↓Skip to main content

RAIE: Region-Aware Incremental Preference Editing with LoRA for LLM-based Recommendation

We study the problem of user preference drift in LLM-based recommendation and propose RAIE, a region-aware incremental editing framework. Instead of global updates or instance-level edits, RAIE introduces preference regions as structured units for localized adaptation. This design enables efficient continual learning while preserving stable preferences.

Core Problem: Existing adaptation strategies either update the entire model (causing interference and forgetting) or operate at the instance level (failing to capture distributional shifts). There is a lack of middle-level granularity for stable and precise preference adaptation.

📄 Paper 📁 Code

Main Contributions #

  • Region-based Preference Modeling: Introduces preference regions as semantically coherent clusters of user behaviors.
  • Region-aware Incremental Editing: Designs three editing operations (Update / Expand / Add) to model different drift patterns.
  • Region-specific LoRA Adaptation: Assigns each region a dedicated LoRA adapter for localized parameter updates.

Method #

1. Overview #

RAIE decomposes preference adaptation into three stages:
(1) knowledge region construction, (2) region-aware editing, and (3) routing-based inference.

RAIE framework
Figure 1. The overall architecture of our proposed RAIE. It consists of three core modules: (a) Knowledge Region Construction in the setup phase (S); (b) Region-aware Editing and LoRA Adaptation in the incremental learning finetune phase (F); and (c) Region-aware Routing in the inference test phase (T).

2. Knowledge Region Construction #

User interaction sequences are segmented into subsequences and encoded by a frozen LLM. These representations are clustered via spherical k-means to form preference regions:

  • each region has a center ( c_k ) and radius ( R_k )
  • each region corresponds to a specific interest pattern

3. Region-Aware Editing and Adaptation #

For each incoming subsequence, RAIE:

  1. computes similarity to existing regions
  2. selects an editing operation based on confidence
  • Update: refine region center (small drift)
  • Expand: enlarge region boundary (moderate drift)
  • Add: create a new region (emerging preference)

Each region is associated with a dedicated LoRA adapter, trained only on its regional data.

4. Inference Routing #

At inference time:

  • map sequence → region
  • activate the corresponding LoRA adapter

This enables dynamic, context-aware adaptation without modifying the backbone model.

Experimental Setup #

Setup: We evaluate RAIE under a time-sliced continual learning protocol (Set-up → Finetune → Test), measuring both retention and adaptability.
  • Datasets: MovieLens-10M, Yelp
  • Backbones: BERT4Rec, SASRec, TiSASRec, OpenP5
  • Baselines: LoRA, Replay, LwF, LSAT, MoLE, E-BPR
  • Metrics: Recall@10, NDCG@10

Representative Findings #

  • Superior Adaptation: RAIE achieves the best performance on future data (Test split).
  • Strong Retention: Maintains competitive performance on historical data (Set-up split).
  • Balanced Learning: Effectively resolves the stability–plasticity trade-off.
  • Interpretable Adaptation: Visualization shows that RAIE preserves global region structure while performing localized adjustments, leading to controlled and interpretable preference updates.

Main results
Figure 2. RAIE consistently outperforms baselines across datasets and backbones.
Region visualization
Figure 3. Visualization of preference regions before and after editing, showing localized geometric adjustments with preserved global structure.
Key Insight: Preference drift should be handled via structured, localized editing, rather than global parameter updates.

Citation #

WWW 2026

Jin Zeng, Yupeng Qi, Hui Li, Chengming Li, Ziyu Lyu, Lixin Cui, Lu Bai. 2026. RAIE: Region-Aware Incremental Preference Editing with LoRA for LLM-based Recommendation. arXiv preprint arXiv:2603.00638.

BibTeX
@article{zeng2026raie,
  title={RAIE: Region-Aware Incremental Preference Editing with LoRA for LLM-based Recommendation},
  author={Zeng, Jin and Qi, Yupeng and Li, Hui and Li, Chengming and Lyu, Ziyu and Cui, Lixin and Bai, Lu},
  journal={arXiv preprint arXiv:2603.00638},
  year={2026}
}