Research map / Reinforcement learning / Visual and representation RL
Meta Learning for Image Captioning
Nannan Li, Zhenzhong Chen, Shan Liu
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 meta learning (separate adaptation steps per task) instead of summing MLE and RL losses (ablation) | Up-Down captioner with CIDEr+SPICE reward · MS COCO Karpathy test | CIDEr: 119.3 -> 121.0 | Improved |
| use meta learning instead of summing MLE and RL losses (ablation) | Up-Down captioner · MS COCO self-retrieval | R@1: 42.6 -> 56.9 | Improved |
| add ground-truth supervision via meta learning to RL (ablation) | RL-trained captioner · MS COCO self-retrieval | R@1: 38.9 -> 56.9 | Improved |
| add ground-truth supervision via meta learning to RL (ablation) | RL-trained captioner · MS COCO SPICE relations | SPICE Relations F-score: 5.8 -> 7.4 | Improved |
| add ground-truth supervision via meta learning to RL (ablation) | RL-trained captioner · MS COCO Karpathy test | SPICE: 21.3 -> 21.7 | Improved |
| add SPICE to the CIDEr reward (ablation) | RL captioner · MS COCO Karpathy test | SPICE/CIDEr: SPICE 21.0 -> 21.3; CIDEr 120.9 -> 120.4 | Mixed |
| naively sum MLE and RL losses (ablation) | RL captioner · MS COCO Karpathy test | CIDEr: 120.4 -> 119.3 | No clear effect |
| RL optimization on top of MLE (ablation) | Up-Down captioner · MS COCO self-retrieval | R@1: 55.6 -> 38.9 | Hurt |
| meta-learned model vs Up-down ensemble on test server | proposed model · MS COCO online test server c5 | CIDEr: 117.9 -> 117.4 | Hurt |
Closest papers by meaning
- BalCapRL : A Balanced Framework for RL-Based MLLM Image Captioning (NeurIPS 2026 · similarity 0.90)
- Temporal-Difference Learning With Sampling Baseline for Image Captioning (AAAI 2018 · similarity 0.89)
- CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026 · similarity 0.89)
- CCCaption: Dual-Reward Reinforcement Learning for Complete and Correct Image Captioning (CVPR 2026 · similarity 0.88)
- Self-Critical N-Step Training for Image Captioning (CVPR 2019 · similarity 0.88)
- RePIC: Reinforced Post-Training for Personalizing Multi-Modal Language Models (NeurIPS 2025 · similarity 0.86)
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.
