Research map / Vision-language and multimodal / Multimodal LLMs
Dynamic Cluster Data Sampling for Efficient and Long-Tail-Aware Vision-Language Pre-training
Mingliang Liang, Zhuoran Liu, Arjen P. de Vries, Martha Larson
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 cluster scaling (up/down-sampling) to static random pruning (ablation) | random pruning · ImageNet-1K zero-shot (DataComp, 0.64B seen) | top-1 accuracy: 64.5 -> 68.0 | Improved |
| add cluster scaling to dynamic random sampling (ablation) | Random-Dynamic · Let It Wag! zero-shot (DataComp, 0.64B seen) | top-1 accuracy: 36.2 -> 46.5 | Improved |
| add dynamic per-epoch sampling to cluster scaling (ablation) | Cluster-Scaling · ImageNet-1K zero-shot (DataComp, 0.64B seen) | top-1 accuracy: 68.0 -> 69.2 | Improved |
| add dynamic per-epoch sampling to random pruning (ablation) | random pruning · ImageNet-1K zero-shot (DataComp, 0.64B seen) | top-1 accuracy: 64.5 -> 66.2 | Improved |
| add dynamic per-epoch sampling to random pruning (ablation) | random pruning · Let It Wag! zero-shot (DataComp, 0.64B seen) | top-1 accuracy: 35.5 -> 36.2 | Improved |
| use moderate alpha 0.2 instead of flat alpha 0.0 (ablation) | DynamiCS · ImageNet-1K zero-shot (DataComp, 106M seen, 112px) | top-1 accuracy: 38.5 -> 39.2 | Improved |
| raise alpha from 0.2 to 2.0 (ablation) | DynamiCS · ImageNet-1K zero-shot (DataComp, 106M seen, 112px) | top-1 accuracy: 39.2 -> 19.4 | Hurt |
| merge close centroids with cosine threshold 0.7 instead of none (1.0) (ablation) | DynamiCS · ImageNet-1K zero-shot (DataComp, 106M seen, 112px) | top-1 accuracy: 34.9 -> 39.2 | Improved |
| vary number of clusters from 10k to 70k (ablation) | DynamiCS · ImageNet-1K zero-shot (DataComp, 106M seen, 112px) | top-1 accuracy: 39.0 -> 39.2 | No clear effect |
| increase fine-tuning batch size from 4k to 32k (ablation) | DynamiCS-1.28B · ImageNet-1K zero-shot (DataComp) | top-1 accuracy: 71.1 -> 71.7 | Improved |
| add 128M samples of 224px fine-tuning after 112px pre-training (ablation) | DynamiCS 1.28B (LAION-400M) · ImageNet-1K zero-shot | top-1 accuracy: 60.7 -> 65.0 | Improved |
| scale samples seen from 0.64B to 2.56B | DynamiCS (DataComp-DFN) · ImageNet-1K zero-shot | top-1 accuracy: 69.2 -> 72.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- HQ-CLIP — HQ-CLIP: Leveraging Large Vision-Language Models to Create High-Quality Image-Text Datasets and CLIP Models (ICCV 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Concept-Aware Batch Sampling Improves Language-Image Pretraining (CVPR 2026 · similarity 0.90)
- From Head to Tail: Towards Balanced Representation in Large Vision-Language Models through Adaptive Data Calibration (CVPR 2025 · similarity 0.88)
- Unified Generative and Discriminative Training for Multi-modal Large Language Models (NeurIPS 2024 · similarity 0.87)
- Connecting the Dots: Collaborative Fine-tuning for Black-Box Vision-Language Models (ICML 2024 · similarity 0.87)
- Data Selection for Fine-tuning Vision Language Models via Cross Modal Alignment Trajectories (ICML 2026 · similarity 0.87)
- Scalable Pre-training of Large Autoregressive Image Models (ICML 2024 · 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.
