Trajectory Straightening Loss for World Model Planning
ylecun · x · 2026-07-15
This post outlines the core idea of a world model planning research: reducing the difficulty of planning optimization by "locally straightening" latent trajectories.
The implementation is straightforward: increase the cosine similarity between adjacent velocity vectors by applying the loss
Lcurv = 1 - cos(vt, v(t+1)).
The post also shares two observations:
- Applying this loss to patch tokens is more effective than applying it to the CLS token.
- The method remains highly challenging for long-horizon planning; in the mentioned setup (planning horizon of 50, PushT dataset), the best success rate is only 36.00%.
The author also notes that this work predates Le-WM and wonders if anyone has tested this loss on Le-WM.
More from Research
- NeurIPS 2026 workshop will focus on on-device intelligence and local execution — YiMaTweets · 2026-07-21
- NeurIPS 2026 workshop calls papers on on-device intelligence — YiMaTweets · 2026-07-21
- AI Security Institute says every tested model tried to cheat in cyber evaluations — connoraxiotes · 2026-07-21
- AI companies are buying old books to avoid training on AI-generated slop — CackleRooster · 2026-07-21
- Sakana says multiple diffusion models plus MCTS beat test-time scaling on coding and math — SakanaAILabs · 2026-07-21
- Soofi S 30B-A3B releases a full pretraining report and claims open-model leads in English and German — abursuc · 2026-07-21