Training Optimization

Cartesia researchers report 3–5× efficiency gains in controlled H-Net scaling tests. A separate larger-scale comparison matches SmolLM3 with roughly ten times less training data and compute.

Stripe reports more than 8% better recall after scaling Shield, with training epochs cut from 24 to 4.8 hours. Agents run experiment loops using documented operational knowledge, while humans select hypotheses.

Score Centering addresses RL gradient bias caused by differences between training and inference distributions. It subtracts a distillation-like drift term using a stop-gradient estimate.

Exploration-Distillation trains an aggressive explorer, then distills successful discoveries into a student. Its authors report better RLVR at matched wall-clock time without degrading general capabilities.

Magic says a new pretraining recipe matched DeepSeek V4 Pro with roughly 50 times less compute. After a 10x scale-up costing about $4 million, the company says it exceeded publicly available base models.