Research map / Vision-language and multimodal / Cross-modal learning
Uncovering Grounding IDs: How External Cues Shape Multi-Modal Binding
Amirmohammad Izadi, Hosein Hasani, Fatemeh Askari, Mobin Bagherian, Sadegh Mohammadian, Mohammad Izadi, Mahdieh Soleymani Baghshah
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 aligned symbol cues to both image and prompt (vs. text-only or image-only cues) (ablation) | Qwen2.5-VL-7B · Synthetic scene description, 20 objects | F1: 0.21 -> 0.62 | Improved |
| add cues to the image only (ablation) | Qwen2.5-VL-7B · Synthetic scene description, 10 objects | accuracy: 0.42 -> 0.38 | Hurt |
| add structured cues with Grounding IDs instead of VISER | Qwen2.5-VL-7B · Counting | accuracy: 43.33 -> 53.00 | Improved |
| add structured cues to image and prompt | LLaVA-1.5 · MS-COCO CHAIR | CHAIR_s: 51.60 -> 41.00 | Improved |
| add structured cues to image and prompt | GPT-4o · MS-COCO CHAIR | CHAIR_s: 29.20 -> 23.20 | Improved |
| use row-level symbols with lines rather than per-object symbols (ablation) | Qwen2.5-VL-7B · Synthetic variants, 20 objects | accuracy: 0.19 -> 0.40 | Improved |
| use row separator lines in addition to symbols (ablation) | Qwen2.5-VL-7B · Synthetic variants, 20 objects | accuracy: 0.33 -> 0.40 | Improved |
| use 4x4 grid layout instead of row partitions on synthetic images (ablation) | Qwen2.5-VL-7B · Synthetic variants, 20 objects | accuracy: 0.40 -> 0.22 | Hurt |
| add structured cues on POPE | Qwen2.5-VL · POPE Random | accuracy: 89.73 -> 85.67 | Hurt |
| add structured cues on POPE | LLaVA-1.5 · POPE Random | accuracy: 87.27 -> 89.23 | Improved |
| ensemble five partition layouts with majority voting | Qwen2.5-VL structured · POPE Adversarial | accuracy: 83.73 -> 85.72 | Improved |
| swap object activations between contexts (causal test) | Qwen2.5-VL-7B · Activation swap | accuracy: standard 1.00 -> 0.02; swap accuracy 0.98 | Mixed |
Closest papers by meaning
- GroundingME: Exposing the Visual Grounding Gap in MLLMs through Multi-Dimensional Evaluation (CVPR 2026 · similarity 0.90)
- PostAlign: Multimodal Grounding as a Corrective Lens for MLLMs (ICLR 2026 · similarity 0.90)
- Grounding Multimodal Large Language Models to the World (ICLR 2024 · similarity 0.89)
- Understanding Language Prior of LVLMs by Contrasting Chain-of-Embedding (ICLR 2026 · similarity 0.89)
- Seeing is Understanding: Unlocking Causal Attention into Modality-Mutual Attention for Multimodal LLMs (ICML 2026 · similarity 0.89)
- Your Large Vision-Language Model Only Needs A Few Attention Heads For Visual Grounding (CVPR 2025 · similarity 0.89)
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.
