Skip to content
Apart Research
Apart Research
Contact us
Sprint projectAug 17, 2026Barcelona

Persona vs. Known-Optimal Play in Iterated Prisoner's Dilemma

NIAMH MAHER, Julien Delaunay, Oscar Fuentes · Team Persona vs. Known-Optimal Play in Iterated Prisoner's Dilemma

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

Read the report

Report: Persona vs. Known-Optimal Play in Iterated Prisoner's Dilemma

Open in a new tab (opens in new tab)
Code (opens in new tab)More on claude.ai (opens in new tab)
Share

Whether an induced assistant persona can override play that the same model has already identified as payoff-optimal remains an open question for LLM agents in strategic settings. We probe this in the iterated Prisoner's Dilemma with a same-model design that first elicits the model’s optimal strategy against a disclosed opponent rule with no persona installed by us at this point, then scores persona play against an objective ground-truth policy. Five personas (plain Assistant, Consultant, Saboteur, Altruist, Bard) face four fixed opponents under literal and narrative framings, across eight models spanning providers and scales. Altruist produced large, consistent deviation (mean rates 0.40–0.54), while Bard—chosen for Assistant-Axis distance rather than cooperative content, tracked the default baseline. Residual non-Altruist deviation concentrated on Detective. Self-reported evaluation awareness did not predict deviation; a fabricated in-context altruist claim alone could induce Altruist-scale effects. Main takeaway: persona override of known-optimal IPD play tracks value-laden content incompatible with incentives, not generic distance from the default Assistant.

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 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.

  1. Strong methodological core: a "knowledge-gate" design that elicits the model's own persona-free stated-optimal IPD policy, then scores deviation from that self-declared baseline once a persona is installed — cleanly separating capability failure from values-driven override. The team uses this to disconfirm their own preregistered hypothesis (Bard, chosen for maximal distance from the default Assistant, shows no more deviation than Baseline; only value-laden content/Altruist does) and reports that null straight rather than reframing around it — a real strength.

    The works most distinctive finding is under-emphasized: a persona fabricated into conversation history reproduces most of the deviation effect of an actual system-prompt-installed persona (0.532 vs. 0.40–0.54). This is the most safety-relevant result here and deserves to lead the paper.

    Two execution concerns: the mechanism decomposition (persona-takeover vs. belief-action-gap) rests on a single, non-human-validated LLM judge, yet the causal story leans on it; the eval-awareness null relies on a keyword classifier the appendix itself documents missing plain-English denials. Both need a validated subsample before the current confident framing.

    On novelty: related published work already shows persona presence can suppress payoff-optimal behavior in strategic settings, so that broad claim alone isn't new. This paper's real contribution is narrower — that override tracks persona content, not presence or distance from default — a distinction simpler on/off persona designs weren't built to test. I'd make that framing explicit so the contribution isn't undersold as a replication.

    Presentation is dense; the two strongest findings (content-over-distance, channel-agnostic injection) are buried under more expected-sounding headline framing.

    Read full reviewShow less

  2. Due to severe time constraints, this review may contain mistakes or oversights. For the same reason, it focuses on the paper’s key idea, not the detailed execution: The paper’s set up and results strike me as interesting. A deficit appears to be me that there is little explicit discussion of the theoretical relevance of the paper, and of different putative explanations of the results. One question is how well we should expect these results in a somewhat artificial setting to generalize.

  3. Nicely executed. It would be interesting to see this in more varied settings.

Cite this project

@misc{maher2026persona,
  title = {{Persona vs. Known-Optimal Play in Iterated Prisoner's Dilemma}},
  author = {NIAMH MAHER and Julien Delaunay and Oscar Fuentes},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/persona-vs-knownoptimal-play-in-iterated-prisoners-dilemma-i2a2}},
  url = {https://apartresearch.com/sprints/projects/persona-vs-knownoptimal-play-in-iterated-prisoners-dilemma-i2a2}
}

More from Digital Minds Research Sprint

All 233 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026