Research map / Graph learning / Graph self-supervised learning
DynamicGTR: Leveraging Graph Topology Representation Preferences to Boost VLM Capabilities on Graph QAs
Yanbin Wei, Jiangyue Yan, Chun Kang, Yang Chen, Hua Liu, James Kwok, Yu Zhang
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 |
|---|---|---|---|
| route each question to its preferred GTR instead of one fixed GTR (ablation) | VLM graph QA (GPT-4o) · In-domain graph QA | Avg. GRE: 1.21 -> 1.67 | Improved |
| add DynamicGTR router to GPT-4o on cycle detection | GPT-4o · Cycle detection | accuracy: 68.7 -> 89.3 | Improved |
| add DynamicGTR router on topological sort | GPT-4o · Topological sort | accuracy: 40.9 -> 40.8 | No clear effect |
| add DynamicGTR router for token savings on topological sort | GPT-4o · Topological sort | token consumption: 256.1 -> 176.1 | Improved |
| apply router trained on synthetic algorithm QA to unseen real-world node classification | GPT-4o · PolBlog | accuracy: 63.6 -> 71.9 | Improved |
| use Tmat (adjacency matrix) as the only GTR (ablation) | GPT-4o · In-domain graph QA | Avg. GRE: 1.05 -> 0.56 | Hurt |
| lower alpha from 0.5 to 0 (accuracy only) (ablation) | DynamicGTR · In-domain graph QA | token consumption: 212.0 -> 290.8 | Mixed |
| raise alpha from 0.5 to 1 (ablation) | DynamicGTR · In-domain graph QA | accuracy: 69.5 -> 69.3 | Mixed |
| increase probing trials k from 1 to 10 (ablation) | GTR preference dataset · In-domain graph QA | accuracy: 62.4 -> 69.5 | Improved |
| increase probing trials k from 1 to 10 (ablation) | GTR preference dataset · In-domain graph QA | token consumption: 286.2 -> 212.0 | Improved |
| use Gemini-2.5 Pro's router on GPT-4o instead of native router (ablation) | GTR router · Cross-model transfer, average | accuracy: -0.6 | No clear effect |
| use GPT-4o's router on Gemini-2.5 Pro instead of native router (ablation) | GTR router · Cross-model transfer, average | accuracy: +0.2 | No clear effect |
Closest papers by meaning
- G-Retriever: Retrieval-Augmented Generation for Textual Graph Understanding and Question Answering (NeurIPS 2024 · similarity 0.88)
- DiGraphHal-Bench: Evaluating Multimodal Large Language Models on Complex Directed Graphs (CVPR 2026 · similarity 0.87)
- GraphVLM: Benchmarking Vision Language Models for Multimodal Graph Learning (CVPR 2026 · similarity 0.87)
- Visual Graph Arena: Evaluating Visual Conceptualization of Vision and Multimodal Large Language Models (ICML 2025 · similarity 0.87)
- Talk like a Graph: Encoding Graphs for Large Language Models (ICLR 2024 · similarity 0.87)
- UniGTE: Unified Graph–Text Encoding for Zero-Shot Generalization across Graph Tasks and Domains (NeurIPS 2025 · similarity 0.87)
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.
