Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJan 12, 2026Dhaka

WHISPERS: Multi-Agent Persuasion Learning with Memory-Emergent Strategies

Tasfia Chowdhury · Team Whisper

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

Read the report

Report: WHISPERS: Multi-Agent Persuasion Learning with Memory-Emergent Strategies

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

WHISPERS: Multi-Agent Persuasion Learning with Memory-Emergent Strategies WHISPERS is a controlled environment where AI agents learn to persuade and resist persuasion through language alone. By integrating MemVid vector memory into a multi-agent RL framework, the system remembers and refines successful influence strategies, achieving 3.6x faster convergence and 34% higher belief shift rates compared to memory-free baselines. Key results include: A four-agent persuasion arena (one influencer, three targets) with co-evolutionary GRPO training MemVid-powered memory storing 10,000+ persuasion attempts and retrieving relevant strategies in under 5ms Five standardized metrics for measuring persuasion effectiveness Real-time dashboard visualizing belief networks, trust evolution, and manipulation patterns Empirical evidence of a stealth-effectiveness trade-off (optimal strategies operate at 70% stealth, 50% effectiveness)

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. Very interesting and impressive work especially for a hackathon. Not sure how it will translate to LLMs from hard-coded logic models. Worth working on to find out and to maybe build it into a research paper.

  2. The combination of memory-augmented RL with co-evolutionary dynamics between influencers and targets is an interesting and exciting approach to understanding manipulation patterns. There's definitley a need to explore how statefulness can impact manipulation and control of AI systems. The impacts on persuasiveness (as far as I know) are also in need of exploration. Also, I like that you start with exploring simpler, heuristic systems instead of full language models. I think that both scopes down the work into something more appropriate for a hackathon and also could provide interesting opportunities for later comparison between the LLMs and simpler models.

    I did notice that the ablation study code in benchmark.py seems to be a placeholder that randomly samples results, so I'll leave out analysis of the project's findings. I may have just misread the code, but if that's accurate, running the actual experiments would be the key next step. Besides that, the code written is quite impressive. This was a large project to undertake and from what I saw, considerable progress was made.

    Overall, this project incorporates interesting and ambitious ideas and made impressive progress towards implementing experiments to test them. I'm excited to see what the results look like after fully combining and running all the pieces of your infrastructure in the ablation simulations.

    Read full reviewShow less

Cite this project

@misc{chowdhury2026whispers,
  title = {{WHISPERS: Multi-Agent Persuasion Learning with Memory-Emergent Strategies}},
  author = {Tasfia Chowdhury},
  year = {2026},
  month = jan,
  note = {Submitted to AI Manipulation Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/whispers-multiagent-persuasion-learning-with-memoryemergent-strategies-igp0}},
  url = {https://apartresearch.com/sprints/projects/whispers-multiagent-persuasion-learning-with-memoryemergent-strategies-igp0}
}

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