Improving Complex Reasoning over Knowledge Graph with Logic-Aware Curriculum Tuning

TL;DR

This work presents Logic-Aware Curriculum Tuning (LACT), a fine-tuning framework that enables 7B-scale LLMs to perform complex logical reasoning over incomplete knowledge graphs. By combining Binary Tree Decomposition to convert $EFO_{1}$ queries into a logic-rich computation tree and curriculum learning to balance difficulty across query types, LACT achieves state-of-the-art results on FB15K, FB15K-237, and NELL995, with an average MRR improvement of approximately $+5.5\%$. The approach emphasizes knowledge sharing from KGs during training and demonstrates strong transferability across datasets and model scales, offering a scalable solution for KG reasoning without resorting to large closed-source models. Its practical impact lies in enabling accurate, scalable reasoning over incomplete KGs using accessible 7B models, which broadens the deployment potential of KG-aware AI systems.

Abstract

Answering complex queries over incomplete knowledge graphs (KGs) is a challenging job. Most previous works have focused on learning entity/relation embeddings and simulating first-order logic operators with various neural networks. However, they are bottlenecked by the inability to share world knowledge to improve logical reasoning, thus resulting in suboptimal performance. In this paper, we propose a complex reasoning schema over KG upon large language models (LLMs), containing a curriculum-based logical-aware instruction tuning framework, named LACT. Specifically, we augment the arbitrary first-order logical queries via binary tree decomposition, to stimulate the reasoning capability of LLMs. To address the difficulty gap among different types of complex queries, we design a simple and flexible logic-aware curriculum learning framework. Experiments across widely used datasets demonstrate that LACT has substantial improvements~(brings an average +5.5% MRR score) over advanced methods, achieving the new state-of-the-art.

Figures (10)

• Figure 1: Schematic illustration: a) Answering logical query over KG with LACT. b) The framework of Logic-Aware Curriculum Tuning over LLama. We leverage the binary tree decomposition strategy (Seen in Section Methodology) to construct a logic-rich FT corpus and the Curriculum learning strategy (Seen in Section Methodology) to fine-tune a base LLM. c) Performing reasoning using well-designed prompts.
• Figure 2: Performance of scaling LACT on FB15K-237 with different a) model and b) data scales.
• Figure 3: Results of Ablation Studies. (a) Comparison PPL and train loss results of whether to use BTD based on FB15k. (b) Comparison PPL results of whether use CL based on FB15K high-difficulty queries.
• Figure 4: (left) MRR performance of LACT and previous SOTA methods at different difficulties based on FB15K. (right) The correlation between related information completeness and accuracy evaluated on FB15K-237, we selected a 3p query and pi query with the same inference path length as the task types. We assume that the completeness of all simple queries is 1.
• Figure 5: Query structure diagram of 14 types of queries