Aplaud made user-specific LLMs lighter with a shared basis and low-rank corrections

Personalized survey-response prediction has three linked constraints: each user supplies little training data, storing a separate adapted model is expensive, and the model should exploit structure shared across survey questions. A per-user adaptation strategy therefore becomes hard to scale.

The obvious route is to keep a separate fine-tune for every user. Aplaud extends LoRA by freezing a shared low-rank basis and learning only a compact user-specific correction, then adds a rank-one residual for finer personalization. The correction can itself be factorized into an even lower-rank form.

The paper reports efficient, scalable personalization across users, with Aplaud outperforming state-of-the-art LoRA-based personalized LLM approaches in both generalization and inference efficiency.

Share the basis; store each user's correction

A shared low-rank basis amortizes common task structure across users.

A compact correction captures user-specific variation without duplicating the shared adaptation.

The rank-one residual adds finer personalization at small parameter cost.

Further factorization reduces per-user parameters and helps mitigate overfitting.

When users repeat the same task, amortize the shared structure and store only a low-rank deviation per user. Personalization can scale through small user state instead of a separate model copy.

Aplaud reframes personalization as one shared adaptation plus a small per-user state. Its reported generalization and inference-efficiency gains show the intended payoff: personalize a large language model without making storage and inference grow like a separate full adaptation for every user.