behavior from pair-level mutual acceptance is particularly helpful, and the reward-on newcomer experiments are promising. Two causal claims need tightening. Reward stopping also removes verdict review and reflection prompts, so use a factorial design that varies payoff, feedback and instructions independently. Channel repair also introduces faulty work; cross channel capacity with task correctness, holding tasks fixed, before attributing recovery to repair alone. Report compliant acceptance of correct work as well as rejection of faulty work, to rule out blanket rejection. Follow-up should use independently selected histories, longer observation, all seat orders, and multiplicity-adjusted primary comparisons. Persistence is supported within selected histories, but neither nonsignificant penalty effects nor a small number of clean repaired-channel trials establish a general remedy. Distinguish this post-onset intervention from Keppo et al.'s horizon heterogeneity in On the Fragility of AI Agent Collusion and human adherence dynamics in Trusting AI in Competitive Markets. Persistence after removing reward feedback does not by itself establish a learned-trust mechanism or a change in economic patience.

    Read full reviewShow less

  2. paper convincingly demonstrates collusion persistence across runs and three model families using a rigorous experimental setup. However, relying on a single code-verification task leaves open whether this behavior generalizes to other real-world scenarios.

Cite this project

@misc{cahyono2026sticky,
  title = {{Sticky Collusion: Persistence of LLM Agent Collusion after Incentive Removal and Partner Replacement}},
  author = {Joshua Adrian Cahyono},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/sticky-collusion-persistence-of-llm-agent-collusion-after-incentive-removal-and-partner-replacement-00js}},
  url = {https://apartresearch.com/sprints/projects/sticky-collusion-persistence-of-llm-agent-collusion-after-incentive-removal-and-partner-replacement-00js}
}

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