sebastian.pth ✈️ICML

PhD student @ UniBern (CVG) 🎓🤖
Researching world models & video generation ✨

[sebastian.pth ✈️ICML @sstapf2000]

9/ Resources:

🌐 Project: https://t.co/AsUKIlrr2K
🤗 Weights: https://t.co/aOAwT0BCUo
📊 Dataset: https://t.co/jbVb9ulI4z 💻 Code: https://t.co/ccZHGIudfg

#WorldModels #VideoGeneration #ReinforcementLearning #Robotics #GenAI

[sebastian.pth ✈️ICML @sstapf2000]

1/🚀 We’re excited to share World Model Self-Distillation (WMSD)

WMSD trains pretrained video generators to solve general tasks from an image + short instruction; without curated task-execution videos.

It combines self-distillation with VLM-feedback RL, letting the Executor learn from a detailed-solution Demonstrator and improve beyond it.

🌐 Project: https://t.co/WneVN1zQ5B

#WorldModels #VideoGeneration #ReinforcementLearning #SelfDistillation #AI #Robotics

[sebastian.pth ✈️ICML @sstapf2000]

8/ Results: WMSD improves task-solving, agent grounding, and consistency across base models.

Even without task-specific supervision, WMSD reaches competitive performance compared with supervised fine-tuning baselines.

That is important because robotics data is expensive, and scalable task learning needs better alternatives."

[sebastian.pth ✈️ICML @sstapf2000]

@haoqik322 @prime_cai I think he’s referring to stuff like policy gradient RL

[sebastian.pth ✈️ICML @sstapf2000]

@k7agar exactly what we're working rn

[sebastian.pth ✈️ICML @sstapf2000]

@andrewgwils This clears up questions I’ve wrestled with for a long time. Cool stuff.