Research map / Vision-language and multimodal / Multimodal LLMs
Concept-as-Tree: A Controllable Synthetic Data Framework Makes Stronger Personalized VLMs
Ruichuan An, Kai Zeng, Ming Lu, Sihan Yang, Renrui Zhang, Huitong Ji, Hao Liang, Wentao 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 |
|---|---|---|---|
| add synthetic data (CaT + PCS filter) to real data | MC-LLaVA · MC-LLaVA eval | Recognition accuracy: 0.917 -> 0.963 | Improved |
| add synthetic data to real data | Yo'LLaVA · MC-LLaVA eval | Recognition accuracy: 0.841 -> 0.885 | Improved |
| add synthetic data to real data | MyVLM · MC-LLaVA eval | Recognition accuracy: 0.795 -> 0.827 | Improved |
| train only on synthetic data instead of real | MyVLM · MC-LLaVA eval | Recognition accuracy: 0.795 -> 0.788 | Mixed |
| add synthetic data in multi-concept setting | MC-LLaVA · MC-LLaVA eval (multi-concept) | Recognition accuracy: 0.845 -> 0.881 | Improved |
| add PCS score filtering over no filtering (ablation) | Yo'LLaVA + CaT · MC-LLaVA eval | Recognition accuracy: 0.835 -> 0.885 | Improved |
| add PCS score filtering over cosine similarity filtering (ablation) | Yo'LLaVA + CaT · MC-LLaVA eval | Recognition accuracy: 0.862 -> 0.885 | Improved |
| add unfiltered synthetic data (ablation) | Yo'LLaVA · MC-LLaVA eval | Recognition accuracy: 0.841 -> 0.835 | Hurt |
| add attributes to the concept tree (1 to 2 Add operations) (ablation) | Yo'LLaVA + CaT · MC-LLaVA eval | Recognition accuracy: 0.841 -> 0.866 | Improved |
| add a third Add operation to the tree (ablation) | Yo'LLaVA + CaT · MC-LLaVA eval | Recognition accuracy: 0.866 -> 0.834 | Hurt |
| remove attributes from the concept tree (ablation) | Yo'LLaVA + CaT · MC-LLaVA eval | Recognition accuracy: 0.841 -> 0.801 | Hurt |
| add synthetic positives (ablation) | Yo'LLaVA · MC-LLaVA eval | Recognition accuracy: 0.841 -> 0.858 | Improved |
Closest papers by meaning
- MMPB: It’s Time for Multi-Modal Personalization (NeurIPS 2025 · similarity 0.88)
- UniCTokens: Boosting Personalized Understanding and Generation via Unified Concept Tokens (NeurIPS 2025 · similarity 0.85)
- Concept-Aware Batch Sampling Improves Language-Image Pretraining (CVPR 2026 · similarity 0.85)
- Rethinking Model Selection in VLM Through the Lens of Gromov-Wasserstein Distance (CVPR 2026 · similarity 0.85)
- Vision-Language Model Selection and Reuse for Downstream Adaptation (ICML 2025 · similarity 0.85)
- BLEND: Balancing Personalization vs. Generalization in Federated Vision–Language Models (NeurIPS 2026 · similarity 0.85)
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.
