I · The Research

MIRROR

Converging cognitive principles as computational mechanisms for AI reasoning.

21% relative gain across seven models

Theories of human cognition converge on three principles: parallel specialized threads, synthesis into one bounded first-person narrative, and reconstruction each turn instead of accumulation. MIRROR implements them as concrete mechanisms, and the gains land where the theories predict.

YOAO

You Only Align Once: propagating cooperative behaviors in multi-agent systems through seed agents.

24.8% → 62.2% cooperation · 21.6% → 91.5% zero-shot

A single aligned agent, placed among untrained teammates, spreads cooperation through nothing but conversation. Trained only in the Red-Black Game, the same seed transfers zero-shot to Sugarscape. Alignment reframed: not training every agent, but placing the right one.