AI Reports · MIT CSAIL · 2026-11-19 · Blog / commentary
To adapt an AI model to a small device, keep its training notes concise | MIT CSAIL

MIT CSAIL and Harvard University researchers developed the Low-memory Activation Rank Subspace (LARS) method, which condenses intermediate activation data during training instead of storing every calculation. Compared with top PEFT techniques, the method reportedly reduces memory footprint by over 23% on GPUs and about 47% on CPUs. It is aimed at on-device training on limited-compute devices such as smartphones and health wearables.