[Paper] Towards Effective Model Editing for LLM Personalization

Source: arXiv - 2512.13676v1

Overview

Personalization is fast becoming a must‑have feature for large language models (LLMs) that interact with real users. The paper Towards Effective Model Editing for LLM Personalization reframes personalization as a model‑editing problem: instead of costly full‑model fine‑tuning, it makes tiny, targeted edits that align the model with a user’s preferences while keeping its general abilities intact. The authors also release a new benchmark, User Preference Question Answering (UPQA), that tests whether a model can recall and apply specific user preferences in realistic, short‑answer queries.

Key Contributions

• Personalization Editing framework – a lightweight, edit‑based approach that injects user‑specific knowledge via localized weight updates guided by clustered preference representations.
• UPQA dataset – a short‑answer QA benchmark built from real user queries, covering easy to hard preference‑recall scenarios and multi‑turn interactions.
• Comprehensive evaluation – shows that Personalization Editing outperforms traditional fine‑tuning in speed and memory, and beats prompting‑based baselines on multi‑turn and implicit‑preference tasks.
• Analysis of failure modes – highlights how existing persona‑dialog benchmarks miss the information‑seeking aspect of personalization, motivating the new dataset.

Methodology

1. Preference Representation – For each user, the system gathers a small set of preference statements (e.g., “I prefer dark mode”, “My favorite cuisine is Thai”). These statements are embedded and clustered to capture distinct preference facets.
2. Localized Model Editing – Using a technique similar to “model surgery,” the authors identify a small subset of model parameters that are most sensitive to the clustered preference vectors. They then apply a low‑rank update (e.g., LoRA‑style adapters) that nudges the model’s behavior toward the user’s preferences without touching the rest of the network.
3. Edit Validation – After each edit, a lightweight validation set checks that the model still answers generic queries correctly, preventing catastrophic forgetting.
4. Benchmark Construction (UPQA) – Real user queries were collected, annotated with the correct short answer, and labeled by difficulty (explicit vs. implicit preference, single‑turn vs. multi‑turn).

The whole pipeline runs in a few minutes on a single GPU, compared to hours of full fine‑tuning.

Results & Findings

• Higher editing accuracy: The edit‑based method reliably injects the exact user preferences.
• Much faster & lighter: Only a fraction of the parameters are touched, cutting memory and time dramatically.
• Better multi‑turn behavior: Because the edit is persistent, the model retains the personalized context across turns, unlike prompting which often drifts.

Practical Implications

• Rapid onboarding – SaaS platforms can personalize a new user’s LLM assistant in seconds, without the need for a dedicated fine‑tuning pipeline.
• Edge deployment – Since edits are low‑rank, they can be shipped as small patches (a few MB) to devices with limited storage.
• Safety & compliance – Localized edits are easier to audit; you can verify that only the intended preference parameters changed, reducing the risk of hidden regressions.
• Dynamic updates – When a user’s preferences evolve (e.g., new favorite sport), the system can apply an incremental edit rather than re‑training from scratch.
• Better QA assistants – The UPQA benchmark gives product teams a concrete way to measure whether their LLM truly “remembers” user‑specific facts, a step beyond style imitation.

Limitations & Future Work

• Scope of preferences – The current approach assumes a modest, well‑defined set of explicit preferences; handling large, noisy preference histories remains open.
• Edit granularity – While low‑rank updates are efficient, they may struggle with highly complex or contradictory preferences that require deeper model changes.
• Evaluation breadth – UPQA focuses on short‑answer QA; extending to richer tasks (code generation, recommendation) would test the limits of edit‑based personalization.
• Long‑term stability – The paper notes a slight drift after many successive edits; future work could explore regularization strategies to keep the base model’s knowledge stable over many user updates.

Bottom line: By treating personalization as a targeted model‑editing problem, the authors deliver a fast, memory‑efficient way to make LLMs truly user‑aware—opening the door for scalable, on‑device, and continuously adaptable AI assistants.

Authors

• Baixiang Huang
• Limeng Cui
• Jiapeng Liu
• Haoran Wang
• Jiawei Xu
• Zhuiyue Tan
• Yutong Chen
• Chen Luo
• Yi Liu
• Kai Shu

Paper Information

• arXiv ID: 2512.13676v1
• Categories: cs.CL
• Published: December 15, 2025
• PDF: Download PDF