AI Paper Atlas

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

ICLR 2026 · Graph learning › Graph self-supervised learning · approach: LLMs and graph foundation models

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 changedWhereMeasuredEffect
add the structural token SOG_k to the LLM input (ablation)SOG_k (LLaMA3-3B) · HIVAUC-ROC: 53.0 -> 75.7Improved
add structural token to LLM input (ablation)SOG_k (LLaMA3-3B) · BBBPAUC-ROC: 61.2 -> 76.9Improved
use token from the tokenizer instead of a random token (ablation)SOG_k (LLaMA3-3B) · BBBPAUC-ROC: 57.3 -> 76.9Improved
use graph-specific token rather than a static token (ablation)SOG_k (LLaMA3-3B) · ClinToxAUC-ROC: 52.5 -> 85.5Improved
use random structural token instead of tokenizer output (ablation)SOG_k (LLaMA2-7B) · HIVAUC-ROC: 83.4 vs 83.2No clear effect
use SOG_k instead of soft promptLLaMA3-3B · BBBPAUC-ROC: 27.9 -> 76.9Improved
add soft prompt to the small LLMLLaMA3-3B · five MoleculeNet datasetsAUC-ROC: performance drops of 8.6 to 24.2 points versus zero-shotHurt
use all three hybrid structure QAs rather than a subset (ablation)SOG_k · five MoleculeNet datasetsAUC-ROCMixed
set structural vocabulary size K to 256 instead of 64, 128 or 512 (ablation)SOG_k · five datasets averageAUC-ROC avg: 60.1 / 59.3 / 66.8 -> 81.7Improved
choose anchor node by degree instead of random (ablation)SOG_k · 7 MoleculeNet tasks averageAUC-ROC avg: 72.2 -> 78.4Improved
choose anchor node by PageRank instead of degree (ablation)SOG_k · BBBP and ClinToxAUC-ROC: BBBP 76.9 -> 82.2; ClinTox 77.8 -> 99.6Mixed
extend SOG_k with per-node structural tokens for node classificationSOG_k LLaMA3-3B · Coraaccuracy: 68.62 (GPT-4o) -> 91.58Improved

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.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy