Research map / Vision-language and multimodal / Cross-modal learning
CMDR: Contextual Multimodal Document Retrieval
Ryota Tanaka, Taku Hasegawa, Kyosuke Nishida
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 contextual joint multi-page encoding and CMCL instead of finetuning only (ablation) | ColQwen · CMDR-Bench | nDCG@5 overall: 40.4 -> 56.6 | Improved |
| add contextual joint multi-page encoding and CMCL instead of finetuning only (ablation) | ColPali · CMDR-Bench | nDCG@5 overall: 36.3 -> 49.8 | Improved |
| add contextual joint encoding and CMCL (ablation) | ColQwen · CMDR-Bench | Recall@5 overall: 58.8 -> 75.1 | Improved |
| finetune backbone on CMDR-Synth (ablation) | ColQwen · CMDR-Bench | nDCG@5 overall: 33.2 -> 40.4 | Improved |
| add CMCL loss with in-chunk and in-document hard negatives (ablation) | CMDR-Embed · CMDR-Bench | nDCG@5 | Improved |
| use late interaction multi-vector instead of single-vector mean pooling (ablation) | CMDR-Embed (ColPali) · CMDR-Bench | nDCG@5: 19.1-point drop with last-token pooling | Improved |
| add Qwen3-VL reranker to the contextual model (ablation) | CMDR-Embed (ColPali) · CMDR-Bench | nDCG@5 overall: 49.8 -> 41.0 | Hurt |
| add Qwen3-VL reranker to non-contextual model (ablation) | ColPali · CMDR-Bench | nDCG@5 overall: 32.2 -> 39.5 | Improved |
| use GOT-OCR2.0 instead of Tesseract OCR (ablation) | BGE text retriever · CMDR-Bench | nDCG@5 overall: 29.8 -> 30.0 | No clear effect |
| add ViDoRe data via multitask learning (ablation) | CMDR-Embed (ColPali) · ViDoRe | nDCG@5 average: 80.9 -> 85.6 | Improved |
| add ViDoRe data via multitask learning (ablation) | CMDR-Embed (ColPali) · CMDR-Bench | nDCG@5 overall: 49.8 -> 47.1 | No clear effect |
| apply hierarchical token pooling factor 5 (ablation) | CMDR-Embed (ColPali) · CMDR-Bench | vectors / speed / performance: 80.0% fewer vectors, 3.57x faster search, 97.3% performance retained | Mixed |
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.
- ColPali — ColPali: Efficient Document Retrieval with Vision Language Models (ICLR 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
- Contextual Document Embeddings (ICLR 2025 · similarity 0.88)
- MetaEmbed: Scaling Multimodal Retrieval at Test-Time with Flexible Late Interaction (ICLR 2026 · similarity 0.87)
- MM-EMBED: UNIVERSAL MULTIMODAL RETRIEVAL WITH MULTIMODAL LLMS (ICLR 2025 · similarity 0.87)
- Your Embedding Model Is SMARTer Than You Think (NeurIPS 2026 · similarity 0.86)
- Recurrence-Enhanced Vision-and-Language Transformers for Robust Multimodal Document Retrieval (CVPR 2025 · similarity 0.85)
- Bridging Modalities: Improving Universal Multimodal Retrieval by Multimodal Large Language Models (CVPR 2025 · 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.
