A hypernetwork that turns a user's example texts into a small personalized LoRA adapter matched or exceeded three other customization methods across tests with 1-billion- and 2-billion-parameter Gemma models. After shared training, generating each user's adapter required forward passes rather than gradient-based tuning on the device.
Personalizing a local language model usually creates one of two costs. In-context learning places example conversations in every query, lengthening the prompt and its attention computation. Parameter-efficient fine-tuning can produce a compact low-rank adaptation, or LoRA, but normally requires a separate optimization run for every user.
The proposed hypernetwork combines those steps differently. It reads a set of query-and-preferred-response examples, uses the device's frozen base model plus a learned embedder LoRA to summarize the user, and maps that representation to pairs of low-rank matrices. Those matrices modify selected attention weights in the same base model during later requests. The examples are processed when the personal adapter is created, so they do not need to remain in each inference prompt.
One shared model generates many personal adapters
The expensive phase remains population-level training. The researchers first trained the embedder LoRA, then optimized the full hypernetwork so that adapters derived from one portion of a user's history predicted another portion. A distillation term also encouraged the generated adapters to follow LoRAs tuned separately for individual users.
Evaluation used instruction-tuned Gemma 3 1B and Gemma 1 2B base models with adapter ranks four and eight. The tasks were personalized Amazon-review writing, Reddit-post writing and scholarly-title generation from the LongLaMP-3, LongLaMP-4 and LaMP-5 datasets. Comparisons used adapters of the same rank for a non-personal baseline, in-context learning, per-user fine-tuning and the hypernetwork. The in-context-learning baseline used a shared LoRA adapter trained to work with user examples in the prompt, rather than an adapter fitted separately to each user.
Average ROUGE-1, a word-overlap score against a reference answer, was highest or nearly highest for the generated adapters in every model, rank and dataset combination shown. On the Amazon and Reddit tasks, the method improved ROUGE-1 over the best non-personal baseline for more than three quarters of users in almost every configuration. Scholarly-title generation was harder to personalize for all methods.
The evaluation controlled an important decoding variable. The researchers swept generation temperature for each method and dataset, used temperature 1.0 for the Amazon and Reddit tasks and greedy decoding for titles, and averaged ten generations per test query when sampling was enabled. Even so, ROUGE overlap is a proxy for personalization; the paper does not report human judgments of whether users preferred the generated style or content.
Phone performance remains untested
The on-device claim describes the adapter-generation architecture, not a measured mobile deployment. The paper reports no phone latency, energy use, peak memory or adapter-generation time. Its shared training experiments used clusters of 32 to 128 accelerators, each with 32 GiB of memory, and typically ran for several hours. That cost occurs before deployment, but it still matters to anyone reproducing or updating the system.
The authors also identify a population boundary. A hypernetwork learns how to map examples into adaptations seen across its training users, so a user far outside that distribution may receive a poor adapter. Direct per-user fine-tuning could serve such a case better. Production training would also need privacy protections for real user histories; the paper proposes federated learning and trusted execution environments, while leaving differential-privacy training for future study.
Sources
- Paper record and abstract, submitted September 21, 2026
- Full paper with architecture, training procedure, experiments and limitations
- Lead image: Figure 3 from the paper, comparing average ROUGE-1 across personalization methods, published under the paper's CC BY 4.0 license.