Research map / Vision-language and multimodal / Human and embodied modeling
ModernVBERT: Towards Smaller Visual Document Retrievers
Paul Teiletche, Quentin Macé, Max Conti, António Loison, Gautier Viaud, Pierre Colombo, Manuel Faysse
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 |
|---|---|---|---|
| use bidirectional attention (MLM encoder) instead of causal decoder for late-interaction retrieval (ablation) | ModernVBERT ablation models · ViDoRe multi-vector | nDCG@5: 54.2 -> 64.8 (+10.6) | Improved |
| use bidirectional attention instead of causal attention for single-vector retrieval (ablation) | ModernVBERT ablation models · ViDoRe single-vector | nDCG@5 | No clear effect |
| remove causal mask / switch to MLM only during later training (dec-to-enc, dec with MLM) (ablation) | decoder-based retriever · ViDoRe multi-vector | nDCG@5 | No clear effect |
| early fusion of vision encoder with pretrained language model vs fine-tuned SigLIP2 dual encoder (ablation) | vision-language retriever · document retrieval vs natural image tasks | nDCG@5 / accuracy: +6.1 nDCG@5 on document retrieval; worse on natural image tasks | Mixed |
| scale modality alignment data (ablation) | enc · document retrieval vs natural images | retrieval score | Mixed |
| increase input image resolution from 512px to 2048px (ablation) | enc · document retrieval | nDCG@5: 30.7 -> 43.8 | Improved |
| increase input image resolution from 512px to 2048px (ablation) | enc · image classification | accuracy: 41.4 -> 33.9 | Hurt |
| add high-resolution cooldown at end of modality alignment (ablation) | enc 2048px · document retrieval | nDCG@5: 43.8 -> 45.8 | Improved |
| add text-only query-document pairs to contrastive mix (ablation) | enc · document retrieval | nDCG@5: 43.9 -> 45.6 | Improved |
| add text-only query-document pairs to contrastive mix (ablation) | enc · image/caption retrieval | score: 57.2 -> 53.2 | Hurt |
| add NatCap synthetic image-caption pairs (ablation) | enc · image classification | accuracy: 35.7 -> 49.9 | Improved |
| add NatCap synthetic image-caption pairs (ablation) | enc · document retrieval | nDCG@5: 45.6 -> 45.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- ColPali — ColPali: Efficient Document Retrieval with Vision Language Models (ICLR 2025) · better on 1 of 4 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
- ColPali: Efficient Document Retrieval with Vision Language Models (ICLR 2025 · similarity 0.89)
- Document Haystacks: Vision-Language Reasoning Over Piles of 1000+ Documents (CVPR 2025 · similarity 0.88)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 2025 · similarity 0.87)
- Revisiting the Role of Language Priors in Vision-Language Models (ICML 2024 · similarity 0.87)
- Towards Natural Language-Based Document Image Retrieval: New Dataset and Benchmark (CVPR 2025 · similarity 0.86)
- Your Embedding Model Is SMARTer Than You Think (NeurIPS 2026 · similarity 0.86)
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.
