Research map / Vision-language and multimodal / CLIP and zero-shot
Learning What Matters: Prioritized Concept Learning via Relative Error-driven Sample Selection
Qian Yang, Shivam Chandhok, Oscar Mañas, Kanishk Jain, Aishwarya Agrawal, Leonid Sigal
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 |
|---|---|---|---|
| select clusters by relative accuracy improvement instead of easiest-first (ablation) | PROGRESS · LLaVA-665K 20% | relative performance (%): 92.3 -> 98.8 | Improved |
| select clusters by relative accuracy improvement instead of medium-first (ablation) | PROGRESS · LLaVA-665K 20% | relative performance (%): 91.1 -> 98.8 | Improved |
| select clusters by relative accuracy improvement instead of hardest-first (ablation) | PROGRESS · LLaVA-665K 20% | relative performance (%): 88.5 -> 98.8 | Improved |
| add progress-driven selection after warmup (ablation) | PROGRESS · LLaVA-665K 20% | relative performance (%): 94.6 -> 98.8 | Improved |
| use a random warmup instead of the default warmup (ablation) | PROGRESS · LLaVA-665K 20% | relative performance (%): 98.8 -> 97.1 | Mixed |
| lower softmax temperature from 1.0 to 0.3 (ablation) | PROGRESS · LLaVA-665K 20% | relative performance (%): 98.8 -> 92.8 | Hurt |
| raise softmax temperature to 1.2 (ablation) | PROGRESS · LLaVA-665K 20% | relative performance (%) | Hurt |
| swap COINCIDE features for the unsupervised DINO-BERT features (ablation) | COINCIDE · LLaVA-665K 20% | relative performance (%): 97.8 -> 96.9 | No clear effect |
| use PROGRESS instead of COINCIDE selection | LLaVA-v1.5-7B · LLaVA-665K 20% | relative performance (%): 97.8 -> 98.8 | Improved |
| apply PROGRESS to LLaVA-v1.5-13B without hyperparameter tuning | LLaVA-v1.5-13B · LLaVA-665K 20% | relative performance (%): 95.9 (COINCIDE) -> 96.8 | Improved |
| apply PROGRESS to Qwen2-VL-7B | Qwen2-VL-7B · LLaVA-665K 20% | relative performance (%): 99.6 (COINCIDE) -> 100.0 | Improved |
| apply PROGRESS on the Vision-Flan dataset | LLaVA-v1.5-7B · Vision-Flan-191K 16.7% | relative performance (%): 95.8 (COINCIDE) -> 99.0 | Improved |
Closest papers by meaning
- Adapt-$\infty$: Scalable Continual Multimodal Instruction Tuning via Dynamic Data Selection (ICLR 2025 · similarity 0.89)
- CoIDO: Efficient Data Selection for Visual Instruction Tuning via Coupled Importance-Diversity Optimization (NeurIPS 2025 · similarity 0.89)
- Filter Images First, Generate Instructions Later: Pre-Instruction Data Selection for Visual Instruction Tuning (CVPR 2025 · similarity 0.88)
- Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves (CVPR 2025 · similarity 0.88)
- Active Prompt Learning in Vision Language Models (CVPR 2024 · similarity 0.88)
- Concept-Aware Batch Sampling Improves Language-Image Pretraining (CVPR 2026 · similarity 0.88)
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.
