Towards Effective Model Editing for LLM Personalization
Baixiang Huang, Limeng Cui, Jiapeng Liu, Haoran Wang, Jiawei Xu, Zhuiyue Tan, Yutong Chen, Chen Luo, Yi Liu, Kai Shu
TL;DR
This paper presents Personalization Editing, a localized model editing framework for efficient, precise, and scalable personalization of large language models, outperforming fine-tuning and prompting in various user preference tasks.
Contribution
The paper introduces a novel model editing approach for personalization that is more efficient and accurate than existing methods, with a new benchmark for evaluating preference recall.
Findings
Higher editing accuracy than fine-tuning
Greater computational efficiency
Outperforms prompting in multi-turn and implicit queries
Abstract
Personalization is becoming indispensable for LLMs to align with individual user preferences and needs. Yet current approaches are often computationally expensive, data-intensive, susceptible to catastrophic forgetting, and prone to performance degradation in multi-turn interactions or when handling implicit queries. To address these challenges, we conceptualize personalization as a model editing task and introduce Personalization Editing, a framework that applies localized edits guided by clustered preference representations. This design enables precise preference-aligned updates while preserving overall model capabilities. In addition, existing personalization benchmarks frequently rely on persona-based dialogs between LLMs rather than user-LLM interactions, or focus primarily on stylistic imitation while neglecting information-seeking tasks that require accurate recall of…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
1. The paper proposes a novel framework that conceptualizes LLM personalization as a model-editing task, so that the existing model-editing methods can be adopted to tackle the LLM personalization problem. 2. The paper presents a new dataset, UPQA, together with clearly defined metrics, which can serve as a new benchmark for LLM personalization methods. 3. The paper extensively experiments with existing model-editing methods and presents comprehensive experimental results.
1. Although the paper presents definitions and high-level formulas for the Personalization Editing framework, such as input, output, objective, etc., it lacks sufficient details about how existing model-editing methods are used to be compatible with the framework and the evaluation datasets. The constrained objective is conceptually clear, but integration details per editor (e.g., layer selection, masking strategy) are not specified. 2. The paper primarily augments existing editors, and no new
* The proposed Personalization Editing is an interesting and novel editing task. * The proposed framework demonstrated and robust strong performance on editing
* The formatting in Section 3 looks a bit disorganized with few sentences in a subsection, I recommend that the authors summarize the problem setups in fewer subsections and elaborate a bit more on the objective function, say, what exactly is the loss function $\mathcal{L}$? * The overall novelty of this work is limited, the theoretical and empirical insight may not benefit broader audience beyond this task.
+ This work frames personalization as model editing, enabling efficient and persistent parameter-level updates that outperform prompting in multi-turn and implicit-query settings. + This work introduces UPQA, a realistic short-answer benchmark with explicit, rephrased, implicit, and recommendation queries plus synonym clusters for robust evaluation. + This work proposes clustering-based preference representations that generalize to paraphrases and implicit requests, improving accuracy and mult
- Personalization metric relies on short-answer accuracy, failing to capture nuanced preferences, trade-offs, or long-term satisfaction across diverse tasks and interactions. - No human evaluation prevents measuring perceived relevance, satisfaction, or harms from personalized edits in real users. - Focus on short answers limits applicability to open-ended tasks, multi-turn dialogues, or creative personalization needs—unclear generalization to longer responses. - Parameter edits may cause uni
1. The paper clearly articulates a valid and important problem: the failure of prompting/RAG-based personalization in multi-turn conversations (as demonstrated in Figure 5) and the high cost of full fine-tuning. 2. The paper conducts a thorough comparison of several representative model editing techniques (ROME, LoRA, FT-M) and baselines (Zero-Shot) on its new benchmark, providing a clear performance landscape. 3. The problem of generalizing from stated preferences to implicit user needs is a
1. The paper's central concept, "Personalization Editing," does not appear to be a novel algorithm. Rather, it is a re-application of existing model editing techniques (LoRA, FT-M, ROME) to the problem of personalization. The primary technical contribution, "clustering-based preference representation," is a data augmentation strategy, not a new editing mechanism. This frames the paper more as an evaluation study than a significant methodological advancement. 2. The paper's main contribution, th
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsTopic Modeling · Information Retrieval and Search Behavior · Expert finding and Q&A systems