Reproduce “Cooperation and Control in Delegation Games” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Oliver Sourbut, Lewis Hammond, Harriet Wood
The University of Oxford research introduces a unified framework for "delegation games" to analyze the performance of multi-agent AI systems acting on behalf of human principals. It quantifies how individual alignment, individual capabilities, collective alignment, and collective capabilities influence principal welfare, demonstrating empirically that higher values in each measure correlate with increased welfare.
Key insights
- Achieving high principal welfare requires addressing all four facets—individual alignment, individual capabilities, collective alignment, and collective capabilities—as perfect individual control alone is insufficient due to cooperation problems.
- Principal welfare regret is bounded by a combination of individual and collective alignment, individual and collective capabilities, and a "fairness remainder" term highlighting the importance of proportional utility magnitudes between principals and agents.
- Small individual irrationalities or capability limitations in agents can compound in multi-agent environments, significantly degrading overall welfare and underscoring the need for robust system design.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.