Coalition Formation Demo
Given agents with differing policy positions, this algorithm finds a majority coalition that agrees on an AI-generated compromise. Read the paper
How it works:
Each agent has an ideal policy sentence. Starting from a status quo, the algorithm iterates:
(1) agents vote yes if the current proposal is closer to their ideal than the status quo
(measured by cosine dissimilarity of sentence embeddings);
(2) if a majority coalition is reached, the proposal wins;
(3) otherwise, two agents are selected and an LLM generates a compromise sentence as the next proposal.
The algorithm guarantees that every agent in the winning coalition genuinely prefers the result over the status quo.