Research map / Reinforcement learning / Policy gradient and actor-critic
Self-Critical Sequence Training for Image Captioning
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 SCST with the test-time greedy inference reward as the REINFORCE baseline | Att2all captioning model · Karpathy test split | CIDEr: 110.4 -> 113.8 | Helped |
| Use SCST instead of MIXER-CL | Att2all captioning model · Karpathy test split | CIDEr: 113.2 -> 113.8 | Helped |
| Use SCST instead of XE training | FC-2K captioning model · Karpathy test split | CIDEr: 90.9 -> 106.3 | Helped |
| Optimize CIDEr rather than BLEU | FC-2K captioning model · Karpathy test split | CIDEr: 94.4 -> 106.3 | Helped |
| Optimize CIDEr rather than ROUGEL | FC-2K captioning model · Karpathy test split | CIDEr: 97.7 -> 106.3 | Helped |
| Optimize CIDEr rather than METEOR | FC-2K captioning model · Karpathy test split | CIDEr: 80.5 -> 106.3 | Helped |
| Add curriculum learning to SCST training (ablation) | MSCOCO captioning · MSCOCO | CIDEr | No clear effect |
| Use time-dependent SCST instead of the basic SCST baseline (ablation) | MSCOCO captioning · MSCOCO | CIDEr | No clear effect |
| Use True SCST with sampled future tokens and inference completion (ablation) | MSCOCO captioning · MSCOCO | CIDEr | No clear effect |
| Add a learned control variate for the SCST baseline (ablation) | MSCOCO captioning · MSCOCO | CIDEr | No clear effect |
| Feed the attention feature into all LSTM gates, in addition to the cell input (ablation) | Att2in attention model · MSCOCO | captioning performance | No clear effect |
| Feed the attention feature into the gates and outputs (ablation) | Att2all attention model · MSCOCO | captioning performance | Helped |
Closest papers by meaning
- Self-Critical N-Step Training for Image Captioning (CVPR 2019 · similarity 0.91)
- CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026 · similarity 0.88)
- BalCapRL : A Balanced Framework for RL-Based MLLM Image Captioning (NeurIPS 2026 · similarity 0.85)
- Learning by Correction: Efficient Tuning Task for Zero-Shot Generative Vision-Language Reasoning (CVPR 2024 · similarity 0.85)
- TTRV: Test-Time Reinforcement Learning for Vision Language Models (CVPR 2026 · similarity 0.85)
- RePIC: Reinforced Post-Training for Personalizing Multi-Modal Language Models (NeurIPS 2025 · similarity 0.84)
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.