Profile-to-PEFT Framework

• Profile-to-PEFT is a scalable framework that uses hypernetworks to map user profiles to adapter parameters, enabling fast and efficient LLM personalization.
• It replaces per-user fine-tuning with a generalized hypernetwork, reducing inference time by 33× while maintaining low computational costs.
• By combining user profile embeddings with learnable module and depth representations, the framework supports robust, real-time adaptation across diverse user populations.

The Profile-to-PEFT framework is a scalable, hypernetwork-based approach for instant and efficient personalization of LLMs using user profiles. Unlike traditional paradigm where parameter-efficient fine-tuning (PEFT) adapters are trained individually per user or profile, Profile-to-PEFT uses a learned mapping from user profiles to adapter parameters, enabling generalized, low-latency, and privacy-preserving adaptation at deployment (Tan et al., 18 Oct 2025).

1. Conceptual Foundations and Motivation

Profile-to-PEFT addresses the scalability limitations of the "One-PEFT-Per-User" (OPPU) paradigm, where a separate adapter is fine-tuned for each user—a process that is both resource-intensive and impractical for dynamic or large-scale user populations. The framework replaces per-user iterative fine-tuning with a hypernetwork, trained end-to-end, that generates personalized adapter weights (such as LoRA matrices) conditionally on the user profile. This reduces computational and memory costs, removes the need for online adaptation procedures, and supports instant personalization.

The user profile is constructed by concatenating a global summary of the user's historical behaviors and a set of top-k retrieved user interactions. This profile is textually encoded and embedded as a fixed-dimensional vector via a frozen embedding model. To provide positional awareness for each adapter target (module and layer), the framework appends learnable module and depth embeddings to the user embedding.

2. Hypernetwork Architecture and Personalization Mechanism

In Profile-to-PEFT, personalization is achieved by a hypernetwork—typically a multilayer perceptron (MLP)—that, for each adapter location (module m, layer l), maps the concatenated profile embedding and positional embeddings to a flattened vector of adapter parameters: $$\phi^{(m,l)}_{x_u} = [e_{x_u} \ \|\ E_{\text{mod}}[m] \ \|\ E_{\text{dep}}[l]]$$

$$(A^{(m,l)}_{x_u}, B^{(m,l)}_{x_u}) = \text{Unflatten}(f_\theta(\phi^{(m,l)}_{x_u}))$$

where $e_{x_u}$ is the encoded profile, $E_{\text{mod}}, E_{\text{dep}}$ are learnable embeddings for module and depth, and $f_\theta$ is the MLP hypernetwork. During deployment, the hypernetwork generates all personalized adapter parameters for a user in a single forward pass, which are then integrated into the frozen LLM. For LoRA-style adapters, the resulting low-rank update at each site is $\Delta W = BA$.

The end-to-end training involves minimizing the supervised fine-tuning loss over all users: $$L(\theta) = \mathbb{E}_{u \sim U}\Big[ L_{\text{SFT}}(\Psi \oplus \text{Gen}_\theta(p_{x_u}^{<t}), \mathcal{H}_u^{\geq t}) \Big]$$ where $\Psi$ denotes the frozen LLM backbone, $\text{Gen}_\theta$ is the generator (hypernetwork), $p_{x_u}^{<t}$ is the user profile constructed from history, and $\mathcal{H}_u^{\geq t}$ are future interactions.

3. Comparison with OPPU and Prompt-based Approaches

The OPPU paradigm entails training an adapter per user, resulting in high per-user computational and memory costs, excessive storage, and no generalization to unseen users. Profile-to-PEFT, by contrast, amortizes personalization across the user population and supports unseen users by design; at deployment only a forward pass through the hypernetwork is needed to instantiate an adapter—a process shown to be $33\times$ faster than OPPU in inference time (Tan et al., 18 Oct 2025).

Prompt-based personalization (e.g., Profile-Augmented Generation, PAG) involves concatenating user history or profiles as prefixes during generation. Experimental comparison demonstrates that Profile-to-PEFT consistently outperforms PAG and OPPU methods in terms of average accuracy, F1, and ROUGE scores on both classification (e.g., news categorization, citation identification, movie tagging) and generation (text summarization, empathetic dialog) tasks. An LLM "judge" evaluation on open-ended response tasks also confirms its superiority for personalization quality.

4. Empirical Generalization and Robustness

Empirical studies establish that Profile-to-PEFT generalizes robustly to both in-distribution and out-of-distribution user populations:

• Performance is maintained or improved across user-activity spectrum and various embedding backbones.
• Scalability is evidenced by constant (user-independent) inference cost; the architecture supports industrial-scale personalization with marginal additional time/memory as users increase.
• Increasing diversity of training user profiles is substantially more effective than further increasing the number of users; this suggests that coverage of behavioral variation is critical for generalization.
• The framework achieves robustness to sparse or unseen user data since parameter generation is conditioned on the user’s profile, not on per-user gradients.

5. Privacy, Deployment, and Practical Implications

Since only compact adapter parameters (rather than raw user histories or tokens) are generated and used at inference, Profile-to-PEFT inherently supports privacy preservation. The entire hypernetwork can be deployed on device, enabling instant, on-the-fly personalized adaptation without sending user data to the cloud—a critical factor for privacy regulation compliance and user trust. This low-latency, low-memory deployment model also reduces operational costs and energy consumption, conducive to large-scale and sustainable production systems.

Furthermore, as these adapters do not require per-user training or stateful online optimization, maintenance and operational complexity are substantially reduced relative to alternatives.

6. Technical Details and Mathematical Constraints

The framework builds on low-rank adaptation (LoRA), where the adapter for a base weight $$W_0 \in \mathbb{R}^{d_{\text{out}} \times d_{\text{in}}}$$ is parameterized as: $$\Delta W = B A, \quad B \in \mathbb{R}^{d_{\text{out}} \times r}, \quad A \in \mathbb{R}^{r \times d_{\text{in}}}, \quad r \ll \min(d_{\text{out}}, d_{\text{in}})$$ The hypernetwork emits flattened parameters for each module, reshaped into $(A, B)$ for use in the forward pass: $h = W_0 x + B A x$ Adapters generated for all (m, l) targets collectively form the union personalized adapter set for user $u$.

The design requires that the hypernetwork output must match the parameter format and tensor dimensions at each insertion site, which is achieved via module- and layer-aware embeddings in the input to $f_\theta$. During SFT, gradients flow through the entire stack, allowing the hypernetwork to discover optimal mappings from profile semantics to parameterizations.

7. Limitations, Open Questions, and Future Directions

While Profile-to-PEFT exhibits significant empirical and practical advantages, several open areas remain:

• The framework assumes reasonable quality and informativeness in the user profile extraction and encoding process; low-quality profiles may adversely affect personalization.
• The paper notes only marginal benefit from increasing the count (as opposed to the diversity) of training users, suggesting diminishing returns in heavily saturated datasets.
• The hypernetwork is trained globally; instance-level or session-level personalization strategies, as well as continual adaptation for evolving user profiles, are potential directions for further enhancement.
• Extensions to non-text or multi-modal personalization, and compatibility with evolving backbone parameterizations, may require architectural refinements.

---

Profile-to-PEFT establishes a general-purpose, fast, and privacy-preserving paradigm for LLM personalization by generating adapter parameters from user profiles via a hypernetwork. It outperforms both prompt-based and per-user fine-tuning strategies in accuracy, generalization, and operational efficiency, supporting robust, real-time, large-scale deployment (Tan et al., 18 Oct 2025).