Core Concepts
LLMsはグラフ生成タスクで有望な初期能力を示す。
Abstract
大規模言語モデル(LLMs)のグラフ生成能力を評価するためにLLM4GraphGenを提案。
LLMsはルールベースと分布ベースのグラフ生成において有望な初期能力を示す。
一部の人気のあるプロンプティング方法(few-shotやchain-of-thought prompting)が一貫して性能向上につながらないことも観察される。
LLMsは特定の特性を持つ分子を生成する初期能力を示す。
Introduction
LLMs have been successful in various domains.
Recent research explores the potential of LLMs in understanding and leveraging graph structures.
Rule-based Graph Generation
GPT-4 shows reasonably good abilities for rule-based graph generation.
The impact of prompts on graph generation varies for different types of graphs.
Distribution-based Graph Generation
LLMs can understand and generate graphs with simple distributions but struggle in complex situations.
Detailed examples and CoT are helpful for distribution-based graph generation.
Property-based Graph Generation
LLMs show preliminary abilities in generating molecules with certain properties.
CoT prompt improves the performance of LLM in generating molecules with specific properties.
Stats
Large language models (LLMs) have achieved great success in many fields, and recent works have studied exploring LLMs for graph discriminative tasks such as node classification.
Graph generation requires the LLM to generate graphs with given properties, which has valuable real-world applications such as drug discovery, while tends to be more challenging.
Our evaluations demonstrate that LLMs, particularly GPT-4, exhibit preliminary abilities in graph generation tasks, including rule-based and distribution-based generation.