Research map / Graph learning / Graph self-supervised learning
<SO$G_k$>: One LLM Token for Explicit Graph Structural Understanding
Jingyao Wu, Bin Lu, Zijun Di, Xiaoying Gan, Meng Jin, Luoyi Fu, Xinbing Wang, Chenghu Zhou
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the structural token SOG_k to the LLM input (ablation) | SOG_k (LLaMA3-3B) · HIV | AUC-ROC: 53.0 -> 75.7 | Improved |
| add structural token to LLM input (ablation) | SOG_k (LLaMA3-3B) · BBBP | AUC-ROC: 61.2 -> 76.9 | Improved |
| use token from the tokenizer instead of a random token (ablation) | SOG_k (LLaMA3-3B) · BBBP | AUC-ROC: 57.3 -> 76.9 | Improved |
| use graph-specific token rather than a static token (ablation) | SOG_k (LLaMA3-3B) · ClinTox | AUC-ROC: 52.5 -> 85.5 | Improved |
| use random structural token instead of tokenizer output (ablation) | SOG_k (LLaMA2-7B) · HIV | AUC-ROC: 83.4 vs 83.2 | No clear effect |
| use SOG_k instead of soft prompt | LLaMA3-3B · BBBP | AUC-ROC: 27.9 -> 76.9 | Improved |
| add soft prompt to the small LLM | LLaMA3-3B · five MoleculeNet datasets | AUC-ROC: performance drops of 8.6 to 24.2 points versus zero-shot | Hurt |
| use all three hybrid structure QAs rather than a subset (ablation) | SOG_k · five MoleculeNet datasets | AUC-ROC | Mixed |
| set structural vocabulary size K to 256 instead of 64, 128 or 512 (ablation) | SOG_k · five datasets average | AUC-ROC avg: 60.1 / 59.3 / 66.8 -> 81.7 | Improved |
| choose anchor node by degree instead of random (ablation) | SOG_k · 7 MoleculeNet tasks average | AUC-ROC avg: 72.2 -> 78.4 | Improved |
| choose anchor node by PageRank instead of degree (ablation) | SOG_k · BBBP and ClinTox | AUC-ROC: BBBP 76.9 -> 82.2; ClinTox 77.8 -> 99.6 | Mixed |
| extend SOG_k with per-node structural tokens for node classification | SOG_k LLaMA3-3B · Cora | accuracy: 68.62 (GPT-4o) -> 91.58 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GIMLET — GIMLET: A Unified Graph-Text Model for Instruction-Based Molecule Zero-Shot Learning (NeurIPS 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- From Sequence to Structure: Uncovering Substructure Reasoning in Transformers (NeurIPS 2025 · similarity 0.88)
- Multi-View Empowered Structural Graph Wordification for Language Models (AAAI 2025 · similarity 0.88)
- SSTAG: Structure-Aware Self-Supervised Learning Method for Text-Attributed Graphs (NeurIPS 2025 · similarity 0.88)
- FoGE: Fock Space inspired encoding for graph prompting (NeurIPS 2025 · similarity 0.88)
- UniGTE: Unified Graph–Text Encoding for Zero-Shot Generalization across Graph Tasks and Domains (NeurIPS 2025 · similarity 0.88)
- LLaGA: Large Language and Graph Assistant (ICML 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
