Cooperative AI Seminar: Solving AI Game Theory Dilemmas with Safe Pareto Improvements

xuanalogue · x · 2026-08-14

Cooperative AI hosted a seminar featuring Nathaniel Sauerberg (UT Austin), who shared his research on Safe Pareto Improvements (SPIs).

Background: Advanced AI systems face challenges in credibly committing to agreements. Players might not agree on default game dynamics or might strategically posture, preventing consensus.
Method: Instead of agreeing on specific outcomes, the SPI approach seeks commitments that leave the game strategically equivalent ('isomorphic') to the original but with payoffs that constitute a Pareto improvement.
Progress: Sauerberg provided geometric characterizations for when SPIs exist and discussed affordances enabling them in larger classes of games, concluding with open research questions.

More from Safety