Research map / Large language models / Alignment and preferences
Teaching Machines to Describe Images with Natural Language Feedback
Read the paperCheck 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 |
|---|---|---|---|
| Add natural-language feedback and corrected captions to RL with 4 ground-truth captions | Image captioning on MS-COCO · MS-COCO test | Weighted metric: 106.55 -> 107.67 | Helped |
| Add feedback-network rewards with C captions to RL using 3 ground-truth captions | Image captioning on MS-COCO · MS-COCO test | Weighted metric: 106.55 -> 107.02 | Helped |
| Use A captions instead of C captions alongside feedback with 3 ground-truth captions | Image captioning on MS-COCO · MS-COCO test | Weighted metric: 107.02 -> 107.31 | Helped |
| Use feedback and corrected captions rather than a ground-truth caption in the single-caption RL setting | Image captioning on MS-COCO · MS-COCO test | Weighted metric: 104.93 -> 106.03 | Helped |
| Add corrected captions to MLE training | Hierarchical phrase-based RNN · MS-COCO test | Weighted metric: 104.12 -> 105.58 | Helped |
| Add phrase labels to the phrase-based captioner (ablation) | Hierarchical phrase-based RNN · MS-COCO test | Weighted metric: 102.14 -> 103.64 | Helped |
| Use two attention context vectors instead of one (ablation) | Hierarchical phrase-based RNN · MS-COCO test | Weighted metric: +0.6 points | Helped |
| Provide phrase-label linguistic information to attention (ablation) | Hierarchical phrase-based RNN · MS-COCO test | Weighted metric: +1.5 points | Helped |
| Replace the phrase-based captioner with a flat word-level captioner (ablation) | Image captioning on MS-COCO · MS-COCO test | Weighted metric: 102.14 -> 104.78 | Helped |
| Add the reference caption as input to the feedback network (ablation) | Feedback network phrase classification · Held-out feedback data | Accuracy: 73.30 -> 73.24 | No clear effect |
| Add the missing/wrong/redundant error category to the feedback network (ablation) | Feedback network phrase classification · Held-out feedback data | Accuracy: 73.30 -> 72.92 | Hurt |
| Add detailed error type such as action, object, or preposition to the feedback network (ablation) | Feedback network phrase classification · Held-out feedback data | Accuracy: 73.30 -> 74.66 | Helped |
Closest papers by meaning
- Reinforcing an Image Caption Generator Using Off-Line Human Feedback (AAAI 2020 · similarity 0.89)
- Controllable Image Captioning via Prompting (AAAI 2023 · similarity 0.87)
- CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026 · similarity 0.86)
- Deep Reinforcement Learning-Based Image Captioning With Embedding Reward (CVPR 2017 · similarity 0.86)
- 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)
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.