Research map / Image generation / Text-to-image diffusion
SC-Captioner: Improving Image Captioning with Self-Correction by Reinforcement Learning
Lin Zhang, Xianfang Zeng, Kangcong Li, Gang Yu, Tao Chen
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 SFT on RefinedCaps | Qwen2-VL-7B · DOCCI500 | CAPTURE: 57.96 -> 62.05 | Improved |
| add self-correction RL (ours) after SFT instead of SFT alone | Qwen2-VL-7B · DOCCI500 | CAPTURE: 62.05 -> 63.34 | Improved |
| add self-correction RL (ours) instead of DPO after SFT | Qwen2-VL-7B · DOCCI500 | Relations QA: 27.33 -> 30.51 | Improved |
| add self-correction RL (ours) after SFT | LLaVA-1.5-7B · DOCCI500 | CAPTURE: 60.92 -> 62.29 | Improved |
| add DPO self-correction after SFT | Qwen2-VL-7B · DOCCI500 | Objects precision: 79.99 -> 78.78 | Hurt |
| add prompting-only self-correction to SFT model | Qwen2-VL-7B · DOCCI500 | CAPTURE: 62.05 -> 62.02 | No clear effect |
| use mistake punishment in the reward (ablation) | SC-Captioner reward · DOCCI500 | Objects precision: 77.52 -> 80.20 | Improved |
| use correctness bonus and mistake punishment instead of CAPTURE score as reward (ablation) | SC-Captioner reward · DOCCI500 | CAPTURE: 62.34 -> 63.34 | Improved |
| add object reward component only (ablation) | SC-Captioner reward · DOCCI500 | Objects F1: 69.32 -> 71.44 | Improved |
| reward from second-turn correction instead of first turn only (ablation) | SC-Captioner · DOCCI500 | CAPTURE: 63.12 -> 63.34 | Improved |
| use CLIP-based DiscriTune instead of correction reward (ablation) | SFT Qwen2-VL-7B · DOCCI500 | CAPTURE: 62.05 -> 60.37 | Hurt |
| train on RefinedCaps instead of DOCCI train set (ablation) | Qwen2-VL-7B · COCO-LN500 (cross-domain) | Relations QA: 32.01 -> 38.51 | Improved |
Closest papers by meaning
- CCCaption: Dual-Reward Reinforcement Learning for Complete and Correct Image Captioning (CVPR 2026 · similarity 0.92)
- CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026 · similarity 0.91)
- BalCapRL : A Balanced Framework for RL-Based MLLM Image Captioning (NeurIPS 2026 · similarity 0.89)
- RePIC: Reinforced Post-Training for Personalizing Multi-Modal Language Models (NeurIPS 2025 · similarity 0.86)
- CaptionQA: Is Your Caption as Useful as the Image Itself? (CVPR 2026 · similarity 0.86)
- Painting with Words: Elevating Detailed Image Captioning with Benchmark and Alignment Learning (ICLR 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.
