Paper Overview
Research area: ML Authors: Sean Augenstein, Li Ding, Jihwan Lee, Keith Rush, Andrey Zhmoginov Published: 2026-09-21 arXiv: 2609.24979
Summary
On-device large language models (LLMs), e.g. running on mobile phones, are ripe for improvement via personalization. The limited compute resources of mobile devices impose limits on model scale and thus model quality, making any realizable quality gains highly impactful. At the same time, their personal nature (i.e., the close coupling to a particular user) means that a given on-device LLM tends to be used in similar, predictable patterns over time.
This paper presents a novel method for personalizing on-device LLMs: it trains a hypernetwork to map a user's context tokens to a LoRA (low-rank adaptation) well-suited to that user. Once the trained common artifacts are deployed to users' devices, each user can synthesize a personalized LoRA entirely on device.
Key Idea
The method combines the strengths of existing LLM customization approaches while avoiding their weaknesses:
- Like in-context learning (ICL) — and unlike PEFT — the on-device stage is computationally feasible, requiring only a neural network forward pass.
- Like PEFT — and unlike ICL — it modifies the target base LLM through weights (LoRA), avoiding the negative effects of extending the input sequence, such as increased latency.
- On-device compute and latency advantages.
- Because the architecture internally reuses parts of the target LLM's weights, it requires almost no additional storage.
Why It Fits Mobile Devices
Evaluation
The authors validate the LoRA-generating hypernetwork on multiple representative personalization datasets, comparing against baselines such as ICL and PEFT. Notably, the personalization experiments focus on the more challenging and less-studied setting of long-form generation tasks.
--- *Source: arXiv:2609.24979*