Research map / Image generation / Autoregressive and latent generation
Hierarchically Structured Reinforcement Learning for Topically Coherent Visual Story Generation
Qiuyuan Huang, Zhe Gan, Asli Celikyilmaz, Dapeng Wu, Jianfeng Wang, Xiaodong He
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 |
|---|---|---|---|
| use hierarchical topic planning and joint Manager-Worker reinforcement learning instead of flat RL | Visual story generation · VIST test (16 topics) | BLEU-4: 8.16 -> 11.64 | Improved |
| use hierarchical topic planning and joint Manager-Worker reinforcement learning instead of flat RL | Visual story generation · VIST test (16 topics) | CIDEr-D: 7.64 -> 9.73 | Improved |
| jointly train the Manager and Worker rather than train them iteratively (ablation) | HSRL · VIST test (16 topics) | BLEU-4: 11.23 -> 11.64 | Improved |
| jointly train the Manager and Worker rather than train them iteratively (ablation) | HSRL · VIST test (64 topics) | BLEU-4: 12.04 -> 12.32 | Improved |
| train Manager and Worker iteratively rather than independently in cascade (ablation) | HSRL · VIST test (16 topics) | BLEU-4: 10.38 -> 11.23 | Improved |
| train Manager and Worker iteratively rather than independently in cascade (ablation) | HSRL · VIST test (64 topics) | BLEU-4: 11.95 -> 12.04 | Improved |
| use learned topics rather than random topics | HSRL Worker · VIST test (16 topics) | BLEU-4: 4.70 -> 11.64 | Improved |
| provide ground-truth topics rather than learned topics | HSRL Worker · VIST test (16 topics) | BLEU-4: 11.64 -> 13.41 | Improved |
| increase topic count from 16 to 64 (ablation) | HSRL w/ Joint Training · VIST test | CIDEr-D: 9.73 -> 10.71 | Improved |
| increase topic count from 16 to 64 (ablation) | HSRL w/ Joint Training · VIST test | BLEU-4: 11.64 -> 12.32 | Improved |
| weight Worker loss more heavily than Manager loss with gamma_1=0.7 (ablation) | HSRL joint objective · VIST validation | overall performance | Improved |
| assign more weight to RL than MLE using gamma_2=0.9 (ablation) | HSRL joint objective · VIST validation | overall performance | Improved |
Closest papers by meaning
- Intelligent Grimm - Open-ended Visual Storytelling via Latent Diffusion Models (CVPR 2024 · similarity 0.87)
- Sketch Then Paint: Hierarchical Reinforcement Learning for Diffusion Multi-Modal Large Language Models (NeurIPS 2026 · similarity 0.87)
- Towards Unified Multimodal Interleaved Generation via Group Relative Policy Optimization (NeurIPS 2025 · similarity 0.87)
- Breaking the SFT Plateau: Multimodal Structured Reinforcement Learning for Chart-to-Code Generation (ICLR 2026 · similarity 0.86)
- VinaBench: Benchmark for Faithful and Consistent Visual Narratives (CVPR 2025 · similarity 0.86)
- ParaUni: Enhance Generation in Unified Multimodal Model with Reinforcement-driven Hierarchical Parallel Information Interaction (CVPR 2026 · 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.
