Research map / Vision-language and multimodal / Multimodal LLMs
MA-CIR: A Multimodal Arithmetic Benchmark for Composed Image Retrieval
Jaeseok Byun, Young Kyun Jang, Seokhyeon Jeong, Donghyun Kim, Taesup Moon
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 |
|---|---|---|---|
| fine-tune E5-V with MA-CIR-tailored text triplets with hard negatives instead of NLI pairs | E5-V · MA-CIR | Avg R@1 (%): 32.4 -> 48.5 | Improved |
| fine-tune E5-V with MA-CIR-tailored text triplets | E5-V · CIRCO | mAP@5 (%): 17.6 -> 26.5 | Improved |
| fine-tune E5-V with MA-CIR-tailored text triplets | E5-V · CIRR | R@1 (%): 30.9 -> 36.8 | Improved |
| add hard negatives to our text triplets (ablation) | adapted E5-V · MA-CIR | Avg R@1 (%): 43.9 -> 48.5 | Improved |
| add hard negatives to default RTD triplets (ablation) | adapted E5-V · MA-CIR | Avg R@1 (%): 38.5 -> 30.2 | Hurt |
| use our BLIP-style LLM triplets instead of default RTD triplets (no hard negatives) (ablation) | adapted E5-V · MA-CIR | Avg R@1 (%): 38.5 -> 43.9 | Improved |
| use default RTD triplets instead of NLI pairs (ablation) | E5-V · MA-CIR / CIRR | Avg R@1 (%): 32.4 -> 38.5 | Mixed |
| scale CLIP backbone up to ViT-G/14 (ablation) | Slerp and LinCIR · MA-CIR | Avg R@1 (%): less than 3% vs over 10% on CIRR | No clear effect |
| use compositionality fine-tuned CLIP (FSC-CLIP, NegCLIP) (ablation) | Slerp and LinCIR · MA-CIR | Avg R@1 (%): over 9% on CIRR, little to no gain on MA-CIR | No clear effect |
| use LLM-based text encoder (E5-V) instead of CLIP | ZS-CIR · MA-CIR | Avg R@1 (%): 23.3 -> 32.4 | Improved |
Closest papers by meaning
- Do Composed Image Retrieval Benchmarks Require Multimodal Composition? (NeurIPS 2026 · similarity 0.92)
- ConText-CIR: Learning from Concepts in Text for Composed Image Retrieval (CVPR 2025 · similarity 0.91)
- Vision-by-Language for Training-Free Compositional Image Retrieval (ICLR 2024 · similarity 0.91)
- Reason-before-Retrieve: One-Stage Reflective Chain-of-Thoughts for Training-Free Zero-Shot Composed Image Retrieval (CVPR 2025 · similarity 0.91)
- Instance-Level Composed Image Retrieval (NeurIPS 2025 · similarity 0.91)
- RMIR: A Benchmark Dataset for Reasoning-Intensive Multimodal Image Retrieval (CVPR 2026 · similarity 0.90)
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.
