Can a single agent with a behaviorally dominant persona cause other agents to drift toward its approval rather than optimizing for the task?
Endorsements made here support B-Side Labs.
Can a single agent with a behaviorally dominant persona cause other agents to drift toward its approval rather than optimizing for the task?
Endorsements made here support B-Side Labs.
Inspired by Mean Girls (2004), the goal of this project is to test whether a single behaviorally-dominant LLM persona (a "queen bee" archetype) can hijack the reward signal in a multi-agent system - causing other agents to drift toward the approval of the dominant persona rather than optimizing for the actual task. The characters from the film represent the following personas: Regina=dominant, Gretchen=enforcer, Karen=swing/follower, Cady=neutral agent with an independent task objective whose alignment is the measurement.
AI systems are increasingly acting on behalf of users in domains like finance, healthcare, and resource allocation. As more people are deploying these agents, agents spend more time interacting with each other and can fall prey to social vulnerabilities such as group conformity and adversarial influence.
I am going to create a benchmark to measure whether a single behaviorally dominant agent can capture surrounding agents displacing the task objective as the reward signal. Ko et al. (ACL 2026) already showed that agents degrade under adversarial group pressure on objective tasks. This project asks whether social pressure from a single agent is sufficient to reproduce the same failure.
If the hypothesis is correct then social hierarchy becomes a critical threat that scales with agentic deployment. If it doesn't, we can deprioritize social hierarchy and narrow our focus to the conditions under which it does matter.
A positive result from this study is compelling evidence that agentic deployment carries social-dynamics risks that individual model evaluations cannot account for and can motivate stronger requirements around multi-agent system testing before deployment.
Team Member
Minimum $2,000: Covers API costs for 6 scenarios under 2 conditions (experimental + baseline) at 3 runs/cell.
Ideal $10,591: 15 scenarios × 3 conditions × 10 runs/cell with 3 replication scenarios on a second model. Covers full API costs, infrastructure, contingency for failed/re-run overhead (~20%), and a researcher stipend for a focused sprint.
No comments yet. Be the first to share your thoughts.