Field Notes

Enabling Frontier-Scale Training for the GLM-5.3 Family

Bringing GLM-5.3 and GLM-5.3 Flash to SkyRL with five-minute numerical checks, faster adapter synchronization, and early training results.

Intelligence Density

Task performance per unit of compute spent at inference.

Choosing the Trainable Geometry for Agentic RL

Choosing the trainable geometry for agentic reinforcement learning.

Continual Learning Cheap Enough to Meter