Alignment Propagation: Spreading Cooperative Behaviors in Multi-Agent Systems through Seed Agents

Asuka Yuxi Zheng, Nicole Hsing, Yi Zhao, Haoqin Tu, Jen-tse Huang

Keywords: Social Simulation, Multi-Agent Systems, Cooperation, Emergent Behavior, AI Alignment, Game Theory, Social Dynamics

TL;DR: We show alignment can propagate through multi-agent systems, emerging from social dynamics and spreading via trained agents, without training each model individually.

Abstract:

Multi-agent systems typically require exhaustive per-agent alignment to ensure cooperative behavior, an approach that scales poorly in open environments.
In this paper, we introduce **Alignment Propagation**, demonstrating that cooperative rationale can be instilled into a single fine-tuned ``seed'' agent and propagated to untrained, unaligned agents purely through interaction.
We evaluate this propagation across two distinct environments: a discrete social dilemma with broadcast deliberation (**Red-Black Game**) and a continuous resource-competition world with pairwise negotiation (**Sugarscape**).
We show that introducing a single seed agent more than doubles the cooperation rate (from 26\% to 62\%) in Red-Black Game.
Furthermore, these seed agents transfer zero-shot to Sugarscape without retraining, achieving a 91.5\% trade success rate compared to a 21.6\% baseline, outperforming prompt-based frontier models like Gemini-3.0-Pro.
Finally, we establish that propagation efficiency is fundamentally governed by communication topology, requiring only a 20\% seed ratio to shift group behavior in Red-Black Game versus approximately 50\% in Sugarscape.
Our findings indicate that alignment can scale effectively as a transferable social capability.

Submission Number: 191