Research map / Vision-language and multimodal / Multimodal LLMs
RegRet: Enhancing Region-Level Retrieval in Large Multimodal Models
Xun Liang, Honghui Yang, Weihang Pan, Boyuan Pan, Yao Hu, Binbin Lin, Deng Cai, Ruisi Zhao, Wenxiao Wang
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 layer-wise coordination between CE and RAE features (ablation) | RegRet-8B RAE · REGMB | average recall: 82.6 -> 86.2 | Improved |
| use separate projectors for RAE and context encoder (ablation) | RegRet-8B RAE · REGMB | average recall: 66.0 -> 82.6 | Improved |
| use separate projectors for RAE and context encoder (ablation) | RegRet-8B RAE · M-BEIR | average recall: 27.2 -> 57.1 | Improved |
| replace auxiliary-image prompting with RAE (ablation) | RegRet-8B · REGMB | average recall: 83.4 -> 86.2 | Improved |
| add regional contrastive training data (REGMB) to 3B model (ablation) | RegRet-3B · REGMB | average recall: 69.4 -> 79.9 | Improved |
| add image-level (M-BEIR) data to text-only training (ablation) | RegRet-3B · REGMB | average recall: 68.4 -> 69.4 | Mixed |
| train LamRA on REGMB (ablation) | LamRA · REGMB | average recall: 67.7 -> 80.9 | Improved |
| add regional contrastive training to zero-shot RegRet-8B | RegRet-8B · Public region benchmarks | average: 48.3 -> 56.1 | Improved |
| use RAE instead of prompting strategies (crop, visual mark, auxiliary image) on LamRA (ablation) | zero-shot REGMB · REGMB | average recall: 63.1 -> 78.8 | Improved |
| mixed visual prompting in Stage III (RAE-only for tasks 1-3, CE+RAE for task 4) (ablation) | RegRet-3B · REGMB | average recall: 78.4 -> 79.9 | Improved |
| train RegRet on region data, global retrieval | RegRet-8B vs LamRA · M-BEIR | average recall: 57.1 -> 57.3 | No clear effect |
| RegRet-8B vs RzenEmbed on context-level xgoods task | RegRet-8B · RegMB Task 4 xgoods | R@5: 98.2 -> 93.6 | Hurt |
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.
- LamRA — LamRA: Large Multimodal Model as Your Advanced Retrieval Assistant (CVPR 2025) · better on 3 of 3 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
- LamRA: Large Multimodal Model as Your Advanced Retrieval Assistant (CVPR 2025 · similarity 0.89)
- Your Embedding Model Is SMARTer Than You Think (NeurIPS 2026 · similarity 0.89)
- MRAG-Bench: Vision-Centric Evaluation for Retrieval-Augmented Multimodal Models (ICLR 2025 · similarity 0.89)
- Beyond Global Similarity: Multi-Conditional Retrieval for Fine-Grained Cross-Modal Understanding (CVPR 2026 · similarity 0.89)
- RMIR: A Benchmark Dataset for Reasoning-Intensive Multimodal Image Retrieval (CVPR 2026 · similarity 0.89)
- U-MARVEL: Unveiling Key Factors for Universal Multimodal Retrieval via Embedding Learning with MLLMs (ICLR 2026 · 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.
