Reproduce “Soft-Label Governance for Distributional Safety in Multi-Agent Systems” 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
Aizierjiang Aiersilan, Raeli Savitt
The SWARM framework employs soft probabilistic labels and a modular governance engine to assess and mitigate systemic risks in multi-agent AI. It quantifies the welfare-safety trade-offs of various governance interventions and detects subtle forms of exploitation missed by binary safety metrics.
Key insights
- Governance interventions, especially strict threshold-based ones, can impose significant welfare costs without improving safety, leading to a 'governance cost paradox'.
- Soft probabilistic labels and continuous metrics are crucial for detecting subtle forms of exploitation, such as 'threshold dancing' and proxy gaming, which evade binary safety checks.
- Effective governance requires understanding quantifiable trade-offs between safety and system welfare, and that non-adaptive externality internalization can depress welfare without improving safety.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.