Graph Linearization Methods for Reasoning on Graphs with Large Language Models
Christos Xypolopoulos, Guokan Shang, Xiao Fei, Giannis Nikolentzos, Hadi Abdine, Iakovos Evdaimon, Michail Chatzianastasis, Giorgos Stamou, Michalis Vazirgiannis
TL;DR
This paper proposes new graph linearization methods based on centrality and degeneracy to improve graph reasoning with large language models, enabling better understanding of graph structures through meaningful token sequences.
Contribution
It introduces novel graph linearization techniques tailored for LLMs, enhancing graph representation and reasoning capabilities in multimodal contexts.
Findings
Linearization methods outperform random baselines.
Enhanced linearization improves graph understanding in LLMs.
Methods facilitate integration of graph learning with multimodal models.
Abstract
Large language models have evolved to process multiple modalities beyond text, such as images and audio, which motivates us to explore how to effectively leverage them for graph reasoning tasks. The key question, therefore, is how to transform graphs into linear sequences of tokens, a process we term "graph linearization", so that LLMs can handle graphs naturally. We consider that graphs should be linearized meaningfully to reflect certain properties of natural language text, such as local dependency and global alignment, in order to ease contemporary LLMs, trained on trillions of textual tokens, better understand graphs. To achieve this, we developed several graph linearization methods based on graph centrality and degeneracy. These methods are further enhanced using node relabeling techniques. The experimental results demonstrate the effectiveness of our methods compared to the random…
Peer Reviews
Decision·Submitted to ICLR 2026
This paper introduces the concept of **graph linearization**, which aims to transform graph structures into linear sequences of tokens. It further proposes two guiding principles for effective linearization: local dependency and global alignment. Local dependency refers to the ability of large language models (LLMs) to predict subsequent or missing tokens within a sequence derived from a single linearized graph. The experimental results demonstrate that the proposed core/degree-based graph linea
1. **Limited Contribution and Novelty**. The contribution of this paper is limited, particularly in terms of novelty. Prior work has already demonstrated the sensitivity of token ordering in LLM-based graph reasoning (e.g., GraphQA, GraphText). Thus, the concept of graph linearization is not new. Moreover, the paper adopts classical heuristics such as Degree/Core/PageRank for node ordering, methods that have been widely used for decades, without proposing any new algorithmic contribution to impr
1. This paper's principal strength lies in its clear, focused, and well-motivated conceptual contribution. It successfully identifies and articulates a fundamental, under-explored problem at the intersection of LLMs and graph learning: the critical importance of the sequence orderwhen linearizing a graph for LLM consumption. The proposal of the two guiding principles—local dependency and global alignment—is both intuitive and grounded in the distributional properties of natural language, provi
1. A primary weakness of this work is its insufficient engagement with the most relevant and powerful baselines, critically undermining its claim of contribution. The paper convincingly demonstrates that its structured linearizations outperform a randomordering of edges. However, in the realm of graph reasoning, the more meaningful comparison is against specialized Graph Neural Networks (GNNs) and graph transformers (e.g., Graphormer), which are architecturally designed to capture graph structur
The paper focuses on an interesting direction—encoding graphs for LLMs. While many studies have explored this in specific domains such as molecules, general methods that work across different types of graphs remain relatively underexplored and deserve further investigation. The idea of considering both local dependency and global alignment is important.
## The evaluation should be extended to real-world data Testing on synthetic data is useful but not sufficient. The current evaluation is quite limited and provides little insight into practical applications. More real-world datasets should be considered, including but not limited to: 1. Social networks at different scales. Social networks can range from hundreds to millions of nodes. It is important to consider how the linearized graph size affects the context window of LLMs, while still pres
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsSemantic Web and Ontologies · Advanced Graph Neural Networks · Natural Language Processing Techniques