Coalition formation clean

[HILLELOH]

Implements the coalition formation algorithm from:

"AI-Generated Compromises for Coalition Formation"
Eyal Briman, Ehud Shapiro, Nimrod Talmon (2024)
https://arxiv.org/pdf/2512.05983

The algorithm finds a majority coalition of agents by iteratively proposing AI-generated compromise sentences. Each agent votes to
accept or reject a proposal based on whether it is closer to their ideal position than the current status quo (using cosine
dissimilarity on sentence embeddings). The process repeats until a proposal wins majority support.

New files

• pref_voting/coalition_formation.py — main implementation
• tests/test_coalition_formation.py — full test suite

Algorithm overview

1. Embed all agents' ideal sentences + status quo using a sentence transformer
2. Generate N candidate compromise sentences via GPT
3. Each agent votes: accept if proposal is closer to ideal than status quo
4. If majority accepts -done; else update status quo and repeat
5. Returns winning sentence + list of supporting agent indices

Programmer

Hillel Ohayon