Contrastive World Models Train World Models Without Pixel Reconstruction
A new arXiv paper proposes Contrastive World Models, which learn latent dynamics via a mutual-information contrastive objective instead of pixel reconstruction, outperforming Dreamer in complex visual environments.
2026-09-26 ~ 2026-09-26 · 3 related posts
- Contrastive World Models: latent-space world models without pixel prediction — bonniesjli · 2026-09-26
- Contrastive World Models: latent-space world models beat pixel reconstruction under distractors — bonniesjli · 2026-09-26
- Contrastive World Models: dropping pixel reconstruction beats Dreamer in visually complex environments — bonniesjli · 2026-09-26