Discovering Cooperative Pipelines: Autoresearch for Sequential Social Dilemmas

A two-level autoresearch system let an outer AI agent modify and evaluate an inner LLM policy-synthesis pipeline for sequential social dilemmas.

In Cleanup and Gathering, it outperformed hand-built and prompt-only baselines across two LLM synthesizers and two welfare objectives.

The system also reduced run-to-run variability and found that fairness-related pipeline changes emerged only under Rawlsian maximin optimization.

The results suggest AI agents can autonomously improve research pipelines for cooperation problems and uncover objective-specific mechanisms.