Stanford-Led Paper Trains Models From Scratch With Zero Real Data via Self-Play

Researchers from Stanford and others propose self-play pretraining with zero data: two randomly initialized models generate and learn from self-produced sequences, showing that general patterns can emerge without any real-world data.

2026-09-25 ~ 2026-09-26 · 2 related posts

• New paper shows models can learn general patterns from zero real-world data — mark_k · 2026-09-25
• Zero-data pretraining: two LMs bootstrap all training data via self-play from random init — ChengleiSi · 2026-09-26