Researchers introduce LARS, a fine-tuning method that decouples memory consumption from sequence length to enable on-device LLM adaptation on resource-constrained hardware.

LARS (Low-memory Activation-Rank Subspace) constrains the activation subspace during training rather than model parameters, directly targeting memory consumption instead of sequence length scaling.

LARS reduces memory footprint by an average of 33.54% on GPUs and 51.95% on CPUs compared to LoRA while maintaining competitive accuracy and throughput across reasoning, understanding, and long-context datasets.

The method has been deployed on Raspberry Pi and consumer-grade CPUs, demonstrating that sophisticated LLM personalization can run on resource-constrained hardware and edge devices.