Research map / Vision-language and multimodal / Multimodal LLMs
Recurrence Meets Transformers for Universal Multimodal Retrieval
Davide Caffagni, Sara Sarto, Marcella Cornia, Lorenzo Baraldi, Rita Cucchiara
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 score fusion in place of late interaction (ablation) | ReT · M2KR (CLIP ViT-L) | average recall: 61.5 -> 61.4 | No clear effect |
| share weights between query and document encoders (ablation) | ReT with score fusion · M2KR (CLIP ViT-L) | average recall: +1.2 points | Improved |
| reduce tokens from 32 to a single token (ablation) | shared ReT · M2KR (CLIP ViT-L) | average recall: +0.2 points | Improved |
| add layer pruning to three layers (ablation) | shared single-token ReT · M2KR (CLIP ViT-L) | average recall: +0.1 points | Improved |
| add global feature injection (ablation) | ReT-2 · M2KR (CLIP ViT-L) | average recall: 62.9 -> 64.1 | Improved |
| add global features to ReT (ablation) | ReT · M2KR (CLIP ViT-L) | average recall: 61.5 -> 62.0 | Improved |
| add recurrent cell (ablation) | ReT-2 · M2KR (CLIP ViT-L) | average recall: 57.3 -> 64.1 | Improved |
| add shared architecture (ablation) | ReT-2 · M2KR (CLIP ViT-L) | average recall: -4.0 points when removed | Improved |
| add score fusion in place of late interaction (ablation) | ReT-2 · M2KR (CLIP ViT-L) | average recall: -2.8 points when removed | Improved |
| add token reduction from 32 tokens (ablation) | ReT-2 · M2KR (CLIP ViT-L) | average recall: 63.6 -> 64.1 | No clear effect |
| add layer pruning (ablation) | ReT-2 · M2KR (CLIP ViT-L) | average recall: 63.8 -> 64.1 | No clear effect |
| combine early, middle and late layers instead of a single depth (ablation) | ReT-2 · M2KR (CLIP ViT-L) | average recall: 62.3 -> 64.1 | 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.
- LamRA — LamRA: Large Multimodal Model as Your Advanced Retrieval Assistant (CVPR 2025) · better on 0 of 1 shared result
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
- Recurrence-Enhanced Vision-and-Language Transformers for Robust Multimodal Document Retrieval (CVPR 2025 · similarity 0.95)
- LamRA: Large Multimodal Model as Your Advanced Retrieval Assistant (CVPR 2025 · similarity 0.90)
- Beyond Global Similarity: Multi-Conditional Retrieval for Fine-Grained Cross-Modal Understanding (CVPR 2026 · similarity 0.89)
- Your Embedding Model Is SMARTer Than You Think (NeurIPS 2026 · similarity 0.89)
- MM-EMBED: UNIVERSAL MULTIMODAL RETRIEVAL WITH MULTIMODAL LLMS (ICLR 2025 · similarity 0.88)
- TIGeR: Unifying Text-to-Image Generation and Retrieval with Large Multimodal Models (ICLR 2025 · similarity 0.88)
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.
