Research map / Vision-language and multimodal / Human and embodied modeling
2.5 Years in Class: A Multimodal Textbook for Vision-Language Pretraining
Wenqi Zhang, Hang Zhang, Xin Li, Jiashuo Sun, Yongliang Shen, Weiming Lu, Deli Zhao, Yueting Zhuang, Lidong Bing
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 LLM refinement of raw ASR text (ablation) | Multimodal Textbook · 7-benchmark 1-shot average | accuracy: 26.2 -> 31.1 | Improved |
| add OCR text from keyframes (ablation) | Multimodal Textbook · 7-benchmark 1-shot average | accuracy: 28.8 -> 31.1 | Improved |
| use SSIM instead of pixel-level keyframe extraction (ablation) | Multimodal Textbook · 7-benchmark 1-shot average | accuracy: 22.1 -> 31.1 | Improved |
| use SSIM instead of CLIP-based keyframe extraction (ablation) | Multimodal Textbook · 7-benchmark 1-shot average | accuracy: 24.6 -> 31.1 | Improved |
| pretrain on Textbook-6.5M instead of OBELICS | Idefics2-8B-base continual · MathVista | accuracy: 27.6 -> 29.7 | Improved |
| pretrain on Textbook-6.5M instead of MMC4 | LLaVA-1.5-7B · ScienceQA-IMG 4-shot | accuracy: 11.6 -> 37.3 | Improved |
| pretrain on Textbook-6.5M instead of OBELICS | LLaVA-1.5-7B · OKVQA 0-shot | accuracy: 13.0 -> 10.2 | Hurt |
| pretrain on Textbook-6.5M instead of OBELICS | LLaVA-1.5-7B · OKVQA 4-shot | accuracy: 37.5 -> 39.9 | Improved |
| pretrain on Textbook-6.5M instead of OBELICS | Idefics2-8B from scratch · OKVQA | accuracy: 10.5 -> 10.1 | No clear effect |
| pretrain on Textbook-6.5M instead of OBELICS | Idefics2-8B-base · MathVista 1-shot cheat test | accuracy: 67.7 -> 94.1 | Improved |
| pretrain on Textbook-6.5M instead of OBELICS then SFT | LLaVA-1.5-7B · MathVista after SFT | accuracy: 25.6 -> 28.7 | Improved |
| shuffle image order in the dataset (20%, 50%, 100%) (ablation) | Textbook-6.5M · 7-benchmark average | accuracy | Hurt |
Closest papers by meaning
- Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-training (ICLR 2026 · similarity 0.90)
- Exploring Visual Pretraining for Learning Language Intelligence (CVPR 2026 · similarity 0.89)
- VILA: On Pre-training for Visual Language Models (CVPR 2024 · similarity 0.89)
- MMIE: Massive Multimodal Interleaved Comprehension Benchmark for Large Vision-Language Models (ICLR 2025 · similarity 0.89)
- The Narrow Gate: Localized Image-Text Communication in Native Multimodal Models (NeurIPS 2025 · similarity 0.89)
- PerceptionLM: Open-Access Data and Models for Detailed Visual Understanding (NeurIPS 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.
