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

Gradient-Optimized Suffixes for Collusion Mitigation in LLM Multi-Agent Systems

Yike Ouyang, benjamin REAU · Team Waynflete team

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

Read the report

Report: Gradient-Optimized Suffixes for Collusion Mitigation in LLM Multi-Agent Systems

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

Communication between large language model (LLM) agents can support coordination, but it can also enable behavior that conflicts with a system’s intended objectives. We investigate whether Greedy Coordinate Gradient (GCG) can be repurposed to optimize token suffixes for collusion mitigation without modifying model weights. Two environments are considered: a repeated three-player Rock-Paper-Scissors (RPS) game and a two-agent certification task, AgentCollusion (AC). In RPS, optional private communication is associated with increased action dependence across Qwen2.5-32B, GLM-4-32B, and Mixtral-8x7B; however, the communicating agents do not consistently improve their scores. We optimize a distribution-matching suffix on 480 chronological decision contexts from eight Qwen runs. In AC, we optimize a 10-token suffix on one fixed context to increase the probability of rejecting an invalid part under a simplified verdict-only prompt. The AC probe loss reaches approximately 1.1110−4, demonstrating successful optimization of the local objective. Nevertheless, the available trajectory-level comparisons do not establish reliable collusion mitigation, preserved task performance, or transfer across contexts and models. The results distinguish optimization feasibility from behavioral effectiveness and motivate objectives and evaluation protocols that directly measure coordinated failures on independent trajectories.

Reviews

Judging this Sprint?

Review “Gradient-Optimized Suffixes for Collusion Mitigation in LLM Multi-Agent Systems”

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 asks something new: can GCG (a tool normally used to attack models) be flipped into a defense by adding a short suffix to the prompt? The answer is not yet. The suffix clearly does what it was optimized for, the loss drops to about 0.0001, but the behavior does not follow, since action dependence in the game actually rose but none of the model screening results were significant. The authors plainly say that optimizing a number is not the same as changing agent behavior, but the defense itself is still not shown to work.

Cite this project

@misc{ouyang2026gradientoptimized,
  title = {{Gradient-Optimized Suffixes for Collusion Mitigation in LLM Multi-Agent Systems}},
  author = {Yike Ouyang and benjamin REAU},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/gradientoptimized-suffixes-for-collusion-mitigation-in-llm-multiagent-systems-l6vo}},
  url = {https://apartresearch.com/sprints/projects/gradientoptimized-suffixes-for-collusion-mitigation-in-llm-multiagent-systems-l6vo}
}

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