Organizations: JIUTIAN Research, Beijing, China · The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China · MIIT Key Laboratory of Data and Decision Intelligence, Beihang University, Beijing, China · Beihang University, Beijing, China
Graph-based tasks in the zero-shot setting remain a significant challenge due to data scarcity and the inability of traditional Graph Neural Networks (GNNs) to generalize to unseen domains or label spaces. While recent advancements have transitioned toward leveraging Large Language Models (LLMs) as predictors to enhance GNNs, these methods often suffer from cross-modal alignment issues. A recent paradigm (i.e., Graph-R1) overcomes the aforementioned architectural dependencies by adopting a purely text-based format and utilizing LLM-based graph reasoning, showing improved zero-shot generalization. However, it employs a task-agnostic, one-size-fits-all subgraph extraction strategy, which inevitably introduces significant structural noise--irrelevant neighbors and edges--that distorts the LLMs' receptive field and leads to suboptimal predictions. To address this limitation, we introduce GraphSSR, a novel framework designed for adaptive subgraph extraction and denoising in zero-shot LLM-based graph reasoning. Specifically, we propose the SSR pipeline, which dynamically tailors subgraph extraction to specific contexts through a "Sample-Select-Reason" process, enabling the model to autonomously filter out task-irrelevant neighbors and overcome the one-size-fits-all issue. To internalize this capability, we develop SSR-SFT, a data synthesis strategy that generates high-quality SSR-style graph reasoning traces for supervised fine-tuning of LLMs. Furthermore, we propose SSR-RL, a two-stage reinforcement learning framework that explicitly regulates sampling and selection operations within the proposed SSR pipeline designed for adaptive subgraph denoising. By incorporating Authenticity-Reinforced and Denoising-Reinforced RL, we guide the model to achieve accurate predictions using parsimonious, denoised subgraphs for reasoning.
Figures & tables
Figure 1. A case of structural noise in the subgraph misleading the model’s reasoning from the Cora dataset. A case of structural noise in the subgraph misleading the model's reasoning from the Cora dataset.
Figure 2. Overview of our proposed GraphSSR framework. Overview of our proposed GraphSSR framework.
Method
Cora
WikiCS
Products
FB15K237
7
2
10
5
47
10
5
10
OFA
28.65
56.92
21.20
35.15
19.37
30.43
39.31
-
GraphGPT
44.65
-
-
-
18.84
-
-
-
UniGraph
69.53
89.74
43.45
60.23
38.45
66.07
75.73
-
ZeroG
64.21
87.83
31.26
48.25
31.24
51.24
71.29
-
LLaGA
51.85
62.73
-
-
23.10
34.15
39.72
-
Table 1. Main Results Comparison. We report Accuracy (ACC) across various tasks. The best results are indicated in bold , and the second-best are underlined . Results of baselines are from the Graph-R1 paper, and “-” denotes results are not available.
Methods
Cora (7)
WikiCS (10)
Products (47)
FB15K237 (10)
DeepSeek-R1-0528
69.59
76.96
70.17
84.12
Qwen3-14B (Thinking)
65.15
78.56
66.54
77.86
Ministral-3-14B-Reasoning-2512
69.30
75.38
59.89
77.67
DeepSeek-R1-distilled-Qwen2.5-14B
68.80
75.26
66.09
73.49
GraphSSR
72.41
79.40
68.49
78.94
Table 2. Comparison between GraphSSR and Large Reasoning Models (LRMs). Best in bold ; second best underlined .
Figure 3. Ablation study of GraphSSR on different datasets. Ablation study of GraphSSR on different datasets.
Figure 4. Parameter Sensitivity of the reward weight λ and candidate group size k . Parameter Sensitivity of the reward weight λ and candidate group size k.
Metric
Strategy
Cora (7)
WikiCS (10)
Products (47)
FB15K237 (10)
Noise Ratio (%)
Raw Subgraph
23.86
68.19
41.25
56.95
Smaller Candidates
12.29
35.05
30.38
49.91
GraphSSR
13.91
35.91
32.61
45.22
Avg. Node Count
Raw Subgraph
18.6
27.3
18.0
32.1
Smaller Candidates
2.9
3.0
3.0
3.6
GraphSSR
5.9
6.8
7.4
8.3
Table 3. Quantitative evaluation of denoising quality and structural trade-offs.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Cora (7)
Cora (2)
WikiCS (10)
WikiCS (5)
Products (47)
Products (10)
Products (5)
FB15K237 (10)
72.32 ± 0.39
89.87 ± 0.20
79.34 ± 0.14
87.89 ± 0.16
68.41 ± 0.18
86.48 ± 0.11
91.90 ± 0.09
78.75 ± 0.08
Appendix
Table 4. Performance stability across multiple runs (reported as mean ± standard deviation).
Method
Cora
Products
Graph-R1
91.63
81.14
GraphSSR
93.50
91.64
Appendix
Table 5. The comparison of accuracy on the link prediction task across unseen datasets.
Filtering Configuration
Cora (7)
WikiCS (10)
Products (47)
FB15K237 (10)
SFT_w/o_Structural_Diversity_Evaluation
69.30
77.94
65.96
77.25
SFT_with_WL_kernel_filter
68.31
78.22
66.66
77.26
SFT_with_text_sim_filter
68.07
78.13
67.09
77.22
SFT_with_hybrid_filter
68.81
78.29
67.06
77.28
SFT_GraphSSR
69.80
78.54
67.83
78.07
Appendix
Table 6. Results on the ablation and variants of the Structural Diversity Evaluation filter.
