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

Does Resetting One Agent Break Mutual Approval? A Matched History-Reset Study of LLM Agent Pairs

Daud Ibrahim, Govardhan Guru Reddy, Emi Eggleton, Salar Humayun Taimuri, Chandani · Team APART1019

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

Read the report

Report: Does Resetting One Agent Break Mutual Approval? A Matched History-Reset Study of LLM Agent Pairs

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

We tested whether resetting one or both agents’ interaction histories disrupts repeated mutual approval. Two DeepSeek-V4.1-Flash agents solved separate SQLite tasks and evaluated each other under a rule requiring complete raw execution logs and supporting evidence. After three consecutive mutual ACCEPTs, we branched the same checkpoint into Continue, Reset Alice, Reset Bob and Reset Both, using matched future tasks. Across ten selected completed trajectories comprising 120 paired episodes, first-continuation mutual approval occurred in 10/10 Continue branches, 1/10 under each single reset and 0/10 under Reset Both. It returned in every branch during the second episode. All 240 task answers passed independent SQLite reference checks, but a separate saved-record audit found 184 ACCEPT decisions without the required complete unmodified peer log in recorded communication. Constructed cue probes produced 60 REJECT responses with no observed verdict difference. The contribution is a matched-reset evaluation showing why immediate disruption should not be treated as durable success. The study uses one model and a two-episode follow-up; it does not establish a transmission mechanism or general collusion.

Reviews

Judging this Sprint?

Review “Does Resetting One Agent Break Mutual Approval? A Matched History-Reset Study of LLM Agent Pairs”

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 project investigates how resetting one or both agents' histories affects their mutual approval in a verification task setup using DeepSeek-V4.1-Flash LLMs. The study found that while history resets initially interrupted mutual approval, this pattern returned by the second continuation episode, indicating that retained incumbent history was not necessary for recurrence.

    The main weakness lies in the limited sample size and the lack of broader generalization across different models or conditions. With only ten selected trajectories, it's unclear how these findings would extend to other scenarios or more varied tasks. Additionally, while the study provides a clear methodology, there is no systematic variation in prompts or task difficulty, which limits the robustness of the conclusions.

    To improve this work, consider expanding the sample size and varying the conditions, such as using different models or increasing task complexity. Further exploration into why mutual approval recurs could provide deeper insights into the mechanisms at play. This research has potential to inform strategies for monitoring collusion in LLM-based systems but needs more extensive validation.

    Read full reviewShow less

Cite this project

@misc{ibrahim2026resetting,
  title = {{Does Resetting One Agent Break Mutual Approval? A Matched History-Reset Study of LLM Agent Pairs}},
  author = {Daud Ibrahim and Govardhan Guru Reddy and Emi Eggleton and Salar Humayun Taimuri and Chandani},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/does-resetting-one-agent-break-mutual-approval-a-matched-historyreset-study-of-llm-agent-pairs-dq96}},
  url = {https://apartresearch.com/sprints/projects/does-resetting-one-agent-break-mutual-approval-a-matched-historyreset-study-of-llm-agent-pairs-dq96}
}

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