LoRA-generating hypernetworks for efficient on-device LLM generative personalization
The paper proposes making each phone generate its own LoRA adapter from the user’s context, without training on the device.
A shared hypernetwork is trained to turn context tokens into a personalized low-rank adaptation for the base LLM. On device, the user only needs forward passes to synthesize that LoRA, which the authors frame as cheaper than fine-tuning and lower-latency than adding long context. The method is designed to keep storage small by reusing parts of the target model’s weights. The authors report tests on personalization datasets, with emphasis on long-form generation tasks. ArXiv · AI/CL/LG's note
A shared hypernetwork is trained to turn context tokens into a personalized low-rank adaptation for the base LLM. On device, the user only needs forward passes to synthesize that LoRA, which the authors frame as cheaper than fine-tuning and lower-latency than adding long context. The method is designed to keep storage small by reusing parts of the target model’s weights. The authors report tests on personalization datasets, with emphasis on long-form generation tasks. ArXiv · AI/CL/LG's note
score 4