Research map / Vision-language and multimodal / Human and embodied modeling
Exploring Diverse In-Context Configurations for Image Captioning
Xu Yang, Yongliang Wu, Mingzhuo Yang, Haokun Chen, Xin Geng
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 |
|---|---|---|---|
| Use MGC-TF@135 captions as anchors to select ground-truth captions rather than simply taking the first GTC | RS in-context image selection on Open-Flamingo v1 9B · MSCOCO test, 4/8/16/32-shot mean | CIDEr: 80.04 -> 82.51 | Improved |
| Combine SIIR-CLIP retrieval with MGC-TF@135 anchors | Open-Flamingo v1 9B · MSCOCO test, 4/8/16/32-shot mean | CIDEr: +20.9 | Improved |
| Use low-quality MGC-TF@66 captions directly instead of ground-truth captions | SIIR-CLIP in-context image selection on Open-Flamingo v1 9B · MSCOCO test | CIDEr: 90.62 -> 66.79 | Hurt |
| Use MGC-VLM(0)@63 instead of syntactically error-prone MGC-TF@66 | Open-Flamingo v1 9B caption assignment · MSCOCO test | CIDEr: 64.24 -> 74.05 | Improved |
| Use MGC-TF@135 anchors instead of first ground-truth captions | Six image selection strategies on Open-Flamingo v1 9B · MSCOCO test | CIDEr: +8.8 | Improved |
| Use MGC-TF@66 as anchors to select GTCs rather than use the first GTC | RS in-context image selection on Open-Flamingo v1 9B · MSCOCO test | CIDEr: 80.04 -> 83.24 | Improved |
| Use similarity-based image retrieval rather than random selection with low-quality captions | MGC-TF@66 caption assignment on Open-Flamingo v1 9B · MSCOCO test | CIDEr: 64.24 -> 66.79 | Mixed |
| Use similar in-context images rather than random images while retaining unrelated captions | Open-Flamingo v1 9B · MSCOCO test; similarity to in-context captions | CIDEr: 54.8 -> 192. 2 | Improved |
| Use similar in-context images rather than random images while retaining unrelated captions | Open-Flamingo v1 9B · MSCOCO test; similarity to test-image ground truth | CIDEr: 58.5 -> 6.7 | Hurt |
| Use more iterations of generated-caption prompting after iteration 3 (ablation) | MGC-VLM(0) iterative prompting on Open-Flamingo v1 9B · MSCOCO test, successive iterations | CIDEr: 79.9 -> 79.3 -> 77.3 | Hurt |
| Use more iterations of generated-caption prompting after iteration 2 (ablation) | MGC-VLM(32) iterative prompting on Open-Flamingo v1 9B · MSCOCO test, successive iterations | CIDEr: 80.5 -> 79.4 -> 78.9 -> 77.1 | Hurt |
| Use DIIR-TT rather than DIIR-TR for diverse image retrieval (ablation) | Open-Flamingo v1 9B with GTC captions · MSCOCO test, 4/8/16/32-shot mean | CIDEr: 84.82 -> 86.00 | Improved |
Closest papers by meaning
- VL-ICL Bench: The Devil in the Details of Multimodal In-Context Learning (ICLR 2025 · similarity 0.90)
- How to Configure Good In-Context Sequence for Visual Question Answering (CVPR 2024 · similarity 0.89)
- CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026 · similarity 0.88)
- Semantic and Expressive Variations in Image Captions Across Languages (CVPR 2025 · similarity 0.88)
- Modeling Caption Diversity in Contrastive Vision-Language Pretraining (ICML 2024 · similarity 0.87)
- Large Language Models are Good Prompt Learners for Low-Shot Image Classification (CVPR 2024 · 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.
