LLMs + Persona-Plug = Personalized LLMs

TL;DR

<3-5 sentence high-level summary> The paper addresses the challenge of personalizing LLM outputs without per-user fine-tuning or heavy reliance on retrieval-based prompts. It introduces PPlug, a plug-and-play persona embedding learned via a lightweight User Behavior Encoder and an Input-aware Personal Aggregator, producing a single embedding $P^u$ that conditions a fixed LLM by augmenting its input with task-specific instructions $\mathbf{I}$ and history-derived representations. Through end-to-end optimization on the LaMP benchmark, PPlug significantly outperforms both fine-tuned and retrieval-based personalized LLMs, with improvements ranging up to $35.8\%$ in some tasks and robust performance across LLM/encoder variants. The work also demonstrates that integrating retrieval cues with PPlug can yield additional gains, and it discusses practical considerations, such as efficiency, privacy, and potential extensions to history granularity and selective usage of histories.

Abstract

Personalization plays a critical role in numerous language tasks and applications, since users with the same requirements may prefer diverse outputs based on their individual interests. This has led to the development of various personalized approaches aimed at adapting large language models (LLMs) to generate customized outputs aligned with user preferences. Some of them involve fine-tuning a unique personalized LLM for each user, which is too expensive for widespread application. Alternative approaches introduce personalization information in a plug-and-play manner by retrieving the user's relevant historical texts as demonstrations. However, this retrieval-based strategy may break the continuity of the user history and fail to capture the user's overall styles and patterns, hence leading to sub-optimal performance. To address these challenges, we propose a novel personalized LLM model, \ours{}. It constructs a user-specific embedding for each individual by modeling all her historical contexts through a lightweight plug-in user embedder module. By attaching this embedding to the task input, LLMs can better understand and capture user habits and preferences, thereby producing more personalized outputs without tuning their own parameters. Extensive experiments on various tasks in the language model personalization (LaMP) benchmark demonstrate that the proposed model significantly outperforms existing personalized LLM approaches.

Figures (4)

• Figure 1: The comparison of our proposed personalized LLM and previous personalized LLM approaches
• Figure 2: The overall framework of the proposed PPlug model.
• Figure 3: Overall performance of ablation models.
• Figure 4: Performance of PPlug selecting only top-$K$ user histories.