Research map / Vision-language and multimodal / Human and embodied modeling
SGEITL: Scene Graph Enhanced Image-Text Learning for Visual Commonsense Reasoning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 SGEITL to VL-BERT | VL-BERT · VCR validation | accuracy: 72.9 -> 74.9 (Q→A); 75.3 -> 77.2 (QA→R); 54.9 -> 57.8 (Q→AR) | Helped |
| Add scene-graph predicate features (ablation) | VL-BERT on VCR · VCR validation | accuracy: 72.9 -> 73.1 (Q→A); 75.3 -> 75.6 (QA→R); 54.9 -> 55.3 (Q→AR) | Helped |
| Add weakly supervised, domain-relevant predicates (ablation) | SGEITL with conventional scene-graph predicates · VCR validation | accuracy: 73.1 -> 73.5 (Q→A); 75.6 -> 76.2 (QA→R); 55.3 -> 56.0 (Q→AR) | Helped |
| Use triplet-based scene-graph-aware pre-training (ablation) | SGEITL with SceneGraph+ · VCR validation | accuracy: 73.5 -> 74.4 (Q→A); 76.2 -> 76.9 (QA→R); 56.0 -> 57.2 (Q→AR) | Helped |
| Add multihop graph Transformer (ablation) | SGEITL with SceneGraph+, Pretrain-V · VCR validation | accuracy: 74.4 -> 74.9 (Q→A); 76.9 -> 77.2 (QA→R); 57.2 -> 57.8 (Q→AR) | Helped |
| Use one-hop attention with identity mapping instead of vanilla Transformer (ablation) | SGEITL with Pretrain-V · VCR validation | accuracy: 73.80 -> 73.48 (-0.32) | Hurt |
| Use three-hop attention with identity mapping (ablation) | SGEITL with Pretrain-V · VCR validation | accuracy: 73.80 -> 73.60 (-0.20) | Hurt |
| Use a Gaussian kernel for three-hop attention (ablation) | SGEITL with Pretrain-V · VCR validation | accuracy: 73.80 -> 73.89 (+0.09) | Helped |
| Use a Rational Quadratic kernel for three-hop attention (ablation) | SGEITL with Pretrain-V · VCR validation | accuracy: 73.80 -> 74.91 (+1.11) | Helped |
| Increase Rational Quadratic-kernel attention from three to six hops (ablation) | SGEITL with Pretrain-V · VCR validation | accuracy: 74.91 -> 74.28 | Hurt |
| Add scene-graph predicate features and multihop graph Transformer (ablation) | Baseline · GQA validation | accuracy: 51.63 -> 53.25 (+1.62) | Helped |
| Add scene-graph predicates, multihop graph Transformer, and SceneGraph+ (ablation) | Baseline · SNLI-VE validation | accuracy: 74.06 -> 75.31 (+1.25) | Helped |
Closest papers by meaning
- Scene Graph Thinking: Reinforcing Structured Visual Reasoning for Multimodal Large Language Models (ICML 2026 · similarity 0.90)
- Multimodal Commonsense Knowledge Distillation for Visual Question Answering (Student Abstract) (AAAI 2025 · similarity 0.89)
- Compositional Chain-of-Thought Prompting for Large Multimodal Models (CVPR 2024 · similarity 0.88)
- Visual Graph Arena: Evaluating Visual Conceptualization of Vision and Multimodal Large Language Models (ICML 2025 · similarity 0.88)
- Synthetic Visual Genome (CVPR 2025 · similarity 0.87)
- VGR: Visual Grounded Reasoning (ICLR 2026 · 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.