Research map / Vision-language and multimodal / Multimodal LLMs
Unbalanced Optimal Transport for Efficient Visual Document Retrieval
Hoyeon Shin, Jeongyeon Kim, Yeong Jun Koh, Yeoneung Kim, Hanul Kim
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 |
|---|---|---|---|
| replace Light-ColQwen2 hard clustering with UOT compression at MF=50 | ColQwen2 · ViDoRe V3 cross-lingual | average nDCG@5: +11.8 | Improved |
| replace Light-ColQwen2.5 clustering with UOT compression at MF=50 | ColQwen2.5 · ViDoRe V3 cross-lingual | average nDCG@5: +19.0 | Improved |
| replace Light-ColQwen2 with UOT compression at MF=50 | ColQwen2 · ViDoRe V2 | average nDCG@5: +13.9 | Improved |
| replace Light-ColQwen2.5 with UOT compression at MF=50 | ColQwen2.5 · ViDoRe V2 | average nDCG@5: +23.7 | Improved |
| UOT compression versus pruning at MF=5 (Qwen2.5-VL-3B) | ColQwen2.5 · ViDoRe V2 | average nDCG@5: 49.5 -> 53.1 | Improved |
| UOT compression versus pruning at MF=5 (Qwen2-VL-2B) | ColQwen2 · ViDoRe V3 cross-lingual | average nDCG@5: 47.6 -> 48.1 | Mixed |
| use hierarchical clustering to compress rather than UOT at MF=25 (V3 Qwen2) | Light-ColQwen2 · ViDoRe V3 cross-lingual | average nDCG@5: 36.2 -> 42.9 | No clear effect |
| increase compression factor from 25 to 50 | UOT method · Indexing time | ms/doc: 11.4 -> 9.4 | Improved |
| increase compression factor from 25 to 50 | Light-ColQwen2 · Indexing time | ms/doc: 196.0 -> 200.1 | Hurt |
| lower marginal penalty rho toward 0 (ablation) | UOT compression · ViDoRe V1 academic | retrieval accuracy | Hurt |
| relax rho from 100 (balanced OT) to 5 (ablation) | UOT compression · ViDoRe V1 academic | retrieval accuracy | Improved |
| increase entropy gamma to 0.06 or higher (ablation) | UOT compression · ViDoRe V1 academic | retrieval accuracy | Hurt |
Closest papers by meaning
- ApET: Approximation-Error Guided Token Compression for Efficient VLMs (CVPR 2026 · similarity 0.88)
- Prune Redundancy, Preserve Essence: Vision Token Compression in VLMs via Synergistic Importance-Diversity (ICLR 2026 · similarity 0.88)
- ColPali: Efficient Document Retrieval with Vision Language Models (ICLR 2025 · similarity 0.87)
- CORE: Compact Object-centric REpresentations as a New Paradigm for Token Merging in LVLMs (CVPR 2026 · similarity 0.87)
- EvoComp: Learning Visual Token Compression for Multimodal Large Language Models via Semantic-Guided Evolutionary Labeling (CVPR 2026 · similarity 0.87)
- LensVLM: Selective Context Expansion for Compressed Visual Representation of Text (NeurIPS 2026 · 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.
