Method Cuts AI Memory Needs for Fine-Tuning

Researchers at MIT and Harvard describe a technique that reduces the memory required to fine-tune AI models, as outlined in a preprint on arXiv.

Updated on Nov. 22, 2026 in Artificial Intelligence

Method Cuts AI Memory Needs for Fine-Tuning

Live Poll

Would you use AI features that run directly on your own phone or device?

Researchers from MIT and Harvard developed LARS, a method that lowers the memory needed for fine-tuning AI systems by compressing the information they temporarily store during training. The approach is described in a preprint published on the arXiv scientific repository.

Why it matters

Intermediate calculations during training can cause sudden spikes in memory use that exhaust device memory, which limits where advanced AI can run. A more compact approach could make such tasks feasible on smaller hardware.

Benchmarks show LARS saved around 10 gigabytes of memory on more powerful GPUs and reduced power and energy consumption by around 22 per cent in some processor tests. Its accuracy trade-off relative to LoRA in certain tests has not been resolved.

The players

Massachusetts Institute of Technology

A private research university in Cambridge, Massachusetts, known for its engineering and computer science programs.

Harvard University

A private research university in Cambridge, Massachusetts, and one of the oldest institutions of higher learning in the United States.

The details

LARS retains a more compact representation of intermediate calculations made during training, and it modifies a method that adjusts only a small proportion of model values without eliminating temporary training computations. The projected applications include running more capable AI features directly on smartphones, wearables and compact computers, which could reduce delays and limit how much data leaves a device.

Timeline

  1. Article published on 22 November 2026 describing the LARS method.

The Tech Race

LARS builds on LoRA, a widely used method that already reduced the cost of adapting large models by adjusting only a small share of their values. The new work targets the memory LoRA still consumes during training, which is the constraint that keeps many fine-tuning tasks off consumer hardware.

If adopted, the technique could let phones, wearables and small computers run more capable AI features without sending data to remote servers. Users may see faster responses and less information leaving their devices, though practical availability depends on future implementation.

The takeaway

Lower memory requirements for fine-tuning could widen who can build and run customized AI models. Readers following AI hardware should watch whether the accuracy trade-offs narrow as the method matures.

Further reading

Explore more on Artificial Intelligence developments shaping how models are trained and deployed.

Source note: This article includes information reported by BGNES: Breaking News, Latest News and Videos.

Live Poll

Would you use AI features that run directly on your own phone or device?