AgentGL: Towards Agentic Graph Learning with LLMs via Reinforcement Learning
Yuanfu Sun, Kang Li, Dongzhe Fan, Jiajin Liu, Qiaoyu Tan
2026-04-09
Summary
This paper introduces a new way for Large Language Models (LLMs) to work with information that isn't just plain text, specifically data organized as graphs – think of things like social networks or knowledge bases where items are connected to each other. It's about making LLMs better at understanding relationships between pieces of information.
What's the problem?
Current LLMs that try to act like 'agents' – meaning they can search for information and use tools – treat all external data as simple text. This is a problem because a lot of real-world information is structured as graphs, where the connections *between* data points are just as important as the data itself. Existing methods don't take advantage of these connections, limiting how well LLMs can reason about complex information.
What's the solution?
The researchers developed a framework called Agentic Graph Learning (AGL), and a specific implementation called AgentGL. AgentGL essentially gives the LLM 'graph-native tools' to explore this connected data. It uses reinforcement learning, a type of machine learning where the model learns by trial and error, to teach the LLM how to navigate these graphs efficiently and accurately. A key part is 'search-constrained thinking,' which helps the LLM balance getting the right answer with not wasting time on irrelevant information. They also use a special training method to help the LLM learn over long sequences of steps without needing constant guidance.
Why it matters?
This work is important because it shows a promising way to make LLMs much better at understanding and reasoning about complex, relational data. By allowing LLMs to directly work with graph structures, they can achieve significantly better performance on tasks like classifying items within a graph or predicting connections between them. This could lead to improvements in areas like knowledge discovery, recommendation systems, and more.
Abstract
Large Language Models (LLMs) increasingly rely on agentic capabilities-iterative retrieval, tool use, and decision-making-to overcome the limits of static, parametric knowledge. Yet existing agentic frameworks treat external information as unstructured text and fail to leverage the topological dependencies inherent in real-world data. To bridge this gap, we introduce Agentic Graph Learning (AGL), a paradigm that reframes graph learning as an interleaved process of topology-aware navigation and LLM-based inference. Specifically, we propose AgentGL, the first reinforcement learning (RL)-driven framework for AGL. AgentGL equips an LLM agent with graph-native tools for multi-scale exploration, regulates tool usage via search-constrained thinking to balance accuracy and efficiency, and employs a graph-conditioned curriculum RL strategy to stabilize long-horizon policy learning without step-wise supervision. Across diverse Text-Attributed Graph (TAG) benchmarks and multiple LLM backbones, AgentGL substantially outperforms strong GraphLLMs and GraphRAG baselines, achieving absolute improvements of up to 17.5% in node classification and 28.4% in link prediction. These results demonstrate that AGL is a promising frontier for enabling LLMs to autonomously navigate and reason over complex relational environments. The code is publicly available at https://github.com/sunyuanfu/AgentGL.