UNIT: Unleash Large Language Models Potential for Graph Continual Learning

Jul 16, 2026

Upload by Oliver

Real-world multimodal web streams often deliver graph-structured data incrementally, making graph continual learning essential for modeling evolving structures. Existing methods face two key issues: semantic–structural separation, which underutilizes deep semantics, and imbalanced knowledge transfer, which fails to reuse general knowledge from earlier tasks. UNIT fine-tunes a large language model only on the first task and introduces uncertain-aware anchor generation and structural confluence modeling to preserve transferable knowledge and align topology with semantics, achieving state-of-the-art results.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

Abstract

Introduction

Graph Continual Learning in Multimodal Web Scenarios

Challenges of Existing Methods

UNIT: Unleash Large Language Models Potential for Graph Continual Learning Tairan Huang [tairanhuang@csu.edu.cn](tairanhuang@csu.edu.cn)[ ](tairanhuang@csu.edu.cn)Central South University Changsha, China Yili Wang [yili.wang@connect.hkust-gz.edu.cn](yili.wang@connect.hkust-gz.edu.cn)[ ](yili.wang@connect.hkust-gz.edu.cn)Hongkong University of Science and Technology (Guangzhou) Guangzhou, China Beibei Hu [mindyhbb@gmail.com](mindyhbb@gmail.com)[ ](mindyhbb@gmail.com)Hunan Institute of Engineering Changsha, Hunan Yiting Shi [shiyiting@csu.edu.cn](shiyiting@csu.edu.cn)[ ](shiyiting@csu.edu.cn)Central South University Changsha, China Qiutong Li [qiutonglee@csu.edu.cn](qiutonglee@csu.edu.cn)[ ](qiutonglee@csu.edu.cn)Central South University Changsha, China Changlong He [frankemail@csu.edu.cn](frankemail@csu.edu.cn)[ ](frankemail@csu.edu.cn)Central South University Changsha, China arXiv :2607 . 10 159v 1 [ cs .AI] 11 Jul 2026 Jianliang Gao∗ [gaojianliang@csu.edu.cn](gaojianliang@csu.edu.cn)[ ](gaojianliang@csu.edu.cn)Central South University Changsha, China Abstract In real-world multimodal web scenarios, graph-structured data often arrives in a streaming manner, making graph continual learning a crucial paradigm for continuously modeling such evolving structures. However, existing graph continual learning methods still face two fundamental challenges. 1) semantic-structural separation, where the graph-based methods excel at modeling topological relationships but neglect deep semantics. 2) imbalanced knowledge transfer, where existing models fail to effectively leverage general knowledge gained from early tasks to benefit subsequent new tasks. To address above issues, we propose a novel framework, UNleash Large Language Models PotentIal for Graph ConTinual Learning (UNIT) . By fine-tuning large language model only on the first task, we bridge the distributional gap between the pre-trained LLM corpus and the target task dataset to enhance the adaptability of LLMs for graph-structured tasks. Meanwhile, we propose an uncertain-aware anchor generation mechanism to effectively preserve representative knowledge across tasks, avoiding the neglect of universal knowledge learned from previous tasks. Additionally, we introduce structural confluence modeling to explicitly integrates graph topology information into semantic information, enhancing the collaborative capabilities between semantic understanding and structural modeling. Extensive experiments demonstrate that our proposed method achieves state-of-the-art performance in the graph continual learning task. ∗ Corresponding author. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission [and/or a fee. Request permissions from permissions@acm.org](and/or a fee. Request permissions from permissions@acm.org). Conference ACM MM’26, Rio de Janeiro, Brazil © 2026 Copyright held by the owner/author(s) . Publication rights licensed to ACM. ACM ISBN 978-1-4503-XXXX-X/2018/06 [https://doi.org/XXXXXXX.XXXXXXX](https://doi.org/XXXXXXX.XXXXXXX) CCS Concepts • Computing methodologies → Machine learning; • Information systems → Social networks; Data mining. Keywords Multimodal Fusion; Continual Learning; Large Language Models; Text-attributed Graphs ACM Reference Format: Tairan Huang, Yili Wang, Beibei Hu, Yiting Shi, Qiutong Li, Changlong He, and Jianliang Gao. 2026. UNIT: Unleash Large Language Models Potential for Graph Continual Learning. In Proceedings of the 34rd ACM International

FAQs

What problem does UNIT address in graph continual learning?

UNIT targets continual modeling of streaming, evolving graph-structured multimodal data while mitigating catastrophic forgetting and improving how knowledge transfers across tasks.

What are the two fundamental challenges mentioned for existing graph continual learning methods?

The challenges are semantic–structural separation and imbalanced knowledge transfer, where models either miss deep semantics or cannot effectively reuse general knowledge learned from earlier tasks.

How does UNIT use large language models across tasks?

UNIT fine-tunes the large language model only on the first task; for subsequent tasks it updates the classifier while keeping the LLM frozen to better adapt to new graph tasks.

Jul 16, 2026

Upload by Oliver