**Exploring the Design Space of Reward Backpropagation for Flow Matching**

> Aligning text-to-image flow matching models with human preferences via direct reward backpropagation is sample-efficient but hampered by two well-known pathologies: activations cannot be stored across the full sampling trajectory at modern model scale, and chained Jacobian products across steps inflate the reward gradient as it travels back to early indices. Connector-based methods, such as LeapAlign, address these i...

**来源信息**
- **来源**：HuggingFace Daily Papers（社区热门论文）
- **分类**：ai-models
- **发布时间**：2026-06-23 08:00（北京时间）
- **原文**：[打开原文](https://huggingface.co/papers/2606.11075)