DEP: Difference-aware Embedding Personalization

• DEP is a personalization framework that explicitly models inter-user differences via embeddings to capture unique user traits.
• It leverages user history and peer embedding differences, refined through sparse autoencoders, to construct soft prompt tokens for LLMs.
• DEP architectures yield improved metrics in text generation and federated recommendation by integrating user uniqueness with minimal retraining.

Difference-aware Embedding-based Personalization (DEP) frameworks constitute a family of personalization methods that leverage explicit modeling of inter-user differences in embedding space to enhance the specificity and fidelity of outputs generated by large models, particularly in the context of LLMs and recommendation systems. By constructing and injecting embeddings that represent not only a user's historical behaviors or preferences but also their relative deviation from peer users, DEP architectures aim to distill and utilize “what makes this user unique,” thereby enabling more precise and effective personalization beyond summary-based or retrieval-only baselines (Qiu et al., 28 Jul 2025).

1. Theoretical Motivation and Formalization

DEP arises from the limitation of conventional personalization methods that rely exclusively on user-centric representations, typically summarizing or retrieving a user's own historical data without situating the user within a broader population-level context. This overlooks the critical comparative signal required to identify and amplify distinguishing characteristics essential for true personalization. DEP frameworks generalize the modeling objective to the following problem:

Given a corpus of user-item interactions $\mathcal{D} = \{ (u, i, y_u^i) \}$, for a target user $u'$, item $i'$, and output $y_{u'}^{i'}$, the objective is to generate $\hat{y}_{u'}^{i'}$ that not only matches the user's general writing style and preferences (user history), but also integrates inter-user differences with respect to peers who have interacted with $i'$ or similar items, thereby informing the model about the user's unique profile (Qiu et al., 28 Jul 2025).

2. Core Methodological Components

DEP methodologies are characterized by the following stages:

1. User History Embedding: For each reference user and their key historical items, generate user-specific embeddings. Formally, with a frozen encoder $f_\text{emb}$,

$$e_\text{his}^i = f_\text{emb}(y_{u'}^i) \in \mathbb{R}^d$$

2. Peer Embedding and Difference Computation: Identify peer users $\{u_j\}_{j=1}^m$ who have experience with the same or similar items $i$. Compute peer-specific embeddings and derive a difference-aware embedding as

$u'$0

This operation quantifies how the target user's textual output deviates relatively from population norms on a per-item basis (Qiu et al., 28 Jul 2025).

1. Latent Compression and Refinement: Both $u'$1 and $u'$2 are compressed and refined via a sparse autoencoder (SAE), yielding bottleneck representations $u'$3.
2. Soft Prompt Construction: Compressed representations are projected into the LLM input space via a trainable lightweight MLP, and prepended to the autoregressive context as “soft prompt tokens”:

$u'$4

These tokens, demarcated by special token boundaries, inform the downstream model of both baseline and difference-aware user profiles.

1. Personalized Generation: The downstream task (e.g., review generation) is performed by a frozen LLM, conditioned on these soft prompt tokens (Qiu et al., 28 Jul 2025).

3. Architectural Instantiations and Algorithmic Details

Several variants of DEP have been proposed, capturing the central ethos of difference-awareness via diverse architectural mechanisms:

• Latent Inter-User Difference Modeling (DEP): Embeddings are learned via a frozen encoder (e.g., BGE-m3). Both user-history and difference-aware embeddings are sparsified via an autoencoder. Only the autoencoder and projection layers are trained; the LLM remains frozen. No language-based difference extraction is used, avoiding verbosity and unreliability associated with LLM prompting. The entire process is formalized with reconstruction and sparsity regularizers:

$u'$5

Total loss integrates standard generation loss:

$u'$6

(Qiu et al., 28 Jul 2025).

