AI training just got a lot less memory-hungry. 🔥 New research from @NVIDIA introduces LoGRA, cutting average reinforcement-learning training memory by up to 45.7% while preserving performance in the tested reasoning tasks. The trick: compress the learning signals into compact "gradient sketches" and control the size of each update to keep training stable. The model still learns through updates to its weights, with much less information held in GPU memory during training. They trained a 27B model for over 1,100 steps on a single node with eight H100 GPUs. On the same hardware, the standard Adam setup ran out of memory. More AI training fitting on existing hardware means more room to experiment. Efficiency improvements like this deserve more attention.