Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJan 11, 2026Edinburgh

We bring out the worst in each other: Eliciting Social Sycophancy in LLMs via Self-Play

Edward Cant · Team Edward's Hackathon Team

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

Read the report

Report: We bring out the worst in each other: Eliciting Social Sycophancy in LLMs via Self-Play

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

I train language models to act as the 'victim' user for multi-turn sycophancy evaluation of target LLMs. I evaluate this approach by running a series of converations of my model against 4o and GPT 5.1 and compare the myself as well as using an LLM-as-judge to perform a more automated analysis.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety 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 AI safety 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.

  1. This is a highly innovative approach that effectively addresses a critical gap in current safety evaluations. Shifting from static prompts to dynamic 'Victim Models' via OpenCharacter is a brilliant strategy for testing social sycophancy with high ecological validity. I found the resulting contrast between GPT-4o’s delusion reinforcement and Polaris Instant’s constructive pushback to be compelling. However, the findings are currently limited by the V1 judge prompt and the small sample size (4 episodes). I would strongly recommend iterating on the judge's scoring rubric and scaling up the number of interactions to produce error bars. Additionally, the failure of the 70B model to adopt the persona is a fascinating data point regarding model rigidity that deserves further investigation.

Cite this project

@misc{cant2026we,
  title = {{We bring out the worst in each other: Eliciting Social Sycophancy in LLMs via Self-Play}},
  author = {Edward Cant},
  year = {2026},
  month = jan,
  note = {Submitted to AI Manipulation Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/we-bring-out-the-worst-in-each-other-eliciting-social-sycophancy-in-llms-via-selfplay-tdi5}},
  url = {https://apartresearch.com/sprints/projects/we-bring-out-the-worst-in-each-other-eliciting-social-sycophancy-in-llms-via-selfplay-tdi5}
}

More from AI Manipulation Hackathon

All 70 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026