← changmin-lee.com
ICML 2026
EPIC

From Volume to Value: Preference-Aligned Memory Construction for On-Device RAG

Changmin Lee  ·  Jaemin Kim  ·  Taesik Gong
UAI Lab, UNIST

📄 arXiv 💻 Code 🖼 Poster
TL;DR

EPIC is a preference-aligned memory construction framework for on-device RAG that selectively retains preference-relevant information and aligns retrieval toward user preferences, achieving 18.79 pp higher accuracy with 2,404× lower memory usage and 32.17× lower latency than the best baseline.

🎬 Demo video coming soon
Abstract

Retrieval-Augmented Generation (RAG) on personal devices requires managing memory under tight resource budgets while aligning with user-specific preferences. Existing approaches focus on volume-based memory management, retaining information based on recency or frequency, without considering what users actually prefer. We propose EPIC, a preference-aligned memory construction framework that selectively retains preference-relevant information and aligns retrieval toward user preferences. EPIC first identifies preference-relevant information from user interactions and constructs a compact preference-aligned memory. During retrieval, EPIC prioritizes memory entries that align with inferred user preferences. Experiments on four benchmarks show that EPIC improves preference-following accuracy by 18.79 percentage points on average while reducing memory usage by 2,404× and retrieval latency by 32.17× compared with the strongest baseline.

BibTeX
@inproceedings{lee2026epic, title = {From Volume to Value: Preference-Aligned Memory Construction for On-Device {RAG}}, author = {Lee, Changmin and Kim, Jaemin and Gong, Taesik}, booktitle = {International Conference on Machine Learning}, year = {2026} }