Graph learning under distribution shift presents a persistent challenge, particularly when models must adapt to new graphs with limited or no supervision. Recent graph-LLM approaches have moved towards label-efficient prediction by linearizing graphs into prompts and utilizing large language models (LLMs) as predictors, adopting Chain-of-Thought (CoT) prompting to leverage LLM's multi-step reasoning capabilities. However, existing CoT-based graph-LLM methods generate intermediate thoughts while conditioning on fixed graph tokens, limiting the step-wise refinement of structural cues.
This paper introduces CoEvoT, a simple yet effective co-evolving CoT prompting framework for graph-LLM reasoning. CoEvoT couples text-to-graph token rewriting and graph-to-text reasoning guidance in a closed loop: each intermediate textual thought is used to update the graph token evidence state via a lightweight condition network, and the updated tokens are fed back into the next-step instruction to guide subsequent LLM reasoning.
This approach enables step-wise, state-aware evidence refinement instead of reasoning over a fixed graph snapshot. Extensive experiments across eight datasets demonstrate that CoEvoT consistently outperforms state-of-the-art baselines.
Blogger's Review: The CoEvoT framework opens new possibilities in graph learning by dynamically updating graph tokens. Its step-wise, state-aware reasoning enhances model adaptability and provides fresh insights for future graph-LLM research. Notably, its superior performance across multiple datasets indicates a broad range of applications.