Configuration
Cora (7)
WikiCS (10)
Products (47)
FB15K237 (10)
SFT_Teacher_32B (Quality)
69.88
75.50
62.21
64.76
SFT_20%_Data (Quantity)
68.61
78.08
67.50
76.66
SFT_GraphSSR
69.80
78.54
67.83
78.07
Appendix
Table 7. Impact analysis regarding the training data configurations.
Method
Cora (7)
WikiCS (10)
Products (47)
FB15K237 (10)
Graph-R1-Select
70.93
76.99
62.59
76.08
Graph-R1
71.53
78.68
66.59
75.17
GraphSSR
72.41
79.40
68.49
78.94
Appendix
Table 8. Performance comparison between GraphSSR and the static selecting baseline.
Table 13Table 14
Domain
Dataset
#Nodes
#Edges
#Classes
Split
Social Network
Instagram
11,339
155,349
2
Training
Knowledge Graph
WN18RR
40,943
93,003
11
Training
FB15K237
14,541
310,116
237
Test (Zero-shot)
Citation
Arxiv
169,343
1,166,243
40
Training
Citeseer
3,186
8,554
6
Training
Pubmed
19,717
88,648
3
Training
Appendix
Table 9. Comprehensive statistics of the multi-domain benchmarks. We employ a zero-shot setting, ensuring that datasets marked for testing remain entirely inaccessible during the training phase to rigorously assess model generalizability.
Zero-shot reasoning on text-rich networks (TRNs) remains a challenging frontier, as models must integrate textual semantics with relational structure without task-specific supervision. While graph neural networks rely on fixed label spaces and supervised objectives, recent large language model (LLM)-based approaches often overlook graph context or depend on distillation from larger models, limiting generalisation. We propose TRN-R1-Zero, a post-training framework for TRN reasoning trained solely via reinforcement learning. TRN-R1-Zero directly optimises base LLMs using a Neighbour-aware Group Relative Policy Optimisation objective that dynamically adjusts rewards based on a novel margin gain metric for the informativeness of neighbouring signals, effectively guiding the model toward relational reasoning. Unlike prior methods, TRN-R1-Zero requires no supervised fine-tuning or chain-of-thought data generated from large reasoning models. Extensive experiments across citation, hyperlink, social and co-purchase TRN benchmarks demonstrate the superiority and robustness of TRN-R1-Zero. Moreover, relying strictly on node-level training, TRN-R1-Zero achieves zero-shot inference on edge- and graph-level tasks, extending beyond cross-domain transfer. The codebase is publicly available at https://github.com/superallen13/TRN-R1-Zero.
Yilun Liu, Ruihong Qiu, Zi Huang
School of Electrical Engineering and Computer Science The University of Queensland Brisbane, Queensland, Australia
Graph neural networks (GNNs) tightly couple their input-output parameters to dataset-specific feature spaces and target sets, exhibiting limited transferability across different datasets. In contrast, language models (LMs) generalize flexibly via a unified input-output interface, motivating recent attempts to adapt LMs to graph tasks. However, existing methods struggle to encode whole-graph information, leading to potential information loss and suboptimal graph understanding. In this work, we propose a novel weight-level information injection paradigm for adapting LMs to graph tasks. This paradigm injects whole-graph information by generating task-specific weight updates that interact directly with hidden representations. Instantiating this paradigm following low-rank adaptation (LoRA), we introduce GaRA, a Graph-aware LoRA generation model. GaRA constructs low-rank weight updates conditioned on the original graph structures and constrains the norm of the generated updates, thus injecting whole-graph information and avoiding the optimization bias in the weight generation. Empirical studies demonstrate that GaRA consistently outperforms baselines on zero-shot graph learning tasks.
Junshu Sun, Wanxing Chang, Qingming Huang +1
State Key Lab. of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China · University of Chinese Academy of Sciences, Beijing, China · DAMO Academy, Alibaba Group, Hangzhou, China
The advent of Large Language Models (LLMs) has fundamentally reshaped the way we interact with graphs, giving rise to a new paradigm called GraphLLM. As revealed in recent studies, graph learning can benefit from LLMs. However, we observe limited benefits when we directly utilize LLMs to make predictions for graph-related tasks within GraphLLM paradigm, which even yields suboptimal results compared to conventional GNN-based approaches. Through in-depth analysis, we find this failure can be attributed to LLMs' limited capability for processing graph data and their tendency to overlook graph information. To address this issue, we propose LoReC (Look, Remember, and Contrast), a novel plug-and-play method for GraphLLM paradigm, which enhances LLM's understanding of graph data through three stages: (1) Look: redistributing attention to graph; (2) Remember: re-injecting graph information into the Feed-Forward Network (FFN); (3) Contrast: rectifying the vanilla logits produced in the decoding process. Extensive experiments demonstrate that LoReC brings notable improvements over current GraphLLM methods and outperforms GNN-based approaches across diverse datasets. The implementation is available at https://github.com/Git-King-Zhan/LoReC.
Hongyu Zhan, Qixin Wang, Yusen Tan +6
The Hong Kong University of Science and Technology (Guangzhou) · Tongji University · Westlake University +2