Research map / Reinforcement learning / RL from feedback
Partial Off-Policy Learning: Balance Accuracy and Diversity for Human-Oriented Image Captioning
Jiahe Shi, Yali Li, Shengjin 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 max-CIDEr reward optimization (ablation) | Up-Down baseline · MSCOCO Karpathy test | self-CIDEr: 16.0 -> 20.7 | Improved |
| add max-CIDEr reward optimization (ablation) | Up-Down baseline · MSCOCO Karpathy test | CIDEr: 111.9 -> 111.4 | No clear effect |
| add max-CIDEr reward optimization (ablation) | Up-Down baseline · MSCOCO Karpathy test | Pearson correlation: 0.253 -> 0.266 | Improved |
| add off-policy sampling model (initial model as behavior policy) (ablation) | Up-Down baseline · MSCOCO Karpathy test | self-CIDEr: 16.0 -> 30.6 | Improved |
| add off-policy sampling model (ablation) | Up-Down baseline · MSCOCO Karpathy test | CIDEr: 111.9 -> 107.2 | Hurt |
| add off-policy sampling model (ablation) | Up-Down baseline · MSCOCO Karpathy test | Pearson correlation: 0.253 -> 0.251 | No clear effect |
| combine max-CIDEr and sampling model (full method) (ablation) | Up-Down baseline · MSCOCO Karpathy test | Pearson correlation: 0.253 -> 0.337 | Improved |
| combine max-CIDEr and sampling model (full method) (ablation) | Up-Down baseline · MSCOCO Karpathy test | self-CIDEr: 16.0 -> 69.3 | Improved |
| combine max-CIDEr and sampling model (full method) (ablation) | Up-Down baseline · MSCOCO Karpathy test | CIDEr: 111.9 -> 89.9 | Hurt |
| raise the on-policy selection probability, eps 0.1 -> 0.01 (ablation) | Partial off-policy learning · MSCOCO Karpathy test | CIDEr: 89.9 -> 111.9 | Improved |
| raise off-policy probability eps 0.1 -> 0.9 (ablation) | Partial off-policy learning · MSCOCO Karpathy test | self-CIDEr: 69.3 -> 80.6 | Mixed |
| apply partial off-policy learning instead of on-policy training | AoA · MSCOCO Karpathy test | self-CIDEr: 51.6% | Improved |
Closest papers by meaning
- BalCapRL : A Balanced Framework for RL-Based MLLM Image Captioning (NeurIPS 2026 · similarity 0.88)
- CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026 · similarity 0.88)
- RePIC: Reinforced Post-Training for Personalizing Multi-Modal Language Models (NeurIPS 2025 · similarity 0.87)
- Modeling Caption Diversity in Contrastive Vision-Language Pretraining (ICML 2024 · similarity 0.86)
- Learning by Correction: Efficient Tuning Task for Zero-Shot Generative Vision-Language Reasoning (CVPR 2024 · similarity 0.86)
- Painting with Words: Elevating Detailed Image Captioning with Benchmark and Alignment Learning (ICLR 2025 · similarity 0.85)
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.
