Research map / Vision-language and multimodal / Human and embodied modeling
CoCo-IR: Conversational Composed Image Retrieval
Shengcao Cao, Tanmaya Dabral, Zhongli Ding, Madhuri Shanbhogue, Kaifeng Chen, Zhe Li, Mojtaba Seyedhosseini, Liang-Yan Gui, Yu-Xiong 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 LMM-verified hard negatives to mixed-source training data (ablation) | TIE-4B · Fashion IQ | R@10: 33.7 -> 40.1 | Improved |
| add LMM-verified hard negatives to mixed-source training data (ablation) | TIE-4B · CoCo-IR 1-turn | R@1: 67.7 -> 75.4 | Improved |
| add hard negatives to feature-based data (ablation) | TIE-4B · CIRCO | mAP@5: 33.2 -> 38.0 | Improved |
| add hard negatives to metadata-based data (ablation) | TIE-4B · CIRR | R@1: 36.6 -> 34.3 | Hurt |
| add hard negatives to metadata-based data (ablation) | TIE-4B · CoCo-IR 1-turn | R@1: 60.7 -> 56.4 | Hurt |
| mix feature-based and metadata-based data instead of feature-based only (with hard negatives) (ablation) | TIE-4B · Fashion IQ | R@10: 38.0 -> 40.1 | Improved |
| mix feature-based and metadata-based data instead of feature-based only (with hard negatives) (ablation) | TIE-4B · CIRCO | mAP@5: 38.0 -> 37.8 | No clear effect |
| mix sources instead of metadata-only data (no hard negatives) (ablation) | TIE-4B · CoCo-IR 1-turn | R@1: 60.7 -> 67.7 | Improved |
| scale training data from 320K to 2.56M samples (ablation) | TIE-4B · CIRR | R@1: 34.8 -> 37.7 | Improved |
| scale training data from 320K to 2.56M samples (ablation) | TIE-4B · CIRCO | mAP@5: 36.7 -> 37.8 | Mixed |
| scale training data from 1.28M to 2.56M samples (ablation) | TIE-4B · Fashion IQ | R@10: 40.0 -> 40.1 | No clear effect |
| scale model from TIE-4B to TIE-12B | TIE · CoCo-IR 4-turn | R@1: 31.1 -> 44.1 | Improved |
Closest papers by meaning
- Instance-Level Composed Image Retrieval (NeurIPS 2025 · similarity 0.89)
- Context-I2W: Mapping Images to Context-Dependent Words for Accurate Zero-Shot Composed Image Retrieval (AAAI 2024 · similarity 0.89)
- ConText-CIR: Learning from Concepts in Text for Composed Image Retrieval (CVPR 2025 · similarity 0.89)
- INQUIRE: A Natural World Text-to-Image Retrieval Benchmark (NeurIPS 2024 · similarity 0.88)
- Highlighting What Matters: Promptable Embeddings for Attribute-Focused Image Retrieval (NeurIPS 2025 · similarity 0.88)
- Do Composed Image Retrieval Benchmarks Require Multimodal Composition? (NeurIPS 2026 · 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.