• Difference-aware Personalization Learning (DPL): Contrasts with DEP by relying on LLM-based text prompts for extracting difference signals. For each key history, the target review is compared against peer reviews (clustered to maximize coverage) via an LLM prompt along preset axes (writing style, emotional tone, content). Structured differences are then summarized and injected into generation as explicit, human-readable instructions (Qiu et al., 4 Mar 2025).
• Additive Personalization in Federated Settings (FedRAP): In federated recommendation, a global embedding matrix $u'$7 (for all items) is shared across clients, while each user maintains a private local update $u'$8. The final embedding is

$u'$9

where $i'$0 is global and $i'$1 is a user-specific deviation (“difference”). A Frobenius regularizer encourages local views to focus on deviations rather than duplicating global knowledge. Sparsity and curriculum schedules are used to balance learning efficiency and communication cost (Li et al., 2023).

Method	Difference Extraction	Compression/Refinement	Personalization Target
DEP (Qiu et al., 28 Jul 2025)	Latent (embedding subtraction)	Sparse autoencoder	LLM generation
DPL (Qiu et al., 4 Mar 2025)	LLM text-based prompt	None	LLM generation
FedRAP (Li et al., 2023)	Additive (global+local embedding)	None	Federated recommender

4. Regularization, Optimization, and Training Protocols

DEP approaches commonly employ regularization schemes to enforce the semantic distinctiveness and sparsity of difference-aware representations:

• Sparsity Penalty: The average activation of SAE latent dimensions is penalized via KL divergence to a target ρ (e.g., 0.05).
• Reconstruction Loss: SmoothL1 on both user-history and difference embeddings ensures information retention through compression.
• Curriculum Scheduling (FedRAP): Regularization strengths are gradually increased using $i'$2-based schedules to prevent hindering early learning dynamics.

Most approaches keep downstream LLM parameters frozen to reduce retraining cost and facilitate plug-and-play personalization. Only the modules responsible for embedding compression, difference computation, and soft-prompt projection are updated (Qiu et al., 28 Jul 2025, Li et al., 2023).

5. Empirical Performance and Comparative Evaluation

DEP consistently sets state-of-the-art results in benchmarks requiring personalized text or recommendation outputs:

• Text Generation (LLMs): On Amazon Reviews 2023 (Books, Movies & TV, CDs & Vinyl) with Qwen2.5-Instruct 7B backbone, DEP achieves ROUGE-1 0.3334, METEOR 0.2634, BLEU 8.24, and BERTScore 0.5107. These are 11.4% higher in ROUGE-1 and 17.9% higher in BLEU than next-best DPL baselines (Qiu et al., 28 Jul 2025).
• Federated Recommendation: On MovieLens-100K, FedRAP achieves HR@10 of 97.1% and NDCG@10 of 87.8%, outperforming prior federated methods by substantial margins (Li et al., 2023).
• Ablation Analyses: Removing difference-aware signals or compressive refinement leads to substantial metric drops. Combining user-history and difference-aware embeddings yields 10–15% improvements over using either alone (Qiu et al., 28 Jul 2025).
• User Uniqueness: DEP provides especially large gains for highly unique users, underscoring its superiority in leveraging distinctiveness for personalization.

6. Practical Implications and Extensions

DEP’s advantages include minimal retraining cost, robust adaptation to new users, and ease of extension to other modalities (e.g., multimodal embeddings). Its latent-space contrastive approach circumvents verbosity and inconsistency in LLM-prompted difference extraction. In recommendation settings, additive and explicitly regularized decompositions afford interpretability and strong privacy/communication guarantees in federated contexts (Li et al., 2023).

A plausible implication is that the explicit modeling of difference-aware embeddings, via either latent subtraction or additive decomposition, constitutes a generally applicable principle in personalization—potentially extendable via attention or gating mechanisms, and to a variety of user-item data regimes.

7. Limitations and Future Directions

Current DEP frameworks depend on the availability of high-quality peer reviews or population-level data for meaningful difference extraction, and may encounter expressivity bottlenecks if user uniqueness is not well-captured in the embedding space. Further, the interaction between coarse-grained (e.g., sentiment) and fine-grained (e.g., syntax) difference signals invites more sophisticated feature disentanglement, potentially through richer multi-view or multimodal supervision.

Opportunities remain for developing unified frameworks that synergize latent-space and text-prompted difference modeling, improve interpretability of soft prompts, and extend DEP principles via dynamic user clustering or online adaptation for continually evolving user personas (Qiu et al., 28 Jul 2025, Qiu et al., 4 Mar 2025, Li et al., 2023).