Research map / Graph learning / Graph self-supervised learning
Enhancing LLMs for Graph Tasks via Graph-aware LoRA Generation
Junshu Sun, Wanxing Chang, Qingming Huang, Shuhui Wang
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 norm rescaling of generated low-rank matrix (ablation) | GaRA, full tuning data · Cora (zero-shot) | accuracy: 0.164 -> 0.251 (+53.04%) | Improved |
| add norm rescaling (ablation) | GaRA, full tuning data · AmazonPhoto link prediction | accuracy: 0.539 -> 0.708 | Improved |
| add norm rescaling (ablation) | GaRA, partial tuning data · WikiCS (zero-shot) | accuracy: 0.611 -> 0.664 | Improved |
| add norm rescaling (ablation) | GaRA, partial tuning data · AmazonPhoto link prediction | accuracy: 0.527 -> 0.529 | Improved |
| increase tuning data from Partial to Full (ablation) | GaRA without norm rescaling · Cora (zero-shot) | accuracy: 0.203 -> 0.164 | Hurt |
| add weight-level injection to input-level subgraph injection (input+weight) (ablation) | UniGTE-based framework · zero-shot datasets (Partial setup) | performance | Improved |
| inject weight update at layer 21 instead of layers 1, 11, 31, 32 (ablation) | GaRA on Vicuna-7B · average over node and link datasets | average performance | Improved |
| generate the augmented left low-rank matrix instead of the right matrix (ablation) | GaRA · average over node and link datasets | average performance | Improved |
| generate weights from whole-graph representation instead of subgraph, task-description or random inputs (ablation) | GaRA · zero-shot datasets | performance | Improved |
| generate weights from subgraph features instead of non-graph features (ablation) | GaRA · zero-shot datasets | performance | No clear effect |
| add GaRA weight generation compared to UniGTE | UniGTE · training and inference cost | s/sample: 1.28 -> 1.46 train; 0.67 -> 0.71 inference | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- UniGTE — UniGTE: Unified Graph–Text Encoding for Zero-Shot Generalization across Graph Tasks and Domains (NeurIPS 2025) · better on 7 of 9 shared results
- GOFA — GOFA: A Generative One-For-All Model for Joint Graph Language Modeling (ICLR 2025) · 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
- UniGTE: Unified Graph–Text Encoding for Zero-Shot Generalization across Graph Tasks and Domains (NeurIPS 2025 · similarity 0.88)
- LoReC: Rethinking Large Language Models for Graph Data Analysis (NeurIPS 2026 · similarity 0.87)
- Large Language Model Meets Graph Neural Network in Knowledge Distillation (AAAI 2025 · similarity 0.87)
- A Unified Graph Language Model for Multi-Domain Multi-Task Graph Alignment Instruction Tuning (NeurIPS 2026 · similarity 0.87)
- LLM Enhancers for GNNs: An Analysis from the Perspective of Causal Mechanism Identification (ICML 2025 · similarity 0.87)
- LLaGA: Large Language and Graph Assistant (ICML 2024 · similarity 0.86)
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.
