Agentic Graph Token Reasoning Enhances LLM Graph Analysis

Graphs are fundamental for modeling relational data across numerous scientific and industrial applications, often featuring nodes with rich textual information. While Large Language Models (LLMs) are increasingly applied to graph analysis, existing methods typically use graph tokens in a static, single-shot manner. This means a predefined graph view is encoded once, limiting the LLM's ability to perform step-by-step, adaptive reasoning. This research proposes "agentic graph token reasoning," which redefines graph tokenization as an integral part of the LLM's reasoning process. In this paradigm, the model dynamically decides which graph view to encode and at what level of detail at each step. A graph encoder is invoked on demand to generate corresponding graph tokens, which are then spliced into the LLM's running context. The training pipeline involves self-supervised tasks for reading heterogeneous graph tokens, a token-robust trajectory stage, and preference optimization. This approach has shown significant performance improvements over a wide range of baselines across seven graph domains, demonstrating zero-shot transferability to unseen domains and pushing LLM-based graph analysis towards a more native, agentic paradigm.