MIT, Harvard Method Cuts AI Memory Use Sharply

Researchers introduced LARS, a fine-tuning technique that reduces the memory needed to adapt large language models.

Updated on Nov. 20, 2026 in Artificial Intelligence

MIT, Harvard Method Cuts AI Memory Use Sharply

MIT CSAIL and Harvard researchers developed the Low-memory Activation Rank Subspace (LARS) method for fine-tuning language models. In tests, LARS reduced memory footprint by over 23% on GPUs and about 47% on CPUs compared with top parameter-efficient fine-tuning techniques.

Why it matters

Fine-tuning large models often fails on devices with limited memory because intermediate activations spike memory use during training. LARS pools those activations into a more compact representation rather than storing every calculation, which could make model adaptation practical on smaller hardware.

The research reports memory reductions of over 23% on GPUs and about 47% on CPUs versus leading parameter-efficient fine-tuning methods, with about 22% lower power and energy use on a CPU. The reported tests do not yet address how these gains hold across other model sizes or hardware.

The players

MIT Computer Science and Artificial Intelligence Laboratory (CSAIL)

MIT's CSAIL is one of the largest computer science research labs in the United States, known for work across machine learning and robotics.

Harvard University

Harvard University is a private research university in Cambridge, Massachusetts, and a major center for artificial intelligence research.

The details

LARS works by pooling intermediate activation details into a more concise form instead of storing every calculation during training. On a Raspberry Pi 5 with an 8 GB memory limit, it saved 9% more memory than LoRA at a context length of 128,000, though it was slightly less accurate than LoRA on detailed questions.

Timeline

  1. Research describing LARS was published on arXiv on November 20, 2026.

Roadmap

LARS builds on LoRA, the widely used low-rank adaptation method, by targeting the activation memory that LoRA leaves largely untouched during training. The approach points toward fine-tuning large models on consumer and edge hardware rather than only on high-memory data center GPUs.

If LARS is adopted, developers could fine-tune language models on devices with limited memory, such as small computers or wearables, without needing high-end hardware. The method also reportedly lowers power and energy use, which could matter for battery-powered devices.

The takeaway

Memory, not just compute, is often the bottleneck for running and adapting AI models, so efficiency gains in training can widen where these tools can be used. Readers following AI tools should watch whether memory-saving methods like LARS are validated beyond initial tests.

Further reading

Explore more on Artificial Intelligence developments shaping how models run on everyday devices.

Source note: This article includes information reported by Tech Xplore.