GRASP: Graph Reasoning via Agentic Solving and Probing of LLMs

Xiaojun Guo, Mingxue Tian, Chenheng Zhang, Xiaohan Wang, Jiajun Chai, Guojun Yin, Wei Lin, Yifei Wang, Yisen Wang
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:38316-38336, 2026.

Abstract

Integrating graph knowledge into Large Language Models (LLMs) via passive representation faces critical bottlenecks: limited context windows, unreliable numerical computation, and structural hallucinations. To solve this, we propose GRASP (Graph Reasoning via Agentic Solving and Probing), shifting the paradigm from passive ingestion to proactive agentic exploration. By interleaving Neighbor Retrieval for on-demand probing with Code Interpreter as a deterministic solver, GRASP enables LLMs to autonomously navigate and compute over complex topologies. We employ a staged reinforcement learning strategy (GRPO) that transitions from visible tuning to a structure-blind environment, forcing the agent to develop genuine topological awareness. Evaluated on multi-domain graph reasoning benchmarks, our 4B model achieves a 53.06% average performance boost, surpassing SOTA baselines like DeepSeek-V3.2 and successfully generalizing to unseen tasks, with high potential for tackling sampling on million-node graphs and solving Hard-level LeetCode graph problems. Our implementation is open-sourced at https://github.com/PKU-ML/GRASP, with models hosted on Huggingface collection https://huggingface.co/collections/PKU-ML/grasp.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-guo26t, title = {{GRASP}: Graph Reasoning via Agentic Solving and Probing of {LLM}s}, author = {Guo, Xiaojun and Tian, Mingxue and Zhang, Chenheng and Wang, Xiaohan and Chai, Jiajun and Yin, Guojun and Lin, Wei and Wang, Yifei and Wang, Yisen}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {38316--38336}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/guo26t/guo26t.pdf}, url = {https://proceedings.mlr.press/v306/guo26t.html}, abstract = {Integrating graph knowledge into Large Language Models (LLMs) via passive representation faces critical bottlenecks: limited context windows, unreliable numerical computation, and structural hallucinations. To solve this, we propose GRASP (Graph Reasoning via Agentic Solving and Probing), shifting the paradigm from passive ingestion to proactive agentic exploration. By interleaving Neighbor Retrieval for on-demand probing with Code Interpreter as a deterministic solver, GRASP enables LLMs to autonomously navigate and compute over complex topologies. We employ a staged reinforcement learning strategy (GRPO) that transitions from visible tuning to a structure-blind environment, forcing the agent to develop genuine topological awareness. Evaluated on multi-domain graph reasoning benchmarks, our 4B model achieves a 53.06% average performance boost, surpassing SOTA baselines like DeepSeek-V3.2 and successfully generalizing to unseen tasks, with high potential for tackling sampling on million-node graphs and solving Hard-level LeetCode graph problems. Our implementation is open-sourced at https://github.com/PKU-ML/GRASP, with models hosted on Huggingface collection https://huggingface.co/collections/PKU-ML/grasp.} }
Endnote
%0 Conference Paper %T GRASP: Graph Reasoning via Agentic Solving and Probing of LLMs %A Xiaojun Guo %A Mingxue Tian %A Chenheng Zhang %A Xiaohan Wang %A Jiajun Chai %A Guojun Yin %A Wei Lin %A Yifei Wang %A Yisen Wang %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-guo26t %I PMLR %P 38316--38336 %U https://proceedings.mlr.press/v306/guo26t.html %V 306 %X Integrating graph knowledge into Large Language Models (LLMs) via passive representation faces critical bottlenecks: limited context windows, unreliable numerical computation, and structural hallucinations. To solve this, we propose GRASP (Graph Reasoning via Agentic Solving and Probing), shifting the paradigm from passive ingestion to proactive agentic exploration. By interleaving Neighbor Retrieval for on-demand probing with Code Interpreter as a deterministic solver, GRASP enables LLMs to autonomously navigate and compute over complex topologies. We employ a staged reinforcement learning strategy (GRPO) that transitions from visible tuning to a structure-blind environment, forcing the agent to develop genuine topological awareness. Evaluated on multi-domain graph reasoning benchmarks, our 4B model achieves a 53.06% average performance boost, surpassing SOTA baselines like DeepSeek-V3.2 and successfully generalizing to unseen tasks, with high potential for tackling sampling on million-node graphs and solving Hard-level LeetCode graph problems. Our implementation is open-sourced at https://github.com/PKU-ML/GRASP, with models hosted on Huggingface collection https://huggingface.co/collections/PKU-ML/grasp.
APA
Guo, X., Tian, M., Zhang, C., Wang, X., Chai, J., Yin, G., Lin, W., Wang, Y. & Wang, Y.. (2026). GRASP: Graph Reasoning via Agentic Solving and Probing of LLMs. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:38316-38336 Available from https://proceedings.mlr.press/v306/guo26t.html.

Related Material