K-DeCore: Facilitating Knowledge Transfer in Continual Structured Knowledge Reasoning via Knowledge Decoupling

TL;DR

The paper tackles continual learning for structured knowledge reasoning across heterogeneous schemas by introducing K-DeCore, a framework that fixes backbone parameters and employs knowledge decoupling to separate schema filtering from query construction. It adds a dual-perspective memory and a structure-guided pseudo-data synthesis strategy to retain knowledge and enable generalization to unseen patterns. Across four SKR benchmarks and multiple backbones, K-DeCore achieves state-of-the-art results in accuracy and forward transfer while avoiding parameter growth. The approach has practical implications for real-world systems requiring continual, scalable reasoning over evolving knowledge bases.

Abstract

Continual Structured Knowledge Reasoning (CSKR) focuses on training models to handle sequential tasks, where each task involves translating natural language questions into structured queries grounded in structured knowledge. Existing general continual learning approaches face significant challenges when applied to this task, including poor generalization to heterogeneous structured knowledge and inefficient reasoning due to parameter growth as tasks increase. To address these limitations, we propose a novel CSKR framework, \textsc{K-DeCore}, which operates with a fixed number of tunable parameters. Unlike prior methods, \textsc{K-DeCore} introduces a knowledge decoupling mechanism that disentangles the reasoning process into task-specific and task-agnostic stages, effectively bridging the gaps across diverse tasks. Building on this foundation, \textsc{K-DeCore} integrates a dual-perspective memory consolidation mechanism for distinct stages and introduces a structure-guided pseudo-data synthesis strategy to further enhance the model's generalization capabilities. Extensive experiments on four benchmark datasets demonstrate the superiority of \textsc{K-DeCore} over existing continual learning methods across multiple metrics, leveraging various backbone large language models.

Figures (5)

• Figure 1: a) Overview of the continual SKR task. The backbone LLM is frozen and only the PEFT module if tunable. b) SKR based on knowledge decoupling. Schema filtering refines the scope of a given schema and impacts various SKR performance. The differences between schema filtering in these SKR tasks are minimal, making it potential for knowledge transfer.
• Figure 2: The left panel presents the K-DeCore training framework, organized into two key stages for each SKR task: schema filtering and query building, each supported by specialized PEFT modules. By unifying the schema, the framework aims to bridge the gap between tasks, effectively enabling the knowledge transfer. The right panel illustrates the creation process of structure-guided synthetic pseudo samples, designed to offer a more structurally diverse set of examples.
• Figure 3: AA (%), BWT (%), and FWT (%) till the seen tasks after learning on each task, using Llama3-8B (top row) and QWEN2.5-8B (bottom row). Solid lines represent the mean values across three distinct task sequences, while shaded regions indicate the standard deviation.
• Figure 4: Peformance of K-DeCore with varing memory sizes and synthetic sample percentages.
• Figure 5: Training and testing time of various methods.