Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Learning to Caption Images Through a Lifetime by Asking Questions
Tingke Shen, Amlan Kar, Sanja Fidler
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 student-collected captions improved by asking questions, vs same amount of GT captions only | captioner · MSCOCO test | CIDEr: 91.5 -> 98.0 | Improved |
| train on teacher-improved captions instead of all ground-truth captions | captioner · MSCOCO test | CIDEr: 96.4 -> 98.0 | Improved |
| add question asking (inquisitive vs mute student) (ablation) | lifetime captioner · MSCOCO test | CIDEr: 95.9 -> 98.0 | Improved |
| add question asking at 10% warmup (ablation) | lifetime captioner · MSCOCO val | CIDEr: 95.6 -> 97 | Improved |
| add question asking at 1% warmup (ablation) | lifetime captioner · MSCOCO val | CIDEr: 64.3 -> 77 | Improved |
| reduce warmup from 10% to 1% at about 30% supervision | Inquisitive Student · MSCOCO val | CIDEr: 93.9 -> 83.5 | Hurt |
| use the full learned decision maker instead of asking no question (ablation) | decision maker · MSCOCO val | CIDEr: +6.5 CIDEr | Improved |
| pick time step by max entropy instead of random step (ablation) | decision maker · MSCOCO val | CIDEr: +0.3 CIDEr | No clear effect |
| add word closeness metrics to entropy for uncertainty (ablation) | decision maker · MSCOCO val | CIDEr: +1.0 CIDEr | Improved |
| learn the decision maker with RL instead of hand-set uncertainty rules (ablation) | decision maker · MSCOCO val | CIDEr: +3.1 CIDEr | Improved |
| add RNN caption encoding to question generator (ablation) | question generator · question gen. val | a@1: 37.8 -> 45.9 | Improved |
| add position encoding to question generator (ablation) | question generator · question gen. val | a@1: 37.8 -> 49.2 | Improved |
Closest papers by meaning
- Teaching Machines to Describe Images with Natural Language Feedback (NeurIPS 2017 · similarity 0.89)
- Proactive Agents for Multi-Turn Text-to-Image Generation Under Uncertainty (ICML 2025 · similarity 0.86)
- CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026 · similarity 0.84)
- OpenEQA: Embodied Question Answering in the Era of Foundation Models (CVPR 2024 · similarity 0.84)
- Seeing, Listening, Remembering, and Reasoning: A Multimodal Agent with Long-Term Memory (ICLR 2026 · similarity 0.84)
- Divide-and-Conquer: Tree-structured Strategy with Answer Distribution Estimator for Goal-Oriented Visual Dialogue (AAAI 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.
